<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>用例AI / UseCase AI</title><link>https://ttqq.dev/</link><description>AI use cases real people actually tried, in Chinese and English</description><item><title>Georges Petrov has Claude Code film a demo video for each docs page with open-source Screenplay, and a nightly check goes red when a button is renamed / Georges Petrov 让 Claude Code 用开源的 Screenplay 给文档每一页拍功能演示视频，按钮改名后夜间检查会直接报红</title><link>https://ttqq.dev/en/tool/agent-screenplay-docs-videos-stale-check/</link><guid>https://ttqq.dev/en/tool/agent-screenplay-docs-videos-stale-check/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>A short video on every feature page of SaaS docs helps, but each takes an hour to make and goes stale after the next release. Georges Petrov (@georges_petrov) writes each video as a small script: an agent explores the feature in a demo account and writes the steps, and the open-source tool Screenplay clicks through them in a headless browser, films and edits the take (smoothed cursor, zooms that follow clicks, captions, optional voice). The same script is a test: check replays it fast without filming and fails when a button was renamed or removed, so a nightly CI job tells you which video is out of date. On Linux we installed Screenplay 0.1.7, filmed a 22-second 1080p video of its bundled example app, then renamed the "Take over" button; check failed at step 3 with exit code 2. / SaaS 文档里每个功能页配一段操作视频会更好懂，但录一段要一小时，产品一改版又过时了。Georges Petrov（@georges_petrov）把每段视频写成一个小脚本：agent 先在演示账号里摸清功能、写好步骤，开源工具 Screenplay 在无头浏览器里照着点、录屏并剪辑（平滑光标、跟着点击放大、字幕、可选配音）。同一个脚本也是测试：check 快速跑一遍不录像，按钮被改名或删掉时就失败，放进每晚的 CI 里就能知道哪段视频过期了。我们在 Linux 上装了 Screenplay 0.1.7，用它自带的示例应用完整录了一段 22 秒的 1080p 视频，再把「Take over」按钮改名，check 在第 3 步失败、退出码为 2。</description></item><item><title>MPC gave Claude Code three backend skills and cut critical mistakes on 15 hidden-check backend tasks from 25 in 150 runs to 2 / MPC 给 Claude Code 装上三个后端 skill，在 15 个带隐藏检查的后端任务上把严重失误从 150 次里的 25 次降到 2 次</title><link>https://ttqq.dev/en/tool/backendskills-hidden-checks-25-to-2-critical/</link><guid>https://ttqq.dev/en/tool/backendskills-hidden-checks-25-to-2-critical/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>AI-written backends often work in the demo and break in production on concurrency, permissions and schema changes. backendskills is three open-source skills from MPC (globalmpc): api-endpoint (scope queries to the caller, 4xx on bad input, rate-limit login and costly routes), data-integrity (locks or conditional updates for read-then-write, UNIQUE constraints, idempotency keys on anything that charges or sends) and db-migration (expand then contract, concurrent indexes, batched backfills, down migrations). The authors built 15 backend tasks, each with a hidden check the agent never sees (20 buyers at once for 5 items, reading someone else's order, old and new app versions mid-deploy), and ran Sonnet 5.5 and Haiku 5.5 five times each with and without the skills. Critical mistakes fell from 25/150 to 2/150, mostly thanks to the login-rate-limit and rename-column tasks; each run cost roughly twice as much. We recomputed the published raw runs and ran the benchmark's 22 self-tests against a local PostgreSQL. / 让 AI 写后端，演示时能用，上线后常在并发、权限和改表结构时出事。backendskills 是 MPC（globalmpc）开源的三个 skill：api-endpoint（查询按调用者隔离、坏输入返回 4xx、登录和贵接口限流）、data-integrity（先读后写加锁、唯一约束、扣钱和发消息加幂等键）、db-migration（先扩后缩、并发建索引、分批回填、写回滚迁移）。作者准备了 15 个后端任务，每个任务带一段 agent 看不到的隐藏检查（例如 20 人同时抢 5 件库存、读别人的订单、新旧版本混跑），用 Sonnet 5.5 和 Haiku 5.5 各跑 5 遍对照。严重失误从 25/150 降到 2/150，差距主要来自登录限流和改列名两个任务；代价是每次多花约一倍费用。中文圈由 @lovethelife2333 介绍过来。我们重算了作者公开的原始记录，并在本机 PostgreSQL 上运行了测评框架的 22 个自检，全部通过。</description></item><item><title>Guizang built a xianxia 3D action game, Mist-Locked Mountain Gate, in one afternoon with Claude Fable 5.2, then turned the lessons into an open-source skill / 歸藏用 Claude Fable 5.2 一个下午做出仙侠 3D 动作游戏《雾锁山门》，再把踩过的坑整理成一个开源 skill</title><link>https://ttqq.dev/en/tool/guizang-game-skill-fable-3d-action-game/</link><guid>https://ttqq.dev/en/tool/guizang-game-skill-fable-3d-action-game/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Guizang (@op7418) only asked for "Chinese xianxia style" and had Claude Fable 5.2 build a top-down 3D action game, Mist-Locked Mountain Gate, in one afternoon: one map, a mob fight and a two-phase boss, with a three-hit combo, dodge, block and parry, an ultimate and 8 voiced lines, packed into one ~33 MB HTML file that plays in desktop and mobile browsers. Characters started as concept art in Jimeng, became rigged 3D models in Tripo, and the music came from Suno. He then wrote the workflow and his list of pitfalls into guizang-game-skill: install it in Claude Code or Codex, describe your game, and it asks questions first, connects your asset tools, checks its own screenshots and packages the result. We inspected the repo, ran its local server script and scanned it with SkillSpector; we did not build a game. / 歸藏（@op7418）只说了「要做国风仙侠」，让 Claude Fable 5.2 在一个下午做出俯视角 3D 动作游戏《雾锁山门》：一个场景、一段小怪战、一场两阶段 Boss 战，有三段连招、闪避、格挡弹反、大招和 8 条配音，打包成一个约 33MB 的 HTML，电脑和手机浏览器都能直接玩。角色先用即梦出设定图，再用 Tripo 转成带骨骼的 3D 模型，配乐用 Suno。之后他把这次的流程和踩坑清单写成 guizang-game-skill，装进 Claude Code 或 Codex 后，说出想做的游戏，它会先问清需求、接上你的素材工具、自己截图核对，最后打包。我们检查了仓库、跑了它的本地服务器脚本并用 SkillSpector 扫了一遍，没有自己做游戏。</description></item><item><title>Oil spent one to two months having AI build Oil Studio, a Mac screen recorder and editor, in Swift he didn't know, and wrote up how he keeps the repo from falling apart / 欧呦用不熟的 Swift 让 AI 开发了一两个月 Mac 录屏剪辑工具 Oil Studio，总结出一套不让仓库失控的长期开发流程</title><link>https://ttqq.dev/en/tool/oil-studio-ai-long-project-governance/</link><guid>https://ttqq.dev/en/tool/oil-studio-ai-long-project-governance/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Oil (@I_am_oil_oil, author of oil-ui and oil-tone) had AI build Oil Studio, a macOS screen recorder and editor, from scratch in Swift, a language he hadn't used. Left unchecked over a long project, he found five kinds of mess: every screen with its own styles, hard-coded values everywhere, the same function written twice, old code never deleted, and stale docs misleading later work. His process has six parts: have AI research open-source projects and pick the language before coding; let Claude lead the conversation and write plans while oil-delegate hands critique, coding and review to Codex; develop in parallel with Git Worktree and split tests into layers; write recurring pitfalls into AGENTS.md and Review.md; have AI periodically check docs against the code and prune them; for stubborn bugs, have two strong models investigate and give a confidence level, and gate releases with scripted checks. We ran oil-delegate's unit tests and environment check; the overall process is untested. / 欧呦（@I_am_oil_oil，oil-ui、oil-tone 的作者）用自己以前没接触过的 Swift，让 AI 从零开发了 macOS 录屏剪辑工具 Oil Studio。他发现不加约束地长期迭代，仓库会出现五类混乱：每处单独写一套样式、到处硬编码、重复实现同一个函数、旧方案的代码不删、过期文档误导后续开发。他的做法分六步：动手前让 AI 调研开源项目和语言选型；Claude 做主对话和方案，通过 oil-delegate 把挑刺、写代码和审查交给 Codex；用 Git Worktree 并行开发并把测试分层；把反复踩的坑写进 AGENTS.md 和 Review.md；定期让 AI 对照代码清理过期文档；久修不好的 Bug 让两个强模型交叉排查，发版用程序卡住测试和版本号。我们跑了 oil-delegate 的单元测试和环境检查，整套流程没有实测。</description></item><item><title>Ganesh Iyer wanted to animate a story his daughter wrote, so he used Claude Code with the open-source tool REA to tear down two story-animation apps and see how they work / Ganesh Iyer 想给女儿写的故事做动画，先用 Claude Code 加开源工具 REA 拆解了两款故事动画产品，看清它们是怎么做的</title><link>https://ttqq.dev/en/tool/rea-product-teardown-skill-claude-code/</link><guid>https://ttqq.dev/en/tool/rea-product-teardown-skill-claude-code/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>Ganesh Iyer (@ganeshiyer316) wanted a short animated video of a story his 7-year-old daughter wrote. Existing tools were pricey and didn't fit, so he tore down two of them first. He made one story in each app while the open-source tool REA (Reverse Engineer Anything) recorded what the app sent and received, keeping only field names, and had Claude turn that into a teardown report with his product-teardown skill: front-end steps, API calls, where media comes from and which AI providers are likely. Every finding is labeled Seen (in the traffic) or Best guess. He shares the skill free on GitHub. We ran the skill's bundled domain logger on our server; we did not install REA or tear down a real product. / Ganesh Iyer（@ganeshiyer316）想给 7 岁女儿写的故事做一段动画短片，现成工具又贵又不太合适，于是先拆解两款产品看看它们怎么做。他在每个应用里各做了一个故事，同时让开源工具 REA（Reverse Engineer Anything）记录应用收发的网络请求，只保留字段名、不记录内容，然后让 Claude 按他写的 product-teardown skill 生成一份拆解报告：前端步骤、调用了哪些接口、媒体文件从哪来、可能用了哪家 AI。每条结论分成「看到的」（在流量里出现过）和「推测的」两类。他把这个 skill 免费放在 GitHub 上。我们在服务器上跑了 skill 自带的域名记录脚本，没有装 REA，也没有拆解真实产品。</description></item><item><title>KKman scans unfamiliar agent skills with NVIDIA's open-source SkillSpector before installing; Anthropic's own docx skill scored 86, and reading the flagged lines showed normal file handling / KKman 装陌生人的 skill 之前先用英伟达开源的 SkillSpector 扫一遍，三份样本里官方 docx 技能被打了 86 分，翻看标红的行才确认是正常操作</title><link>https://ttqq.dev/en/tool/skillspector-scan-skills-before-install/</link><guid>https://ttqq.dev/en/tool/skillspector-scan-skills-before-install/</guid><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><description>A skill is an instruction file for your AI assistant, so installing a stranger's skill means letting a stranger give your assistant orders. KKman (@JackLuiKK) cloned NVIDIA's open-source scanner SkillSpector (Apache-2.0) on Linux and ran static checks only on three samples: the repo's clean example scored 0, its malicious example 100, and Anthropic's official docx skill 86 with a DO NOT INSTALL verdict. Opening the flagged lines, he found normal steps for repacking Word files and launching LibreOffice. His takeaway: the score is where you start, and you still read the flagged lines. We reran the same three scans and got the same scores, issue counts and coverage, then scanned two newer skills and saw the same pattern. / skill 是写给 AI 助手的说明书，装一个陌生人写的 skill，等于让陌生人给你的助手下指令。KKman（@JackLuiKK）在 Linux 上克隆 NVIDIA 开源的扫描器 SkillSpector（Apache-2.0），只开静态检查，扫了三份样本：仓库自带的干净示例 0 分、恶意示例 100 分，Anthropic 官方的 docx 技能 86 分、结论「不要安装」。他打开标红的位置看，发现是重新打包 Word 文件和启动 LibreOffice 的正常步骤。他的结论是：分数是起点，标红的行还得自己看。我们照着重跑了这三份，分数、问题条数和覆盖率都和他一致；又多扫了两份新技能，看到同样的情况。</description></item><item><title>Meng Jian built a stick-figure animation skill in Claude Code; one 15-second 60 fps explainer cost $0.54, and changing a line needs no re-generation / 孟健用 Claude Code 自建火柴人动画 Skill，做一条 15 秒 60 帧讲解短片花了 0.54 美元，改台词不用重新抽卡</title><link>https://ttqq.dev/en/tool/claude-code-stickman-skill-explainer-short-054/</link><guid>https://ttqq.dev/en/tool/claude-code-stickman-skill-explainer-short-054/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>Meng Jian (@xjtumj) makes explainer shorts for overseas audiences. With video-generation models, credits ran out fast, characters drifted between shots, and changing one line meant paying to re-generate. He switched to stick-figure animation written as code with Claude Code: first, with no skill, Claude wrote a 15-second animation page in 64 seconds, but the figure's proportions and arms were off; then he had Claude build a stickman-animator skill with a spec, an engine.js that locks the main character's bone lengths, and an export-mp4.js that records the page with the local browser (about 8 minutes, 11 preset poses such as waving, pointing, thinking, panicking); in a new chat he only described the topic and three scenes, and Claude picked up the skill on its own and produced an animation with Chinese–English subtitles; adding floating text and a badge was a code edit; exporting took 52.2 seconds for a 1280×720, 60 fps, 900-frame, 0.81 MB MP4. /cost showed $0.5391 for the whole session. We did not reproduce this, and the article does not publish the skill files. / 孟健（@xjtumj）做出海解说短视频，用生视频模型时额度消耗快、前后镜头容易崩脸，改一句台词就要重新抽卡付费。他改成在本地用 Claude Code 写纯代码的火柴人动画：先不挂任何 Skill 让 Claude 裸写一个 15 秒动画网页（64 秒写完，但火柴人头身比例失调、手臂歪斜）；再让 Claude 建一个 stickman-animator Skill，里面有规范文件、固定主角骨骼比例的渲染库 engine.js 和调用本机浏览器录成 MP4 的 export-mp4.js（约 8 分钟，含挥手、指点、思考、惊慌等 11 个姿势）；新开对话只描述主题和 3 幕分镜，Claude 自动调用这个 Skill 做出带中英字幕的动画；加飘字和徽章只要改代码；最后 52.2 秒导出 1280×720、60fps、900 帧、0.81MB 的 MP4。/cost 显示整轮会话共 0.5391 美元。我们没有复现这个流程，原文也没有公开 Skill 文件。</description></item><item><title>Cristian Calabro ran Google's open-source EGM plugin in Codex on a docs-only change, designing and reviewing it before coding, and it caught three things he would have shipped / Cristian Calabro 在 Codex 里用谷歌开源的 EGM 插件给一个只改文档的小功能先写设计再审，抓出三处他本来会漏掉的问题</title><link>https://ttqq.dev/en/tool/egm-elephant-goldfish-codex-small-feature-review/</link><guid>https://ttqq.dev/en/tool/egm-elephant-goldfish-codex-small-feature-review/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>EGM (Elephant-Goldfish Model) is a design-review plugin Google open-sourced in November: the "elephant" agent holds project context and drafts the spec, "goldfish" reviewer subagents read only the document and referenced code with no earlier conversation, and you decide. Cristian Calabro (@c_calabro) tried it on a tiny request in a repo-checking CLI he's building: make scan --targets show JSON and YAML samples in --help and CLI.md. In Codex CLI he went through brainstorm, draft, review, implementation and a drift report. Results: the brainstorm's seven review lenses warned the four samples could drift apart, so a parity check went into the acceptance criteria; the implementation-plan review noticed the scanner check would write reports into the repo being scanned, against his AGENTS.md, so it moved to a temp folder outside the repo; the closing drift report found the docs never said commit values are compared exactly. On Linux we cloned the EGM repo (0.1.1, Apache-2.0) and ran its offline unit tests: 381 passed, 2 skipped. We did not install the plugin in Codex or Claude Code and run the workflow. / EGM（Elephant-Goldfish Model）是谷歌 11 月开源的设计评审插件：「大象」是掌握项目上下文、负责写规格的 agent，「金鱼」是不带前面对话、只读文档和相关代码的评审子 agent，人来拍板。Cristian Calabro（@c_calabro）在自己写的一个仓库检查 CLI 上试了一个很小的需求：让 scan --targets 的帮助和 CLI.md 里给出 JSON 和 YAML 示例。他在 Codex CLI 里按 brainstorm、draft、review、实现、drift report 走了一遍。结果：头脑风暴阶段 7 个评审角度指出四处示例可能互相走样，于是验收标准里加了一致性检查；实现计划评审发现扫描检查会把报告写进被扫描的仓库，违反他 AGENTS.md 的规定，改成写到仓库外的临时目录；收尾的 drift report 发现文档没写「commit 值按完全一致比较」。我们在 Linux 上克隆了 EGM 仓库（0.1.1，Apache-2.0），离线单元测试 381 个通过、2 个跳过；没有在 Codex 或 Claude Code 里实际装插件走流程。</description></item><item><title>ZinkLu had Claude review 1,050 of his agents' memory recalls one by one; about one in ten was confirmed used, and he drew practical rules for using memory better / ZinkLu 让 Claude 逐个复核自己 1,050 次 AI 记忆召回，能确认被用上的约一成，并总结出几条更省 token 的用法</title><link>https://ttqq.dev/en/tool/nowledge-mem-agent-memory-audit-1050-recalls/</link><guid>https://ttqq.dev/en/tool/nowledge-mem-agent-memory-audit-1050-recalls/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>ZinkLu (@ZinkLu2) uses Codex, Claude Code, OpenCode, Kimi Code and Hermes across a Mac and two Linux machines, all sharing one self-hosted Nowledge Mem memory store. After installing it he wanted to know how much it actually helped. He gathered 3,230 sessions, kept the 1,078 after memory was connected, defined a "recall window" as the stretch from one recall to the next (1,050 in total), and had Claude Opus 5.5 check each one against the prior context, the returned memory and what the agent did next; every "used" verdict needed a verbatim quote, checked by a script. Results: 101 windows confirmed used, 9.6% strict or 12.4% lenient; memory took a median 4.1% of context; empty results and errors were 28% of recalls. The details are the useful part: the first recall in a session was used 19.6% of the time versus 9.5% after; queries over 60 characters were used 18.2% versus 3.3% for under 15; retrying right after an empty result or error rarely helped; memories under a day old were used more than those over 30 days; and on the same 120 real queries, memory search gave directly usable results 25.8% of the time versus 6.7% for searching past sessions. We did not reproduce this analysis. / ZinkLu（@ZinkLu2）在一台 Mac 和两台 Linux 开发机上用 Codex、Claude Code、OpenCode、Kimi Code 和 Hermes，几个 agent 共用一套自己部署的 Nowledge Mem 记忆库。装了记忆系统后他想知道：它到底帮了多少忙？他整理出 3,230 份会话，筛出接入记忆后的 1,078 个，把「一次召回到下一次召回之间」定义为一个窗口，共 1,050 个，让 Claude Opus 5.5 逐个对照召回前的上下文、返回内容和后续动作，判断记忆有没有被用上，每个「采用」都要给出逐字证据并用程序核对引文。结果：101 个窗口确认用上了，严格口径 9.6%、宽松口径 12.4%；记忆内容占上下文的中位数 4.1%；空结果和报错占 28%。更有用的是细节：一个会话里第一次召回的采用率 19.6%，之后只有 9.5%；60 个字符以上的查询采用率 18.2%，15 个字符以内只有 3.3%；搜空或报错后马上再搜基本没用；不到 1 天的新记忆比 30 天以上的更常被用上；同样的 120 条真实查询，搜记忆有 25.8% 能直接用，搜历史会话只有 6.7%。我们没有复现这项分析。</description></item><item><title>ZDNET's David Gewirtz locked OpenClaw on its own Mac Studio with Claude Code as the warden, authorized 11 escape attempts, and all failed / ZDNET 的 David Gewirtz 把 OpenClaw 关在一台单独的 Mac Studio 上，让 Claude Code 当看守，授权它试了 11 种越狱方法，全部失败</title><link>https://ttqq.dev/en/tool/openclaw-mac-studio-claude-code-warden-11-escape-attempts/</link><guid>https://ttqq.dev/en/tool/openclaw-mac-studio-claude-code-warden-11-escape-attempts/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>David Gewirtz (@DavidGewirtz, ZDNET columnist) wanted the local agent OpenClaw running 24/7 without letting it loose. His setup is "warden and prisoner": OpenClaw runs on an old M1 Mac Studio under its own user with no access to any of his accounts, using GLM-5.2 through a $20/month Ollama Cloud plan; Claude Code runs on his own Mac and only tunnels in over SSH to configure and check, then shuts down. Along the way Claude found two problems: its own instruction files first sat where OpenClaw could read them (the warden's notes in the prisoner's cell), and OpenClaw's file-read size limit meant it had never read his two most important rules (ask David before anything leaves the machine; never claim a failed tool call succeeded), which Claude spotted in OpenClaw's own logs. Then, with his authorization, Claude had OpenClaw actively try to break into the admin side. It tried 11 ways (writing admin files, privilege back doors, scripts and links around permissions, pointing its browser at local file URLs) and macOS permissions stopped all of them. For email, he had Claude Code write a script that only mails his one address, and OpenClaw can only drop text files into a folder. We did not reproduce this setup. / David Gewirtz（@DavidGewirtz，ZDNET 专栏作者）想 24 小时跑本地智能体 OpenClaw，又怕它乱来。他的做法是「看守和囚犯」：OpenClaw 装在一台旧 M1 Mac Studio 上，用一个碰不到他任何个人账号的独立用户运行，模型用每月 20 美元的 Ollama Cloud 上的 GLM-5.2；Claude Code 跑在他自己的 Mac 上，只在需要时通过 SSH 隧道进去配置和检查，用完就关。过程中 Claude 发现两个问题：它自己的说明文件最初放在 OpenClaw 能读到的位置（看守的笔记放在了牢房里）；OpenClaw 读配置文件有长度上限，那两条最重要的规则（离开本机的操作先问 David、工具失败不许说成功）根本没读到，是 Claude 从 OpenClaw 的日志里发现的。随后在他的授权下，Claude 让 OpenClaw 主动尝试闯进管理员一侧，它试了 11 种方法（直接写管理员文件、走后门提权、用脚本和链接绕权限、用浏览器打开本地文件地址等），全部被 macOS 权限挡住。发邮件这类能力，他改成让 Claude Code 写一个只往他一个邮箱发信的脚本，OpenClaw 只能往投递文件夹里放文本。我们没有复现这套环境。</description></item><item><title>Gary Zhang pinned a to-do list in cmux's side Dock that opens one Claude Code chat per item, and on day one it caught two things booked at 5 pm / Gary Zhang 在 cmux 右侧常驻一张待办清单，每件事一键开一个 Claude Code 对话，上线第一天就发现两件撞在 17 点的事</title><link>https://ttqq.dev/en/tool/today-board-cmux-todo-dock-agent-workspace/</link><guid>https://ttqq.dev/en/tool/today-board-cmux-todo-dock-agent-workspace/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>Gary Zhang (@GaryZhang124) built a home assistant: an always-on Mac mini turns each day's messages into a briefing. The last step was the problem: what to start on in the morning, when every item was buried in chats. His fix is Today Board: a small local-only web list pinned in the Dock of cmux (a terminal that runs several agents side by side), with five tabs: schedule, in progress, to confirm, done, ignored. Clicking an item's "处理" (handle) opens a cmux workspace named after it, with Claude Code already given the item's text and time. Items a script extracts land in "to confirm" until he approves them. The list itself is one Markdown file. On day one it flagged a 5 pm interview from work messages and a 5 pm class from WeChat as a clash. We tested it on Linux: the 3 bundled tests passed; adding items, changing status and importing an inbox JSON (duplicates skipped) all worked; the page returned 200 and marked two 5 pm items as a time clash. Without macOS and cmux we could not test the Dock or the handle-to-workspace button. / Gary Zhang（@GaryZhang124）给家里搭了个人助理：一台常开的 Mac mini 每晚把当天消息整理成简报。卡在最后一步：早上坐下来先做哪件事？事项全埋在对话里。他的解法是 Today Board：cmux（一个能同时开多个 agent 的终端）右侧的 Dock 里常驻一张只监听本机的小网页清单，分日程、要推进、待确认、已完成、已忽略五栏；点某件事的「处理」，cmux 就新开一个以它命名的工作区，Claude Code 已经拿到这件事的内容和时间。自动抽出来的事项先进「待确认」，他点头才算数。清单本身就是一个 Markdown 文件。上线第一天，它把来自工作消息的 17 点面试和来自微信的 17 点课程标成了时间冲突。我们在 Linux 上测了：3 个自带测试通过，命令行加事项、改状态、从收件箱 JSON 导入（自动跳过重复）都正常，网页返回 200，并对两条同为 17 点的事项显示「时间冲突」。没有 macOS 和 cmux，所以 Dock 挂载和「处理」开工作区没测。</description></item><item><title>Rajesh made a Claude Code skill that turns a tech topic into a 1080p explainer with hand-drawn animation, a cloned voice and sound effects, all on your machine / Rajesh 做了一个 Claude Code 技能，输入一个技术话题就在本机做出带手绘动画、克隆配音和音效的 1080p 讲解视频</title><link>https://ttqq.dev/en/tool/under-the-hood-claude-code-narrated-explainer-video/</link><guid>https://ttqq.dev/en/tool/under-the-hood-claude-code-narrated-explainer-video/</guid><pubDate>Fri, 13 Nov 2026 00:00:00 GMT</pubDate><description>Rajesh (@royalreplaced) open-sourced under the hood: in Claude Code you say "make a video about what happens when you git push", and the skill checks your setup, asks one or two questions (which angle, which voice), shows you the narration and storyboard to approve, then writes a canvas animation, screenshots every scene and fixes overlaps itself, generates two takes per line locally with the open-source TTS model Chatterbox for you to pick, and renders a 1080p MP4 with music that ducks under the voice, a sound on every beat, and loudness at -16 LUFS for X and YouTube. You can also record 20–30 seconds of your own voice to clone. His demo is "what happens when you type cookie into Google". On Linux we rendered the repo's bundled cookie example without narration: a 33.4-second 1080p30 video in about 27 seconds, with complete frames. We did not test narration, voice cloning or Claude writing a new topic. / Rajesh（@royalreplaced）开源了 under the hood：在 Claude Code 里说「做一个讲 git push 发生了什么的视频」，技能会先检查环境，问一两个问题（讲哪个角度、用哪种声音），把旁白稿和分镜给你确认，然后写画布动画、给每个场景截图并自己修重叠，在本机用开源 TTS 模型 Chatterbox 生成每句两条配音让你挑，最后渲染成 1080p MP4：音乐在说话时自动压低，每个动作配音效，响度调到适合发 X、YouTube 的 -16 LUFS。也可以录 20–30 秒自己的声音来克隆。作者的示例是「在 Google 里输入 cookie 时发生了什么」。我们在 Linux 上用仓库自带的 cookie 示例、不带旁白渲染了一遍：33.4 秒的 1080p30 视频用时约 27 秒，抽帧画面完整。没有测配音、声音克隆和 Claude 写新话题的部分。</description></item><item><title>TaXue spent five months building Argo, a search skill that gives Claude Code, Codex and other agents 300 sources and works without an API key / 踏雪寻仙花 5 个月做了 Argo，给 Claude Code、Codex 等智能体装上 300 个搜索源，不配 API Key 也能中英文检索</title><link>https://ttqq.dev/en/tool/argo-agent-search-skill-300-sources/</link><guid>https://ttqq.dev/en/tool/argo-agent-search-skill-300-sources/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Many agents search the web through one or two engines, so a question about a company returns old news and a question about a product returns advertorials. TaXue (@TaXue2025) built Argo for that: a search and deep-research tool for agents, installable as a Skill, an MCP server or a DeepSeek Harness plugin. v2.14.0 lists 300 sources and picks sources by topic and language, de-duplicates and ranks the results, and can also fetch page text, WeChat articles, PDFs and papers. Without API keys it uses only free and local sources; sources with keys join automatically. We cloned v2.14.0 (1345696) and installed only pyyaml, with no keys: 237 sources were callable on our machine; an English query returned 5 results (including the original post and blog) in about 2.8 s, and a Chinese query returned 5 Chinese results from Sogou WeChat and anysearch in about 2.7 s. We did not test deep research, logged-in search or the MCP setup. / 很多智能体的联网搜索只接了一两个搜索引擎，查公司新闻常拿到旧闻，查产品口碑常拿到软文。踏雪寻仙（@TaXue2025）为此做了 Argo：一个给智能体用的搜索与深度研究工具，同时提供 Skill、MCP 和 DeepSeek Harness 插件三种装法，v2.14.0 收录 300 个数据源，按问题领域和语言自动挑源、去重、排序，还能抓网页正文、公众号文章、PDF 和论文。没配 API Key 时只用免费源和本地源，配了 Key 的源会自动加入。我们克隆了 v2.14.0（1345696），只装 pyyaml，不配任何 Key：本机可直接调用的源有 237 个；英文查询约 2.8 秒返回 5 条结果（含原帖和博客），中文查询约 2.7 秒从搜狗微信和 anysearch 返回 5 条中文结果。没有测深度研究、登录态搜索和 MCP 接入。</description></item><item><title>Hamza has Claude Code write an HTML page with a render(t) function, screenshots it frame by frame with Playwright and encodes with ffmpeg, making 30-second motion graphics at no cost beyond his Claude subscription / Hamza 用 Claude Code 写带 render(t) 的网页，再用 Playwright 逐帧截图、ffmpeg 合成，在已有订阅之外不花钱做出 30 秒动效视频</title><link>https://ttqq.dev/en/tool/claude-code-motion-graphics-render-t-playwright-ffmpeg/</link><guid>https://ttqq.dev/en/tool/claude-code-motion-graphics-render-t-playwright-ffmpeg/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>In late September, X filled with Claude motion videos credited to "one prompt," and most copies came out with five fonts, a rainbow palette and every word flying in from a different direction. Hamza (@imhmxi) wrote up the process he actually uses: a brief and a timed scene list first, then a STYLE.md of restraint rules (one type pairing, one accent colour, one easing), then Claude Code writes a page that only draws the frame for a time t, Playwright screenshots it at 30 fps, and ffmpeg encodes the frames. After each render he makes a contact sheet of sampled frames for himself and Claude to review, then adds sound and captions. He says it costs nothing beyond an existing Claude subscription. Following the method, we wrote a 3-second 1080×1920 page (headline fade-in, accent rule, caption), captured 90 JPEG frames with Playwright in about 5 s, encoded a 3-second MP4 with ffmpeg and built a 3×3 contact sheet; each element appeared on schedule. We did not test sound, captions or his full prompt templates. / 9 月底 X 上流行「一句提示词」做出的 Claude 动效视频，但跟着抄提示词的人做出来大多是五种字体、彩虹配色、每个字从不同方向飞进来。Hamza（@imhmxi）写了他实际在用的流程：先写简报和分镜时间表，再写一份限制风格的 STYLE.md（一组字体、一个强调色、一种缓动），然后让 Claude Code 写一个只按时间 t 画画面的网页，用 Playwright 按 30fps 逐帧截图，ffmpeg 编码成视频；每次渲染后拼一张帧缩略图总览给自己和 Claude 检查，最后补声音和字幕。他说这样除了已有的 Claude 订阅不用再花钱。我们照文中的方法写了一个 3 秒的 1080×1920 竖屏页面（标题淡入、强调线展开、说明文字出现），用 Playwright 截了 90 帧 JPEG（约 5 秒），ffmpeg 编码出 3 秒 MP4，再生成 3×3 帧总览图，各元素按设定时间依次出现。没有测声音、字幕和他文中的完整提示词模板。</description></item><item><title>Rohit Kashyap built Dejo, a browser MCP that lets Claude Code learn a website task once and replay it with no model calls; he measured 78% fewer tokens on repeat runs / Rohit Kashyap 做了浏览器 MCP Dejo，让 Claude Code 第一次走完网站任务后记住步骤，之后重跑不调用模型，他实测重复运行省 78% token</title><link>https://ttqq.dev/en/tool/dejo-browser-mcp-remembers-websites/</link><guid>https://ttqq.dev/en/tool/dejo-browser-mcp-remembers-websites/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Driving websites through Playwright MCP means the agent re-reads every page each time, so a task it did yesterday costs full tokens again today. Rohit Kashyap (@rohit_jsfreaky) built Dejo, a browser MCP for agents: the first time, the agent walks the task and Dejo saves each step together with what should be true afterwards; after that the same task is one tool call, and Dejo replays it without reading the page or calling a model. When a site changes, it stops at the step that didn't land and asks the agent to relearn only that step. Against Playwright MCP on 19 tasks across 11 lesser-known sites, he reports 38% fewer tokens on the first run and 78% fewer on every run after, with the data public. We pip-installed dejo-mcp 0.5.1 and dejo doctor passed; acting as the agent through a minimal MCP client, we taught it to read the first paragraph on example.com, then replayed it twice from a fresh process with dejo run: 105 ms and 85 ms, model_calls 0. We did not reproduce his 19-task benchmark. / 用 Playwright MCP 让智能体操作网站，每次都要重新读页面，同一件事昨天做过，今天照样全价付 token。Rohit Kashyap（@rohit_jsfreaky）做了 Dejo：一个给智能体用的浏览器 MCP，第一次由智能体带着走一遍，Dejo 把每一步连同「做完应该看到什么」一起存下来；之后同一任务只要一次工具调用，Dejo 自己按记录重放，不读页面、不调模型。网站改版时，它停在没对上的那一步，只让智能体重学这一步。他用 19 个任务、11 个小众网站和 Playwright MCP 对比：首次运行少用 38% token，之后每次少用 78%，数据公开。我们用 pip 装了 dejo-mcp 0.5.1，dejo doctor 全部通过；自己写了一个最小 MCP 客户端充当智能体，在 example.com 上教会它读取正文第一段，然后在新进程里用 dejo run 重放两次，各用 105 毫秒和 85 毫秒，model_calls 为 0。没有复现他的 19 任务基准。</description></item><item><title>caiman, a visual designer who doesn't code, used Claude Code to get an iOS app approved on the App Store in 11 days, then open-sourced the six-agent workflow as a plugin / 不写代码的视觉设计师 caiman 用 Claude Code 11 天把 iOS 应用送上 App Store，再把 6 个 agent 的流程开源成插件</title><link>https://ttqq.dev/en/tool/designer-app-flow-claude-code-ios-app-store/</link><guid>https://ttqq.dev/en/tool/designer-app-flow-claude-code-ios-app-store/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>caiman (@caiman__) is a brand and motion designer who doesn't write Swift. With Claude Code he built BeanDance, a native iOS app that remembers the grind size for each bag of coffee: 11 days from new project to first approval, 5 releases and 21 TestFlight builds in a month, live in 175 storefronts including mainland China, with 320 rule tests. He says the hard part wasn't code but three questions: who decides, how to know it's right, and how to ship. So he set rules: every screen starts as 2–3 directions on a canvas and code is written only after he picks one; decisions go into docs, and thinking out loud in chat doesn't count; the agent that writes code doesn't approve its own work, because a read-only review agent and a black-box tester that only taps through the simulator check it; uploading, submitting and releasing are triggered only by him. He packaged the roles, doc templates and 150+ lessons as the Claude Code plugin designer-app-flow. We cloned it and ran its init script in an empty folder: the 6 agents and 3 skills were renamed for the new app, and the script listed the placeholders still to fill (Team ID, Bundle ID and so on). We did not build an iOS app or go through App Store submission. / caiman（@caiman__）是做品牌和动效的视觉设计师，不会写 Swift。他用 Claude Code 做了原生 iOS 应用 BeanDance，用来记住每包咖啡豆的研磨度：从建工程到首版过审 11 天，一个月发了 5 个版本、21 个 TestFlight 构建，上架 175 个地区（含中国大陆），有 320 条规则测试。他说难点不在写代码，而在三件事：谁来做决定、怎么知道做对了、怎么上架。于是他定了规矩：任何界面先出 2–3 个方向，他选定后才写代码；决定写进文档，聊天里随口说的想法不算决定；写代码的 agent 不能自己验收，另有一个只读代码的验收 agent 和一个只在模拟器上按用户路径点的试用 agent；上传、送审、发布只能由他本人触发。他把这套分工、文档模板和 150 多条踩坑记录做成了 Claude Code 插件 designer-app-flow。我们克隆了仓库，在空目录里运行了插件自带的初始化脚本，6 个 agent 和 3 个技能都按新名字改好了，脚本也列出了还要手填的占位（Team ID、Bundle ID 等）。没有编译 iOS 应用，也没有走上架流程。</description></item><item><title>Leaf paired the open-source cleaner Mole with Claude to cross-check scans, freeing about 315 GB over three and a half months on a 500 GB MacBook without losing personal files / Leaf 用开源清理工具 Mole 加 Claude 交叉扫描，三个半月从 500G 的 MacBook 上累计清出约 315G，个人文件没有丢</title><link>https://ttqq.dev/en/tool/mole-claude-mac-disk-cleanup-315g/</link><guid>https://ttqq.dev/en/tool/mole-claude-mac-disk-cleanup-315g/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>Leaf (@leaf_sanren) filled a 500 GB MacBook 10 times in three and a half months. His method scans twice: first Mole, tw93's open-source cleanup CLI, with mo clean --dry-run as a preview, then Claude over the home folder, and the two lists are reconciled. Items on both lists are safe; items only Mole found mean Claude missed them (Claude's first pass skipped hidden folders like .npm, about 25 GB); items only Claude found are usually places Mole doesn't know yet, such as 83 GB of CapCut (Jianying) cache in the Movies folder. He decides before anything is deleted, quits the app, deletes one folder at a time, and trusts only the free space df reports. Ten rounds freed about 315 GB. He also sent Mole two PRs for Jianying and Feishu caches, both merged. The article lists four kinds of cache to clear and six things that look like junk but must not be deleted. On Linux we tested Mole's Linux branch: running ./mo clean --dry-run from the repo listed about 102 MB of npm, Bun, pip and uv caches in just over a second, and deleted nothing. We did not test the macOS build or the Claude reconciliation. / Leaf（@leaf_sanren）的 500G MacBook 三个半月满了 10 次。他的做法是两边一起扫：先用 tw93 开源的命令行清理工具 Mole 跑 mo clean --dry-run 预览，再让 Claude 扫家目录，两份清单对账。两边都报的放心删；只有 Mole 报的，说明 Claude 漏了（Claude 第一次漏掉了 .npm 这类隐藏文件夹，少算约 25G）；只有 Claude 报的，多半是 Mole 还不认识的位置，比如剪映把 83G 缓存放在「影片」文件夹里。删之前由他本人拍板，退出相关软件、一次删一个目录，以 df 显示的剩余空间为准。10 次下来累计腾出约 315G。他还给 Mole 提了剪映和飞书缓存的两个 PR，都已合并。文章里列了必清的四类缓存和 6 个看着像垃圾但千万别删的东西。我们在 Linux 上测了 Mole 的 Linux 分支：直接从仓库运行 ./mo clean --dry-run，1 秒多列出 npm、Bun、pip、uv 缓存约 102MB 可清，没有删除任何文件。没有测 macOS 版本，也没有测和 Claude 对账的流程。</description></item><item><title>Chen Dahuang open-sourced his anti-AI-tone writing setup as a Skill: pick one voice card with real human examples per genre, and keep only 20–30 banned words as a guardrail / 陈大黄把自己去 AI 味的写作方法开源成 Skill：按题材挑一张带真人范文的音色卡，只留二三十个禁词兜底</title><link>https://ttqq.dev/en/tool/open-writing-skill-voice-cards-de-ai-tone/</link><guid>https://ttqq.dev/en/tool/open-writing-skill-voice-cards-de-ai-tone/</guid><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><description>A common anti-AI-tone fix is to paste a list of hundreds of banned words. Chen Dahuang (@realchendahuang) argues this backfires: with only prohibitions and no good writing to imitate, the model either swaps in a synonym or strips sentences down to a manual, and putting long-form and short-post rules in one prompt mixes their voices. His setup has three layers: one "voice card" per genre, with the positive style, genre-specific don'ts and two real human-written examples; read only the one card for this piece; and an outer list of just 20–30 of the worst words and patterns. Two hard rules sit on top: don't invent facts, and when editing a draft keep its structure. He stripped personal details and open-sourced it as open-writing-skill, with four cards (long articles, short posts and tool recommendations, community questions, lessons-learned notes) plus a single-file version for web chat; he suggests Gemini with thinking set to Low. We installed it into a Claude Code project folder with npx skills add and checked the entry file and the four cards; we did not write with it or compare before and after. / 很多人去 AI 味的做法是抄几百个禁用词给模型。陈大黄（@realchendahuang）认为这样效果拧巴：只下禁令、不给好文章看，模型要么换个同义词继续端着，要么把句子删成干巴巴的说明书；长文和短帖的要求塞在一个提示词里还会串味。他的做法分三层：每种题材一张「音色卡」，写清正向风格、这类题材专属的禁忌，再放两篇真人写的范文原文；动笔前只读当次题材对应的一张卡；最外层只留二三十个最扎眼的词和句式做兜底。另有两条硬规矩：不编事实，改稿保留原文的逻辑层次。他把这套去掉个人信息后开源成 open-writing-skill，自带深度长文、短评与工具推荐、互动提问、踩坑随笔四张卡，以及网页聊天框用的单文件版；他建议用 Gemini、思考等级调到 Low。我们用 npx skills add 把它装进一个 Claude Code 项目目录，检查了入口文件和四张卡的结构，没有用它实际写稿或做前后对比。</description></item><item><title>Huashu tests ByteDance Seedream 5.5: the person still looks like herself after 20 edits, and he open-sources the sprite-sheet → frames → GIF flow for transparent animated stickers as a skill / 花叔实测字节 Seedream 5.5：连改 20 轮人物还像本人，并把「多格图→切帧→GIF」做透明底动态表情包的流程开源成 skill</title><link>https://ttqq.dev/en/tool/huashu-selfie-sticker-seedream-animated-gif/</link><guid>https://ttqq.dev/en/tool/huashu-selfie-sticker-seedream-animated-gif/</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>ByteDance's Seedream 5.5 leads with multi-turn editing: it claims untouched pixels stay put and a single image can be edited a hundred times. Huashu (@AlchainHust) skipped the hundred-turn marathon as too credit-hungry and instead edited a café-window portrait in 10 places, then reverted them one by one, uploading only the previous output each turn: at turn 20 she was still herself. Small edits scored SSIM 0.983–0.995; a moon earring, a wrist mole and a badge stayed put, while the street outside drifted slightly. He also ran multi-turn edits on a living room, a shop poster and a product shot, checked a 706-character cruise magazine spread (including accented place names) character by character with no typos, and composited the woman with a trilingual coffee bag and a wooden mannequin pose. Transparency is a native PNG alpha, not a later cutout. After reverse-engineering GPT Image 2.6's “draw a sprite sheet, then slice frames” GIF trick, he shipped the open-source skill huashu-selfie-sticker (v1.1.0): lock a character sheet from a selfie or any image, draw a 4×2 or 3×3 sprite sheet, then a script slices, aligns and writes 480px and 240px transparent GIFs plus an APNG. On Jimeng with Seedream 5.5 (16:9, 2K, transparent background) he says 9 images landed in one pass and all 64 captions across 8 sets were typo-free; installing the skill in Doubao took about 30 seconds. We cloned v1.1.0 and ran make_gif.py on a homemade 4×2 transparent sprite sheet (it printed “already a transparent original,” cut 8 frames, and wrote the 480/240 GIFs and APNG). Laying the same sheet on magenta and rerunning with --bg #FF00FF, it reported keying out solid #FF00FF (81% of the image) and again cut 8 frames. The repo has no LICENSE file. We did not call Jimeng, Doubao or the Seedream API to generate a real sticker set. / 字节发布 Seedream 5.5，把多轮编辑放在更新第一条，声称没让改的地方能像素级不动、同一张图可连改上百轮。花叔（@AlchainHust）觉得上百轮太耗积分，就拿一张咖啡馆窗边人像先改 10 处再一处一处改回去，每轮只上传上一轮输出：第 20 轮她还是她。小改动的结构相似度（SSIM）在 0.983–0.995；月亮耳钉、手腕上的痣、徽章位置都还在，窗外街景有一点漂移。他又测了客厅装修、开业海报、商品图多轮改，一张 706 字航线杂志跨页（含特殊字母地名）逐字核对无错字，还把女生与三语文包装、木头人偶动作做了多图融合。透明底是模型直出的 PNG，不是事后抠的。拆完 GPT Image 2.6 的「先画多格再切帧」逻辑后，他写成开源 skill huashu-selfie-sticker（v1.1.0）：自拍或任意图先锁角色设定，再按比例画 4×2 / 3×3 精灵图，脚本切帧对齐，输出 480 与 240 的透明 GIF 和 APNG。他在即梦用 Seedream 5.5（16:9、2K、透明背景）一次出 9 张图，8 套共 64 格字幕无错字；在豆包里装 skill 大约半分钟。我们克隆了 v1.1.0，用自制的 4×2 透明精灵图成功运行了 make_gif.py（认出「原图已是透明底」，切出 8 帧并写出 480/240 GIF 与 APNG）；把同一张图铺到品红底再 --bg #FF00FF，脚本报「抠掉了纯色背景 #FF00FF（背景占 81%）」并同样切出 8 帧。仓库未见 LICENSE 文件。我们没有调用即梦/豆包/Seedream API 生成真实表情包。</description></item><item><title>Simon Willison had Claude Code build Beep Locator, turned phones and tablets into a microphone array, and found the beeping carbon-monoxide alarm with a low battery in his storage room to within 40 cm / Simon Willison 用 Claude Code 写了 Beep Locator，把手机和平板组成麦克风阵列，在 40 厘米范围内找到了储藏室里缺电报警的一氧化碳报警器</title><link>https://ttqq.dev/en/tool/simonw-beep-locator-claude-code/</link><guid>https://ttqq.dev/en/tool/simonw-beep-locator-claude-code/</guid><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><description>After a move, something in a box room started beeping at long intervals—too sparse to locate by ear, and Simon Willison (@simonw) did not want to unpack every box. On his phone he used Claude Code Cloud (Opus 5.5, Extra thinking) against an empty repo to write up the problem, then had it build a Python web app: one device opens a coordinator page, and phones, tablets and laptops join in the browser as listeners, sending audio to the server over WebAudio and WebSockets. Letting the devices locate each other by chirping was inaccurate, so he added a Rough mode for dragging icons into approximate positions, plus a clap test (a different map symbol) to check that localization worked at all. When a beep appeared he could replay each device’s clip and mark “this is the mystery beep,” after which the system tightened the estimate from frequency and interval. He reports it got within about 40 cm, and in that pile of boxes he found an old carbon-monoxide alarm that needed new batteries. Mid-project, Fable 5.2 shipped and he continued on his laptop, chatting with the model while the server ran; listener pages can also hot-swap JavaScript over the WebSocket so you don’t reload and re-grant the microphone. We cloned simonw/beep-locator 0.1.0 (449a488, Apache 2.0); uv run pytest passed all 26 tests in 2.63s (plus one Starlette/httpx deprecation warning). After uv run beep-locator serve on local port 8765, the coordinator page, /join and /sim each returned HTTP 200. We did not triangulate with real devices, and we did not start a Cloudflare tunnel. / 搬家后的储藏室里有东西间歇作响，间隔太长，站在屋里听方向很折磨，Simon Willison（@simonw）又不想把箱子逐个翻开。他先在手机上用 Claude Code Cloud（Opus 5.5，Extra thinking）对着空仓库写问题说明，再让它做一套 Python 网页应用：一台设备打开协调页，手机、平板、笔记本用浏览器加入当监听器，靠 WebAudio 和 WebSocket 把声音汇到服务器。设备互鸣来自动定位并不太准，于是他加了「粗略模式」可把图标拖到大概位置；又加了拍手测试（地图上用另一种符号），方便确认系统真能定位。哔声出现后可逐台回放录音、标记「这就是那声神秘哔」，系统再按频率和间隔收紧位置。按他的记录，最后把声源收在约 40 厘米内，在那一堆箱子里找到一只急需换电池的旧一氧化碳报警器。中途 Fable 5.2 发布，他改在笔记本上继续让模型对着正在跑的服务改功能；监听页还能通过 WebSocket 热更新 JavaScript，免得刷新页面重新要麦克风权限。我们克隆了 simonw/beep-locator 0.1.0（449a488，Apache 2.0），uv run pytest 26 项全部通过（2.63 秒，另有 1 条 Starlette/httpx 弃用警告）；uv run beep-locator serve 在本机 8765 端口启动后，协调页、/join、/sim 均返回 HTTP 200。没有用多台真实设备做麦克风三角定位，也没有开 Cloudflare 隧道。</description></item><item><title>Cloudflare's open-source security-audit skill keeps the hunter agent and the verifier agent separate, so the AI that wrote the code does not approve its own work / Cloudflare 开源的安全审计 Skill 把找漏洞的猎手 Agent 和验证 Agent 强制分开，避免同一个 AI 既写代码又给自己放行</title><link>https://ttqq.dev/en/tool/cloudflare-security-audit-skill/</link><guid>https://ttqq.dev/en/tool/cloudflare-security-audit-skill/</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Asking the same agent session that just shipped a feature to “check security once more” invites self-confirmation. Cloudflare's open-source security-audit skill (GitHub cloudflare/security-audit-skill, MIT) packages their vulnerability-discovery orchestration: six phases (recon → coverage-led hunting → candidate validation → structured output → independent record verification → target-neutral reporting), with hunter/verifier isolation and machine-checked findings.json / coverage-ledger.json via zero-dependency validators. Install with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit (optional --global). Chinese-community post @divid_lu49394 (9 Nov 2026) summarized usage and limits (one run covers only part of the surface, static source only, NEEDS-VALIDATION must be reproduced by a human). We cloned main @ c1c8a8c and ran the shipped tests under skills/security-audit/: validate-findings.test.cjs 34/34 pass, validate-coverage-ledger.test.cjs 31/31 pass; SKILL.md confirms guidance mode by default and full six-phase runs only when the user explicitly asks for a full audit. We did not run a live Claude/Codex audit against an app repo (expensive and sandbox-sensitive). / 用 Agent 写完功能就让同一个会话「再检查一遍安全」，很容易自我证实。Cloudflare 开源的 security-audit skill（GitHub cloudflare/security-audit-skill，MIT）把内部漏洞挖掘编排收成可安装 Skill：六阶段（侦察 → 覆盖率驱动狩猎 → 候选验证 → 结构化输出 → 独立记录复核 → 中立报告），核心规矩是查找与验证必须用隔离的 Agent，并写出 findings.json / coverage-ledger.json，用自带的零依赖校验脚本卡住格式。安装：npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit（可加 --global）。中文社区 @divid_lu49394 在 2026-11-09 介绍了用法与限制（单次大约只能覆盖一部分问题、只看静态源码、NEEDS-VALIDATION 必须人工复现）。我们克隆 main @ c1c8a8c，在 skills/security-audit/ 下跑官方测试：validate-findings.test.cjs 34/34 通过，validate-coverage-ledger.test.cjs 31/31 通过；阅读 SKILL.md 确认默认是 guidance 模式，只有明确要求全面审计时才走六阶段并写报告。没有用真实 Claude/Codex 对业务仓库跑完整审计（会消耗大量额度且需沙箱）。</description></item><item><title>Tencent's open-source TeamAI CLI syncs a team's skills, rules, agents and MCP config from one Git repo, so each person stops keeping their own set / 腾讯开源的 TeamAI CLI 用一个 Git 仓库给团队同步 skills、rules、agents 和 MCP 配置，不再每人各用一套</title><link>https://ttqq.dev/en/tool/tencent-teamai-cli-team-skills-sync/</link><guid>https://ttqq.dev/en/tool/tencent-teamai-cli-team-skills-sync/</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>When every teammate keeps their own Claude Code / Codex / Cursor skills, rules and MCP servers, configs drift fast. Tencent's open-source TeamAI CLI (npm package teamai-cli, MIT) puts those AI assets in a shared Git repo: after one teamai init, later sessions automatically pull the latest skills, rules, agents, hooks, MCP and workflows. You can also install and share through the /teamai skill inside an agent. Chinese-community account TryAI Lab (@TryAiLabHQ) ran init → pull → push locally on 10 Nov 2026 with v0.29.0: after pull, .claude/ held 1 skill, 7 rules, 76 docs, 3 agents and 21 workflows, and a hand-written my-review skill pushed back into the team repo. Two real pitfalls: the first push right after init fails until you commit the scaffold files TeamAI generated; pull also printed twenty-plus workflow schema warnings where the team's own main-branch workflows failed validation against CLI 0.29.0. We installed teamai-cli@0.29.0 from npm (teamai --version matched), confirmed doctor asks for init when uninitialized, and saw init against teamai-hub/teamai-template stop with “GitHub authentication unavailable” in a non-interactive shell without GITHUB_TOKEN. We also cloned the public template and counted 11 SKILL.md files, 11 rules under rules/common, and 3 review agents. We did not finish an authenticated pull/push, and did not open a Claude Code session to verify the auto-sync hook. / 团队里每个人自己攒 Claude Code / Codex / Cursor 的 skills、rules 和 MCP，过几天就各玩各的。腾讯开源的 TeamAI CLI（npm 包 teamai-cli，MIT）把这些 AI 资产放进一个共享 Git 仓库：成员 teamai init 一次之后，开会话时自动 pull 最新 skills / rules / agents / hooks / MCP / workflows。也可以用 /teamai skill 在 Agent 里装和分享。中文社区账号 TryAI Lab（@TryAiLabHQ）在 2026-11-10 用 v0.29.0 在本地完整走了一遍 init → pull → push：pull 后落到 .claude/ 的有 1 个 skill、7 条 rules、76 份 docs、3 个 agents、21 个 workflows；自己写的 my-review skill 也能 push 回团队仓。他踩到两个真坑——刚 init 完第一次 push 会因脚手架未提交被拦；pull 时自家 main 上的 workflow 用 0.29.0 校验刷出二十多条「字段不认识」警告。我们装了 npm 上的 teamai-cli@0.29.0，teamai --version 正确；doctor 在未初始化时会提示要跑 init；在无交互、无 GITHUB_TOKEN 的环境里对 teamai-hub/teamai-template 做 init 会停在「GitHub authentication unavailable」。另克隆了公开模板仓核对结构（11 个 SKILL.md、rules/common 下 11 条规则、3 个 review agents）。没有完成带登录的真实 pull/push，也没有开 Claude Code 会话验证自动同步钩子。</description></item><item><title>Check before you install a skill or MCP: PeepSick's USB+ uses secscan to compare a skill's claims with its body, and mcprobe to watch MCP servers without calling their tools / 装 Skill 或 MCP 前先检查：PeepSick 的 USB+ 用 secscan 对照声明和正文，用 mcprobe 只观察 MCP 服务器、不调用工具</title><link>https://ttqq.dev/en/tool/usbplus-secscan-skill-mcp-trust-layer/</link><guid>https://ttqq.dev/en/tool/usbplus-secscan-skill-mcp-trust-layer/</guid><pubDate>Tue, 10 Nov 2026 00:00:00 GMT</pubDate><description>Skills and MCP servers are copy-pasted everywhere, and it is hard to know before install whether they read .env, shell out, or delete files. PeepSick Labs' USB+ (Universal Systems Bridge, GitHub PeepSick/usbplus, npm @peepsick/usb-cli, MIT) turns “inspect then install” into a CLI: usb secscan statically compares a skill's declared contract with capability signals in the body; usb mcprobe, with explicit consent, starts MCP servers for a handshake-only observation and does not call tools by default; it can also install from a 529-skill catalog into 16 runtimes. Chinese-community posts by @wlzh (and @MirrortekUK) covered the v1.0.8 trust layer on 9 Nov 2026. We installed @peepsick/usb-cli@1.0.8: usb version reported current/latest 1.0.8 and catalog “Universal Systems Bridge Catalog”; usb search postgres returned 16 hits; usb list listed 529 skills; secscan on two local demo skills flagged a sneaky skill whose friendly frontmatter hid .env reads and rm -rf (HIGH secrets-access, MEDIUM destructive-commands, exit 1); secscan --repo cloudflare/security-audit-skill also flagged an fs.rmSync signal in tests (MEDIUM). We did not run usb install for the full 529-skill pack, and did not run mcprobe --yes against a live MCP server. / Skill 和 MCP 服务器现在到处复制粘贴，装之前很难知道它会不会读 .env、调 shell 或删文件。PeepSick Labs 的 USB+（Universal Systems Bridge，GitHub PeepSick/usbplus，npm @peepsick/usb-cli，MIT）把「先验货再安装」做成 CLI：usb secscan 静态比对 skill 声明与正文里的能力信号；usb mcprobe 在征得同意后启动 MCP 做协议握手观察、默认不调用工具；另外还能在 16 种运行时之间安装目录里的 529 个标准化 skill。中文社区 @wlzh（以及 @MirrortekUK）在 2026-11-09 介绍了 v1.0.8 的信任层。我们安装 @peepsick/usb-cli@1.0.8：usb version 回报 current/latest 均为 1.0.8、目录名 Universal Systems Bridge Catalog；usb search postgres 返回 16 条；usb list 列出 529 条；对两个本地演示 skill 跑 secscan——声明友好却正文读 .env / rm -rf 的 sneaky-skill 被标 HIGH secrets-access 与 MEDIUM destructive-commands，exit 1；对 cloudflare/security-audit-skill 远程仓 secscan 也抓到测试文件里的 fs.rmSync 信号（MEDIUM）。未执行 usb install 全量 529、未对真实 MCP 跑 mcprobe --yes。</description></item><item><title>Khazix's Agent Autoloop dispatches CLI agents in the background, exits when they are waiting, and only sends a Feishu or desktop card when it needs your decision / 卡兹克的 Agent Autoloop 在后台给 CLI Agent 派活，干完就退出，需要你拍板时才往飞书或桌面发一张卡片</title><link>https://ttqq.dev/en/tool/agent-autoloop-background-cli-agents/</link><guid>https://ttqq.dev/en/tool/agent-autoloop-background-cli-agents/</guid><pubDate>Mon, 09 Nov 2026 00:00:00 GMT</pubDate><description>With Claude Code or Codex, work usually stops the moment you close the terminal or walk away. Chinese-community author 数字生命卡兹克 / Digital Life Khazix (@Khazix0918, GitHub KKKKhazix) open-sourced the dispatcher he had used for a year as Agent Autoloop (MIT): a local, zero-runtime-dependency kernel that reads rules.json and hands cron jobs, GitHub issues, alerts and other events to a CLI agent you already signed into. The agent runs one round and exits. Deleting data, rebooting, spending money or messaging the outside world becomes a card on Feishu, DingTalk, WeCom, Slack, Discord or the desktop; while you decide, nothing sits around burning tokens, and answering resumes the same session. The kernel claims to do only four things—ingest events, route them, dispatch an agent, and record what happened—with zero-cost suspend, event sourcing, separating instructions from untrusted input, and a hop circuit breaker at 8 layers. Events, agents, channels and delivery are plugins. Bundled recipes cover a morning brief, GitHub issues, alerts, website watch, Feishu weekly notes, inbox triage, feedback and PR delivery. Install by handing the repo to your agent (it is told to read INSTALL.md), or run curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh on macOS/Linux (irm .../install.ps1 | iex on Windows). Node 24+ is required. Windows support landed on 6 November. We cloned main on Linux (fade2ce, after the v1.6.4 release) and ran npm run test:quick with Node 24.11.0: 697 of 719 tests passed, 0 failed, 22 skipped because this machine is not Windows/macOS or has no Claude Code. We did not run the installer, agent-autoloop try against a live Claude/Codex login, or Feishu. / 平时用 Claude Code、Codex 这类命令行 Agent，人一走开或关掉窗口，活就断了。中文社区作者数字生命卡兹克（@Khazix0918，GitHub KKKKhazix）把折腾一年的调度中台开源为 Agent Autoloop（MIT）：本机跑一个零运行时依赖的内核，按 rules.json 把定时任务、GitHub Issue、告警等事件派给已登录的 CLI Agent；Agent 干完一轮就退出，碰到删数据、重启、花钱、对外发消息才通过飞书、钉钉、企微、Slack、Discord 或桌面通知发卡片，等你回答期间不占进程、不烧 Token，点完接上原会话继续。内核自称只做四件事：收事件、按规则分活、派给 Agent、记下来；挂起零消耗、事件溯源、指令与资料分离、超过 8 层跳数熔断。事件、Agent、通道和代码交付都是插件。内置早报、跟进 Issue、告警排查、网站巡检、飞书周报、收件箱分诊、反馈处理和 PR 交付等模板。安装：把仓库交给你的 Agent 读 INSTALL.md，或在 macOS/Linux 执行 curl -fsSL https://raw.githubusercontent.com/KKKKhazix/agent-autoloop/main/install.sh | sh（Windows 用 irm .../install.ps1 | iex），需要 Node 24+。作者 11 月 6 日补上 Windows。我们在 Linux 上克隆 main（fade2ce，晚于已发布的 v1.6.4），用 Node 24.11.0 跑 npm run test:quick：719 项里 697 通过、0 失败、22 项因本机不是 Windows/macOS 或未装 Claude Code 而跳过；没有执行安装脚本，也没有用真实 Claude/Codex 跑 agent-autoloop try，更没有接飞书。</description></item><item><title>Waffle's clear-naming skill reads the surrounding code first, then fixes names and explanations an agent translated too literally, such as turning in_flight into a word-for-word Chinese calque / Waffle 的 clear-naming 技能先读代码上下文，再修改 Agent 硬译出来的命名和说明，比如把 in_flight 译成「在飞任务」这类问题</title><link>https://ttqq.dev/en/tool/clear-naming-agent-jargon-code-names/</link><guid>https://ttqq.dev/en/tool/clear-naming-agent-jargon-code-names/</guid><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><description>When an agent writes or explains code in Chinese, you often get calques like 'wait for flying tasks to drain' for in_flight, and names that don't say whether something is a running task or a request awaiting a response. Chinese-community author Waffle (@Jiaoao1024, GitHub ox01024) open-sourced clear-naming (MIT): an instruction-only skill that reads definitions, call sites and state changes before it decides whether to rename and how. The same in_flight might become running_tasks if the set is already started and not finished, or requests_awaiting_response if those are outstanding HTTP calls; established terms such as HTTP, SQL and transactions can stay. Four modes: clarify this explanation only, review without editing, rename in a stated scope and update references, or merge two short principles into AGENTS.md. Install: npx skills add ox01024/clear-naming. We read SKILL.md and examples.md and checked that the tweet's examples appear in the docs; there is no runtime script to run, and we did not ask an agent to rename a real repo. / 让 Agent 写代码或解释改动时，中文说明常变成「等在飞任务排空后再退出」「provider 抖了重规划没走通」这种机翻黑话，变量名也不分得清是正在执行的任务还是等待响应的请求。中文社区作者 Waffle（@Jiaoao1024，GitHub ox01024）开源 clear-naming（MIT）：纯指令 skill，先读定义、调用和状态变化，再决定要不要改、改成什么。同一个 in_flight，已开始且未结束的任务可考虑 running_tasks，已发出、等待响应的请求则更接近 requests_awaiting_response；HTTP、SQL、事务等有明确含义的术语可以保留。用法分四档：只把当前解释说清楚、先审查不改文件、在指定范围改名并更新引用、把两条原则合并进 AGENTS.md。安装：npx skills add ox01024/clear-naming。我们阅读了 SKILL.md 与 examples.md，核对原帖例句都在文档里；无运行时脚本可跑，也未让 Agent 对真实仓库做重命名。</description></item><item><title>SinFire splits clarifying a task into three steps: align terms, research facts in parallel, then settle decisions in dependency order, so flagship models stop grilling you on trivia / SinFire 把需求澄清拆成三步：先对齐词义，再并行查事实，最后按依赖顺序拍板，避免强推理模型用 grill-me 问个不停</title><link>https://ttqq.dev/en/tool/sinfire-skills-clarify-decisions-before-plan/</link><guid>https://ttqq.dev/en/tool/sinfire-skills-clarify-decisions-before-plan/</guid><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><description>The popular grill-me skill (Matt Pocock) makes an agent interview you before it plans; on today's flagship reasoning models the original 'relentlessly' often means a dozen nitpicky questions. Chinese-community author SinFire (@Sin_Yiyao) reworked grill-me plus his own requirements-clarification practice into sinfire-skills (MIT, v0.1.0): three instruction-only skills, 119 lines of Markdown, no runtime deps. $semantic-clarification (59 lines) only asks what you mean, and will not expand into requirements gathering. $facts-searching (26 lines) dispatches parallel sub-agents to search the web, dedupes, resolves conflicts, and brings back only sourced findings that affect the current decision—unverifiable items stay marked as assumptions. $clarify-decisions (34 lines) stays at the current level, defaults low-impact reversible details, asks only the most consequential frontier questions each round, and recommends an answer only when evidence and stated aims support one. Install: npx skills add SinVC/sinfire-skills --skill clarify-decisions semantic-clarification facts-searching -a codex. We cloned the repo and bun run validate passed for all three skills plus the Codex explicit-invocation policy; we did not run a live Codex/Claude session through the full loop. / 社区热门的 grill-me（Matt Pocock）会让 Agent 动手前先按决策树拷问你；换上现在的强推理旗舰后，原版那句 relentlessly 常让它抓着细枝末节连问十几道。中文社区作者 SinFire（@Sin_Yiyao）把原版和自己在软件工程里打磨的需求澄清拆开重做，开源 sinfire-skills（MIT，v0.1.0）：三个纯指令 skill，合计 119 行 Markdown，0 运行时依赖。$semantic-clarification（59 行）只问「你到底是什么意思」，不擅自扩成需求分析；$facts-searching（26 行）派多个子 Agent 并行联网查证，去重、核对冲突后只带回和当前决定相关、有出处的结论，查不到的标明假设；$clarify-decisions（34 行）守住当前层级、可逆小事给默认值、每轮只问最影响全局的几道，有依据才给推荐答案。安装：npx skills add SinVC/sinfire-skills --skill clarify-decisions semantic-clarification facts-searching -a codex。我们克隆仓库跑 bun run validate，三个 skill 与 Codex「仅明确调用」策略均通过；未开真实 Codex/Claude 会话走完整个流程。</description></item><item><title>sanyam's verso redraws one piece of an existing UI as a Leonardo notebook page, changing only the look and keeping the DOM, handlers and focus order / sanyam 的 verso 把一块现成 UI 画成达芬奇手稿的样子，只改外观，DOM、事件和焦点顺序都不动</title><link>https://ttqq.dev/en/tool/verso-leonardo-notebook-ui-skill/</link><guid>https://ttqq.dev/en/tool/verso-leonardo-notebook-ui-skill/</guid><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate><description>Overseas author sanyam (@sanyampunia) open-sourced verso, a skill for Claude Code / Codex / Cursor that redraws one named slice of your UI as a page from Leonardo's working notebooks: laid rag paper, iron gall ink faded to browns, hatching instead of shadows, hand rules, margin notes and pen sketches. Point it at a component, route or selector—/verso the settings card, or /verso app/blog/page.tsx in the forster style. It reads the project's existing Tailwind / CSS Modules / plain CSS and CLAUDE.md first, then converts using one of four papers: Windsor (forms/settings, default), Leicester (articles/release notes), Arundel (dashboards) and Forster (narrow cards/sidebars). Mirror script is decoration only and gets aria-hidden; body text is claimed at 10.3:1 or better, and every line of real text above 4.5:1. Install: npx skills add sanyampunia/verso. He later posted nine before-and-after conversions. We inspected the bundled examples and the ~30KB manuscript.css: after.html is a default Windsor 'Project settings' card (inputs, buttons, margin notes, mirror script); release-after.html uses the Leicester paper. The repo has no LICENSE file. We did not run a /verso session on a live React/Tailwind project. / 海外作者 sanyam（@sanyampunia）开源 verso：给 Claude Code / Codex / Cursor 用的 skill，把你指定的那一块界面画成达芬奇工作手稿——布纹纸、氧化成棕色的铁胆墨水、用排线代替阴影、手绘格线、页边注释和钢笔速写。你指一个组件、路由或选择器即可，例如 /verso the settings card，或 /verso app/blog/page.tsx in the forster style。它先读项目现有的 Tailwind / CSS Modules / 普通 CSS 和 CLAUDE.md，再按 Windsor（表单/设置，默认）、Leicester（文章/发布说明）、Arundel（看板）、Forster（窄卡片/侧栏）四张纸来转换；镜像字只做装饰并加 aria-hidden，正文对比度声称至少 10.3:1、所有真实文字超过 4.5:1。安装：npx skills add sanyampunia/verso。作者随后又发了九组转换前后对照。我们检查了仓库里的示例页和约 30KB 的 manuscript.css：after.html 是默认 Windsor 的「Project settings」卡片（含输入框、按钮、页边注释和镜像字），release-after.html 使用 Leicester 纸；仓库没有 LICENSE 文件。未在真实 React/Tailwind 项目里跑 /verso 会话。</description></item><item><title>@Feng_YQ's BeRich skill researches how indie products make money, with 1,169 site dossiers and two research methods: Timeline and Causal Chain / @Feng_YQ 的 BeRich（发财 skill）用来调研独立站怎么赚钱，自带 1169 个站点档案和「发财时间线」「因果链」两套调研方法</title><link>https://ttqq.dev/en/tool/berich-indie-product-growth-research-skill/</link><guid>https://ttqq.dev/en/tool/berich-indie-product-growth-research-skill/</guid><pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate><description>Competitive/indie research with agents often invents narratives. Chinese author Max Feng (@Feng_YQ / fengyiqicoder) open-sourced BeRich: 1,169 public dossiers from 闷声发大财 (msfdcszhd.com) plus two methods—Growth Timeline (month-by-month action→metric) and Causal Chain (every step to revenue needs evidence, action before effect, ≥3 peer cases for “it worked”). One-liner for Claude Code / Codex: `curl -fsSL https://raw.githubusercontent.com/fengyiqicoder/BeRich/main/install.sh | bash`. We verified 榜单.csv and data/sites both at 1,169 with 0 missing archives; install.sh symlinked into a temp HOME's .claude/.codex/.agents skills; spot-checked Postiz examples. No live “research this domain” agent session; no 发大财 Pro purchase. / 做竞品/出海调研时，Agent 常编故事缺证据。中文作者 Max Feng（@Feng_YQ / fengyiqicoder）开源 BeRich（发财 skill）：打包闷声发大财（msfdcszhd.com）公开的 1169 个赚钱独立站档案，外加两套方法——发财时间线（逐月营销动作→数字）与发财因果链（动作→收入每环要证据、动作须早于后果、有效手段至少 3 个同类案例）。一条命令装到 Claude Code / Codex：`curl -fsSL https://raw.githubusercontent.com/fengyiqicoder/BeRich/main/install.sh | bash`。我们核对 data/榜单.csv 与 data/sites 均为 1169、档案链接 0 缺失；install.sh 在临时 HOME 链到 .claude/.codex/.agents/skills；并抽查 Postiz 等 examples。未联网跑完整「调研某域名」Agent 会话，未购买发大财 Pro。</description></item><item><title>TokenSave's 288-run study: asking for subagent search left recall almost unchanged and raised the Opus bill by up to 221% / TokenSave 的 288 次对照：写「用子代理搜索」时召回几乎不变，Opus 账单最多多出 221%</title><link>https://ttqq.dev/en/tool/claude-code-subagent-search-cost-288-runs/</link><guid>https://ttqq.dev/en/tool/claude-code-subagent-search-cost-288-runs/</guid><pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate><description>Claude Code 2.1.318 added subagent effort (`CLAUDE_CODE_SUBAGENT_EFFORT`). TokenSave (dataset under johnisking; Japanese write-up by @Yunany90) ran 288 pre-registered headless trials on 12 “find every call site” questions across 6 OSS repos, Haiku 5.5 / Opus 5.5, no-instruction vs forced Explore subagents at low/medium/high. Result: recall stayed ~97–100% (&lt;2 pt gaps); cost rose Haiku +74–116%, Opus +143–221%; with no instruction, 0/72 runs called a subagent. We recomputed the CSV: 288 rows, balanced cells, Opus C2/C0 mean cost ≈3.21× (+221%), C0 subagent_calls sum 0. We did not re-bill a live Claude Code session. / Claude Code 2.1.318 加了子代理 effort（`CLAUDE_CODE_SUBAGENT_EFFORT`）。TokenSave（数据集 johnisking；日文撰稿 @Yunany90）在 6 个开源仓库、12 道「找全调用点」题上，对 Haiku 5.5 / Opus 5.5 做了 288 次预注册 headless 对照（无指示 vs 强制 Explore 子代理 × low/medium/high）。结论：召回全在约 97%–100%，差不到 2 个点；成本 Haiku +74%～+116%，Opus +143%～+221%；无指示时 72 次里主动调子代理 0 次。我们复算 CSV：288 行、条件均衡，Opus C2/C0 均费约 3.21×（+221%），C0 子代理调用合计 0。未重跑 Claude Code 计费会话。</description></item><item><title>Niaz's reanimate records the original animation, then your clone, and compares them frame by frame until the timing is within the default ~80ms / Niaz 的 reanimate 先录下原版动效，再录你的复刻版，逐帧比对，直到时间差在默认的 80ms 以内才算完成</title><link>https://ttqq.dev/en/tool/reanimate-claude-skill-frame-animation-clone/</link><guid>https://ttqq.dev/en/tool/reanimate-claude-skill-frame-animation-clone/</guid><pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate><description>Ask an agent to copy a smooth site animation and it usually makes one pass, calls it done, and misses timing. Overseas author Niaz Morshed (@niazmorshed_) open-sourced reanimate (MIT, ~1.3k★): a Claude Code skill that records a reference from a live URL, screen recording, mp4/gif/Lottie, rebuilds with whatever CSS / Framer Motion / GSAP / three.js your project already uses, then films the clone the same way and compares frame by frame. Default is motion-faithful (±80ms); ask for pixel-faithful (≤2% mean mismatch). Install: `npx skills add NiazMorshed2007/reanimate -g`. On our Linux box we bun-installed deps, installed Playwright Chromium, and ran capture.ts on a local CSS-animated page: 43 frames, detected CSSAnimation rise (1200ms), 91 track samples; self-compare 0% mismatch PASS. No Claude Code session cloning a real marketing site, no full 2–5 round loop. / 看到别人站上的顺滑动效，丢给 Agent 往往做一遍就喊好，节奏和缓动全不对。海外作者 Niaz Morshed（@niazmorshed_）开源 reanimate（MIT，GitHub 约 1.3k★）：Claude Code Skill，可从网页链接、录屏、mp4/gif/Lottie 录参考，再按项目现有 CSS / Framer Motion / GSAP / three.js 写出复刻，并用同一套脚本录复刻版做逐帧对比；默认「动起来一样」（±80ms），要「长得一样」再开像素级（均值像素差 ≤2%）。安装：`npx skills add NiazMorshed2007/reanimate -g`。我们在 Linux 盒子上 bun install 依赖、装 Playwright Chromium，对本地 CSS 动画页跑 capture.ts：43 帧、检出 CSSAnimation rise（1200ms）、track 91 点；compare 自比对 0% mismatch PASS。未开 Claude Code 会话去复刻真实官网，未跑完整 2–5 轮迭代。</description></item><item><title>Connect OpenAI image generation to Grok Bot / Claude / Cursor with @gofelixpy's openai-image MCP: stills and GIFs, about two minutes of pasting config / 用 @gofelixpy 的 openai-image MCP 给 Grok Bot / Claude / Cursor 接上 OpenAI 生图，支持静图和 GIF，粘贴配置约两分钟</title><link>https://ttqq.dev/en/tool/felixpy-openai-image-mcp/</link><guid>https://ttqq.dev/en/tool/felixpy-openai-image-mcp/</guid><pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate><description>ChatGPT Pro/Business users get monthly API credits that cover the Images API. Overseas engineer @gofelixpy open-sourced @felixpy/openai-image (MIT, stdio MCP): tools `openai_image_generate` / `openai_image_edit`, writes absolute-path files, never returns base64; default model gpt-image-2.6-flare, GIF supported. Author says ~2 minutes to attach the MCP to Grok Bot so every agent can use it; sample meme GIF ~$0.06. On the box, npm ci + npm test: smoke 2/2 passed (shebang; after MCP initialize, tools/list exposes exactly generate and edit). No OPENAI_API_KEY set, no real image generated. / ChatGPT Pro/Business 用户有每月 API 额度可覆盖 Images API。海外工程师 @gofelixpy 开源 @felixpy/openai-image（MIT，stdio MCP）：工具 `openai_image_generate` / `openai_image_edit`，写出绝对路径文件，结果不含 base64；默认模型 gpt-image-2.6-flare，支持 GIF。作者称约两分钟把 MCP 挂到 Grok Bot 后全账号 Agent 都能用，样例 meme GIF 约 $0.06。我们在盒子上 npm ci + npm test：smoke 2/2 通过（shebang；MCP initialize 后 tools/list 正好暴露 generate 与 edit）。未设置 OPENAI_API_KEY，未真实出图。</description></item><item><title>@jinchenma_ai open-sources 11 skills for an agent-maintained personal knowledge base: originals protected, sources traceable, Feishu recordings optionally synced / @jinchenma_ai 开源 11 个 Skill，给 AI Agent 搭个人知识库：保护原件、追溯来源，可选同步飞书录音</title><link>https://ttqq.dev/en/tool/jinchenma-knowledge-base-agent-skills/</link><guid>https://ttqq.dev/en/tool/jinchenma-knowledge-base-agent-skills/</guid><pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate><description>The biggest trap building a knowledge base is copying someone else's heavy stack. Chinese author @jinchenma_ai open-sourced jinchenma-knowledge-base: rules in AGENTS.md, 11 skills (capture, media-ingest, meeting-ingestion, ai-dialogue-ingest, knowledge-analysis, query, enrich, entity-propagation, …) for notes, bookmarks, meetings, full AI chats, candidate insights (human confirm), query and entity links; optional Feishu recording sync and ASR. License CC BY-NC 4.0 (non-commercial). On the box with pytest/PyYAML, `scripts/run_tests.py`: 146 passed + 8 subtests; `check_project.py` reported 11 skills and no broken links; `setup.py --mode create` built a full vault template in a temp dir. No real Feishu account or end-to-end ingest. / 搭知识库最大的弯路是照搬别人的重架构。中文作者 @jinchenma_ai（金尘马）把自用底座开源成 jinchenma-knowledge-base：规则在 AGENTS.md，11 个 Skill（capture / media-ingest / meeting-ingestion / ai-dialogue-ingest / knowledge-analysis / query / enrich / entity-propagation 等）负责记想法、收藏外链、入库会议、存完整 AI 对话、提炼候选观点（需人确认）、查询与实体关联；可选飞书录音同步与 ASR。许可证 CC BY-NC 4.0（非商业）。我们在盒子上装 pytest/PyYAML 后跑 `scripts/run_tests.py`：146 passed + 8 subtests；`check_project.py` 报 11 个 Skill、无断链；并用 `setup.py --mode create` 在临时目录建出完整 vault 模板。未接真实飞书账号、未跑端到端采集。</description></item><item><title>vercel-labs/agent-desktop lets agents read and operate the desktop through the OS accessibility tree instead of screenshot pixels (@SimonObserves' hands-on notes) / vercel-labs/agent-desktop 用系统无障碍树给 Agent 读取和操作桌面，不靠截图像素（@SimonObserves 的实测笔记）</title><link>https://ttqq.dev/en/tool/vercel-agent-desktop-a11y-computer-use/</link><guid>https://ttqq.dev/en/tool/vercel-agent-desktop-a11y-computer-use/</guid><pubDate>Fri, 06 Nov 2026 00:00:00 GMT</pubDate><description>Agents usually drive a desktop by staring at screenshot pixels, or by reading the accessibility tree screen readers already use. Vercel Labs' agent-desktop (Apache-2.0, npm 1.2.0) is a model-free CLI/C library: `snapshot` returns JSON with @refs, then `click`/`type`; headless by default so it won't steal focus. Overseas writer @SimonObserves documented a Linux run: almost nothing works without AT-SPI; with it, listing windows and menu clicks work, and terminal typing needs `--headed`. On our headless Linux box after `npm install -g agent-desktop`, `version` was 1.2.0 and `status` reported accessibility=denied (no D-Bus session) with empty supported_surfaces—matching that write-up. We did not wire Claude/Codex MCP to drive a real desktop. / Agent 操控电脑常见两条路：看截图像素，或读屏幕阅读器已经在用的无障碍树。Vercel Labs 开源 agent-desktop（Apache-2.0，npm 1.2.0）：CLI/C 库，不含模型；`snapshot` 给出带 @ref 的 JSON，再 `click`/`type`，默认 headless 不抢焦点。海外作者 @SimonObserves 写了 Linux 实测：无 AT-SPI 时几乎全拒，装上后能列窗、点菜单，终端输入需 `--headed`。我们在无桌面会话的 Linux 盒子上 `npm install -g agent-desktop`：`version` 报 1.2.0；`status` 里 accessibility=denied（无 D-Bus session），supported_surfaces=[]——与笔记一致。未接 Claude/Codex MCP 真控桌面。</description></item><item><title>@GaoXiaoxiao2022 demonstrates the open-source android-ad-cleaner: with a USB cable and AGENTS.md, Codex / Claude turn off Android OEM ad switches over ADB / @GaoXiaoxiao2022 示范开源 android-ad-cleaner：用一根数据线和 AGENTS.md，让 Codex / Claude 通过 ADB 关掉国产安卓的广告开关</title><link>https://ttqq.dev/en/tool/android-ad-cleaner-ai-adb/</link><guid>https://ttqq.dev/en/tool/android-ad-cleaner-ai-adb/</guid><pubDate>Wed, 04 Nov 2026 00:00:00 GMT</pubDate><description>Chinese Android skins bury ad, trending-search and feed toggles. Open-source android-ad-cleaner (MIT, ~6.0k★, author rddmwmddr) needs no root or unlock: ADB flips built-in switches, can suspend ad packages (with consent), ships a desktop client and 35+ device recipes. AGENTS.md tells command-capable agents (Codex, Claude Code, WorkBuddy…) the hard rules and standard flow; a separate prompt guides chat-only AIs step by step. @GaoXiaoxiao2022 posted the clone → read AGENTS.md → ask before each change workflow. We read AGENTS.md/README/CHANGELOG and ran device-free unit tests (channel classification, never-suspend, entrypoint refs); this box had no phone or adbutils, so we did not run a recipe on hardware. / 国产安卓把广告、热搜、资讯流开关藏得很深。开源项目 android-ad-cleaner（MIT，GitHub 约 6.0k★，作者 rddmwmddr）免 Root、免刷机：用 ADB 只拨系统/App 自带开关，可挂起广告包（需同意），带桌面客户端与 35+ 机型脚本。仓库 AGENTS.md 专给能跑命令的 AI 智能体（Codex、Claude Code、WorkBuddy 等）写清硬规则与标准流程；另有给纯聊天 AI 的逐步提示词。@GaoXiaoxiao2022 在 X 上写了「克隆仓库 → 读 AGENTS.md → 逐步征求同意」的实操口吻。我们读了 AGENTS.md/README/CHANGELOG，跑了不依赖真机的单测（渠道分类、never-suspend、入口引用等）；本盒子无安卓设备与 adbutils，未对真机执行 recipe。</description></item><item><title>@komereji12 shares how to batch-process local images in Claude Code with jezweb design-assets' image-processing skill: resize, convert to WebP, trim white edges / @komereji12 分享用 jezweb design-assets 的 image-processing 技能在 Claude Code 里批量处理本地图片：缩放、转 WebP、去白边</title><link>https://ttqq.dev/en/tool/jezweb-image-processing-claude-skill/</link><guid>https://ttqq.dev/en/tool/jezweb-image-processing-claude-skill/</guid><pubDate>Wed, 04 Nov 2026 00:00:00 GMT</pubDate><description>Xiaohongshu posts and web work often need one width, WebP, and logo whitespace trimmed. Overseas repo jezweb/claude-skills (MIT, ~1.1k★) ships design-assets → image-processing: it doesn't generate images, only processes local files; `bin/img-process` (Pillow) covers resize/convert/trim/thumbnail/optimise/og-card/batch. Chinese poster @komereji12 shared the install lines and natural-language triggers. We cloned the repo and ran img-process with Pillow 13: JPG→WebP ~47KB→5KB; transparent logo trim 400×300→261×201; opaque white-canvas "whitespace" did not shrink—trim needs an alpha channel, not visual white. We did not run `/plugin install` inside a live Claude Code session. / 发小红书或做网页时常要统一宽度、转 WebP、裁掉 Logo 白边。海外仓库 jezweb/claude-skills（MIT，约 1.1k★）里的 design-assets 插件带 image-processing 技能：不生成图，只处理本地文件；附带 `bin/img-process`（Pillow）支持 resize/convert/trim/thumbnail/optimise/og-card/batch。中文作者 @komereji12 写了安装句与口语触发示例。我们克隆仓库，用 Pillow 13 直接跑 img-process：JPG→WebP 约 47KB→5KB；透明底 Logo trim 400×300→261×201；不透明白底「白边」图 trim 后尺寸不变——说明 trim 依赖透明通道，不是视觉上的白色。未在 Claude Code 会话里走 /plugin 安装。</description></item><item><title>Writing X Articles in Markdown: @riba2534's Chrome extension and xap bridge turn Markdown into an X Article draft in one click (a Claude Code Skill is included) / 用 Markdown 写 X 长文：@riba2534 的 Chrome 插件配合 xap 桥接，一键排成 X Article 草稿（附 Claude Code Skill）</title><link>https://ttqq.dev/en/tool/markdown-to-x-article-chrome-skill/</link><guid>https://ttqq.dev/en/tool/markdown-to-x-article-chrome-skill/</guid><pubDate>Wed, 04 Nov 2026 00:00:00 GMT</pubDate><description>X's Article editor doesn't keep pasted Markdown—code, tables and math have to be inserted by hand. Chinese author @riba2534 open-sourced Markdown to X Article (MIT, Chrome MV3): write Markdown in a side panel, then write native X headings/lists/code/tables/LaTeX into a draft; Mermaid is rendered locally and uploaded. It only creates drafts—never publishes. A local xap bridge (127.0.0.1:17345) and Claude Code Skill let an agent preflight and create drafts. Needs Chrome 116+, an X login, and X Premium for Articles. On the box after npm ci, vitest: 368/371 passed, 2 skipped; the "200 KB convert under 300ms" case failed at ~757ms here. We did not load the extension or write a real X draft. / X Article 编辑器不认粘贴进来的 Markdown，代码块、表格、公式要一个个从「插入」菜单加。中文作者 @riba2534 开源了 Markdown to X Article（MIT，Chrome MV3）：侧边栏写 Markdown，一键写入 X 原生标题/列表/代码块/表格/LaTeX，Mermaid 本地渲染成图再上传；只创建草稿、从不自动发布。另附本地 xap 桥（127.0.0.1:17345）和 Claude Code Skill，Agent 可预检、建草稿。需要 Chrome 116+、已登录 X，且 X Article 要 Premium。我们在盒子上 npm ci 后跑 vitest：371 项里 368 过、2 跳过，1 项「200 KB 文档转换 &lt;300ms」在本机约 757ms 超时失败；没有装 Chrome 扩展、也没连真 X 账号写草稿。</description></item><item><title>UI designer @double_asa's open-source DSH to Figma lets AI edit a Figma file from one sentence: DeepSeek Harness changes native layers directly and does not use Figma MCP quota / UI 设计师 @double_asa 开源的 DSH to Figma 让 AI 按一句话在 Figma 里改稿：DeepSeek Harness 直接修改原生图层，不占用官方 MCP 额度</title><link>https://ttqq.dev/en/tool/dsh-to-figma-deepseek-harness-figma-editing/</link><guid>https://ttqq.dev/en/tool/dsh-to-figma-deepseek-harness-figma-editing/</guid><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><description>UI designer and indie dev asa (@double_asa) built a DeepSeek Harness (DSH) plugin (MIT): a local service talks to a Figma desktop plugin so DSH can read the selected frame and create/edit native layers, components, variables and auto layout—output you can keep editing by hand, not a flat image or code dump. The author's Oct 19 demo: three sentences took a dark UI → a light version on a new frame → a terminal-style version. We cloned v0.3.2, ran its test suite (80 tests: 79 pass, 1 skipped) and verified the release tarball SHA-256; without DSH or Figma desktop we did no end-to-end edit. / UI 设计师兼独立开发者 asa（@double_asa）做的 DeepSeek Harness（DSH）插件（MIT）：本机服务直连 Figma 桌面端插件，让 DSH 读取选中画板、新建/修改原生图层、组件、变量和自动布局，产出能继续手动改的设计稿，而不是一张图或一堆代码。作者 10 月 19 日实拍：三句话把深色界面读出来 → 新画板做浅色版 → 再做终端风格版。我们克隆 v0.3.2，跑了自带测试（80 项：79 通过、1 跳过），并核对 Release 安装包 SHA-256；没有 DSH 和 Figma 桌面端，未做端到端改稿。</description></item><item><title>@prompt48 cloned the open-source iFixAi to check whether an AI agent oversteps before launch; we reproduced the quick start and also ran all 61 inspections / @prompt48 克隆开源的 iFixAi，在上线前检查 AI 智能体会不会越权；我们复现了快速上手，并多跑了全量 61 项</title><link>https://ttqq.dev/en/tool/ifixai-agent-audit-default-vs-full-run/</link><guid>https://ttqq.dev/en/tool/ifixai-agent-audit-default-vs-full-run/</guid><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><description>iFixAi (Apache-2.0, Python, pip package ifixai 4.3.0) drops your agent into a simulated business, sends a real day's requests—including ones that push it to overstep, pay a declined refund or change a payout account—and a judge model grades each reply with evidence. Overseas creator @prompt48 cloned, installed and ran the 10 default inspections on video, warning that the demo agent's grade A is only a smoke test. We reproduced it on the box from the README: default 10 all pass, grade A 100%; adding --full-run for all 61, the same demo agent fails 5 and drops to 90.9%—still grade A. No real agent or real judge was connected. / iFixAi（Apache-2.0，Python，pip 包 ifixai 4.3.0）把你的智能体放进模拟业务环境，发送真实工作日的请求（包括诱导它越权、退已被拒的款、改收款账户之类），再由评审模型逐条打分并给出证据。海外作者 @prompt48 克隆、安装、跑了默认 10 项检查并录了视频，提醒内置演示 agent 的 A 级只是冒烟测试。我们在盒子上照 README 复现：默认 10 项全过、A 级 100%；再加 --full-run 跑全部 61 项，同一个演示 agent 出现 5 项失败，分数降到 90.9% 但仍是 A。没有接入真实 agent 或真实评审模型。</description></item><item><title>The Photo Relic Editorial Codex skill turns one photo into a vertical piece: the photo on top and a paper print below (Paper Beijing series) / 「纸上留影」Codex Skill 把一张照片做成竖版作品：上面是照片，下面是一幅纸上版画（纸上北京系列）</title><link>https://ttqq.dev/en/tool/photo-relic-editorial-codex-skill-paper-memory/</link><guid>https://ttqq.dev/en/tool/photo-relic-editorial-codex-skill-paper-memory/</guid><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate><description>Wu Haoyang (GitHub @wnby) open-sourced a Codex image skill (MIT): give it a photo and Codex calls an image tool to make a vertical piece—the photo on top, a minimal "paper print" distilled from it below (outline, a few deep-blue ink blocks, one warm accent, a tiny title). The repo ships six "Paper Beijing" samples (Temple of Heaven, Corner Tower, Bird's Nest, China Zun, Great Wall, geese). Chinese creator @xunana1017 recommended it on Nov 1. We cloned it, read SKILL.md and the reference prompt, and inspected the six samples; we did not generate new images in Codex. / 吴昊阳（GitHub @wnby）开源的 Codex 图像创作 Skill（MIT）：给一张照片，让 Codex 调用图像生成工具出一张竖版作品——上半部分是原照片，下半部分是从照片里提炼出的极简「纸上版画」（建筑轮廓、几块深蓝油墨、一点暖色、小标题）。仓库附 6 张「纸上北京」样例（天坛、角楼、鸟巢、中国尊、长城、雁阵）。中文社区 @xunana1017 于 11-01 撰文推荐。我们克隆仓库，通读 SKILL.md 与参考提示词、检查 6 张样例图；没有在 Codex 里实际生成新图。</description></item><item><title>@wangzan101's open-source on-device Android Phone Agent, Mika, operates apps on a background virtual screen while you keep using the phone / @wangzan101 开源的安卓本地 Phone Agent Mika 在后台虚拟屏上操作任意 App，你可以照常用手机</title><link>https://ttqq.dev/en/tool/mika-android-phone-agent/</link><guid>https://ttqq.dev/en/tool/mika-android-phone-agent/</guid><pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate><description>Wang Zan (@wangzan101) open-sourced Mika (Apache-2.0, thousands of stars at launch, latest Release 0.1.33): a Phone Agent that lives on your Android phone. After one sentence it opens apps, taps, types and scrolls on a separate background screen (WeChat messages, group summaries, Xiaohongshu digests, job-site greets, built-in browser shopping…) while you keep using the main display; logins, OTP and payments are handed to you. No required root, no always-on PC; models use your API key or a ChatGPT / xAI / OpenRouter sign-in—the project says there is no Mika server. App control needs Android 12+ (14+ recommended) and a one-time wireless-debugging pair. We cloned commit ebb0e44, read the EN/ZH READMEs, setup guide and privacy notes, and confirmed Releases ship an APK; we had no Android device, installed no APK, and ran no end-to-end model task. / 王赞（@wangzan101）开源 Mika（Apache-2.0，发布时约数千 Star，Releases 最新 0.1.33）：装在安卓手机上的 Phone Agent。一句话交代任务后，它在单独的后台屏上打开 App、点击、输入、翻页（微信发消息、总结群聊、小红书扒内容、BOSS 直聘打招呼、内置浏览器购物等），你主屏幕照常用；登录、验证码、付款会交给你。不强制 Root、不依赖电脑常驻；模型用你自己的 API Key 或 ChatGPT / xAI / OpenRouter 账号，官方称没有 Mika 服务器。操作 App 需 Android 12+（推荐 14+）并配对一次无线调试。我们克隆 commit ebb0e44，通读中英 README、设置指南与隐私说明，核对 Releases 有 APK；没有安卓真机，未安装 APK，也未连接任何模型做端到端任务。</description></item><item><title>@I_am_oil_oil's open-source oil-diff turns a Git Diff of dozens of AI-edited files into an HTML report of 3–7 business changes that a non-coder can read / @I_am_oil_oil 开源的 oil-diff 把 AI 改动的几十个文件（Git Diff）整理成 3～7 件业务事项的 HTML 汇报页，不写代码的人也能看懂</title><link>https://ttqq.dev/en/tool/oil-diff-git-change-visual-report/</link><guid>https://ttqq.dev/en/tool/oil-diff-git-change-visual-report/</guid><pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate><description>oil (@I_am_oil_oil, GitHub oil-oil) open-sourced oil-diff (MIT, 0.1.0), a skill for Claude Code / Codex and similar agents. When an agent touches dozens of files in minutes, raw Git Diff is hard to skim and a free-form HTML write-up from the model often fills with fluff. oil-diff collects the diff with a script, groups it into 3–7 business items, and emits a single-file `.oil-diff/index.html`: a Bento board, before/after diagrams, 🎯 must-read vs 📎 peripheral files, and a side code drawer. Output stays in an untracked `.oil-diff/` folder (built-in `*` gitignore) so the repo status is unchanged. We cloned commit 8f4b40c and ran the author's `python3 -m unittest` suite: 5/5 passed (collect→render→lint, Chinese paths and binaries, blocking scavenged fake screenshots, hype-word lint). We did not have an agent produce a full report on a real product repo. / oil（@I_am_oil_oil，GitHub oil-oil）新开源 oil-diff（MIT，0.1.0）：给 Claude Code / Codex 等 Agent 用的 Skill。Agent 十几分钟改三四十个文件时，直接翻 Git Diff 抓不住重点，让模型从零写 HTML 报告又容易堆满废话。oil-diff 用采集脚本抽出改动，再按业务归纳成 3～7 件事，生成单文件 `.oil-diff/index.html`：Bento 看板、前后对比图示、🎯核心必看 / 📎外围同步、右侧代码抽屉。产物写在未跟踪的 `.oil-diff/`（内置 `*` gitignore），不改仓库状态。我们克隆 commit 8f4b40c，`python3 -m unittest` 跑作者测试 5/5 通过（覆盖采集→渲染→lint、中文路径与二进制、拦截仓库里抄来的假截图、空话词检查）；没有在真实业务仓库上让 Agent 出完整汇报页。</description></item><item><title>@Fastmax_'s open-source Claude Code papercraft skill skips Meshy / Tripo: it unfolds a box net as UVs, has an image model paint it, and folds it back into a low-poly GLB in Blender / @Fastmax_ 开源的 Claude Code 折纸 3D 技能不走 Meshy / Tripo：把纸盒展开成 UV，让图像模型上色，再用 Blender 折回低模 GLB</title><link>https://ttqq.dev/en/tool/origami-3dfolder-ai-papercraft-glb/</link><guid>https://ttqq.dev/en/tool/origami-3dfolder-ai-papercraft-glb/</guid><pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate><description>Fastmax (X @Fastmax_, repo author alias lejrimostfa) open-sourced Origami-3Dfolder-ai (Apache-2.0): a Claude Code skill plus Python/Blender scripts that turn papercraft-friendly props—crates, tables, windows, stalls, water towers, trees—into textured low-poly GLBs with a Fold animation. Flow: write a spec → `net.py` unfolds a UV dieline → an image model paints the template (~$0.10–0.25, with `--dry-run`) → headless Blender folds → `glb_check.py` gates. Eleven pre-painted examples ship so you can skip the paid paint. We cloned commit 4190681, installed numpy/opencv/Pillow: `net.py`+`template.py draw` produced crate template/mask; 87 author tests passed; `glb_check` on shipped `crate.glb` / `crate_fold.glb` reported ok (41 KB / 60 KB, Fold clip with 15 morphs). No Blender on the box, no paid paint call, no Claude session to invent a new asset. / Fastmax（X @Fastmax_，仓库作者署名 lejrimostfa）开源 Origami-3Dfolder-ai（Apache-2.0）：一套 Claude Code Skill + Python/Blender 脚本，把箱子、桌子、窗、摊位、水塔、树等「纸质感」道具做成带贴图的低模 GLB，并附一段 Fold 折起动画。流程是：写 spec → `net.py` 展开 UV 刀版 → 图像模型按模板上色（约 $0.10–0.25，可 `--dry-run`）→ headless Blender 折回 → `glb_check.py` 验收。仓库自带 11 个已上色样例（可跳过付费绘制）。我们克隆 commit 4190681，装 numpy/opencv/Pillow 后：`net.py`+`template.py draw` 能为 crate 写出 template/mask；作者测试 87 项通过；对样例 `crate.glb` / `crate_fold.glb` 的 `glb_check` 报 ok（41 KB / 60 KB，Fold 动画 15 个 morph）。本机无 Blender，没有跑付费上色，也没有开 Claude 会话造新资产。</description></item><item><title>@kevin327g's open-source Bitwarden MCP login broker lets an agent log in without seeing the password; tested with Claude Code / Codex through GitHub 2FA / @kevin327g 用 Bitwarden 做的开源 MCP「登录代办」让 Agent 登录网站但看不到密码，Claude Code / Codex 实测通过了 GitHub 两步验证</title><link>https://ttqq.dev/en/tool/agent-login-broker-bitwarden-mcp-agent-login/</link><guid>https://ttqq.dev/en/tool/agent-login-broker-bitwarden-mcp-agent-login/</guid><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><description>kevin mitsunaga of the Japanese band fhána (@kevin327g, building as qulup) uses Bitwarden, envied 1Password for Agents, and noticed Claude Code and Codex couldn't use it yet—so he built agent-login-broker (MIT, 0.1.2), an MCP server that does exactly one job: logging in. The agent just says "log in to app.example.com"; the server fetches one credential through Bitwarden Agent Access (aac) only after you approve, types it into its own headless Chromium (computing a TOTP code if the item has an authenticator key), drops it, and returns only "logged in to …" to the agent. The author says he tested it with Claude Code, Codex and omp, including GitHub switching from the passkey prompt to an authenticator code. We cloned commit e631a7f, installed deps plus Playwright's pinned Chromium, and ran the author's test suite: all 56 tests passed. We did not connect a real Bitwarden vault or log in to any real site. / 日本乐队 fhána 的 kevin mitsunaga（@kevin327g，开发署名 qulup）是 Bitwarden 用户，羡慕 1Password for Agents 却发现 Claude Code、Codex 还用不上，于是自己做了 agent-login-broker（MIT，0.1.2）：一个只干「登录」这件事的 MCP 服务器。Agent 只说「登录 app.example.com」，服务器通过 Bitwarden 的 Agent Access（aac）在你逐次批准后取出一条凭据，用自己的无头 Chromium 填表（有 TOTP 密钥就顺便算两步验证码），用完丢掉，只回给 Agent「logged in to …」。作者称在 Claude Code、Codex、omp 上测过，GitHub 能从 passkey 切到验证器验证码并完成登录。我们克隆了仓库 commit e631a7f，装好依赖和 Playwright 自带的 Chromium 后跑了作者的测试套件：56 项全部通过；没有接真实 Bitwarden 账号、没有登录任何真实网站。</description></item><item><title>@zhengtang364848 tested OpenMausBot, which is widely promoted on X, with DeepSeek and Zhipu GLM: delegation, memory and schedules worked for about ¥3 a round, but the app has to stay running / @zhengtang364848 用 DeepSeek 和智谱 GLM 实测 X 上热推的 OpenMausBot：派活、记忆、定时都能用，一轮约 ¥3，但程序要一直开着</title><link>https://ttqq.dev/en/tool/openmausbot-deepseek-glm-local-bot-team-test/</link><guid>https://ttqq.dev/en/tool/openmausbot-deepseek-glm-local-bot-team-test/</guid><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><description>OpenMausBot is an open-source, Grok Bot-style local bot-team UI (Apache-2.0): a roster of bots on the left, chat in the middle, and a Chief of Staff that hands work to other bots. Rather than repeat the hype, 秋/十三 (@zhengtang364848) pointed the Claude Code CLI at DeepSeek and Zhipu and tested item by item: three small tasks with one prompt, schedule accuracy and catch-up after downtime, a two-step Chief delegation checked against a known answer, long-term memory after restart—and estimated cost from token counts. Verdict: it's a management layer over the command-line AI already on your machine; quick install, punctual schedules, delegation and memory worked, but it brings no model, must stay running, and every scheduled run is a full conversation. We cloned commit 88d54eb and confirmed two behaviours in source—5-minute minimum interval and catching up only the latest missed run—without installing deps, running it or calling any model. / OpenMausBot 是开源的「Grok Bot 式」本地 bot 团队界面（Apache-2.0）：左边一排 bot、中间聊天，还有能给别的 bot 派活的 Chief of Staff。秋/十三（@zhengtang364848）没有只看宣传，而是把 Claude Code 命令行的出口换成 DeepSeek 和智谱，一项项实测：同一提示词的三个小任务、定时任务的准点与停机补跑、Chief 拆两步任务并核对答案、重启后长期记忆是否生效，并按 token 估了花费。结论：它本质是套在你本机命令行 AI 外面的一层管理界面，安装快、定时准、派活和记忆都能正常工作，但不自带模型、程序得一直开着、每次定时都是一整轮对话。我们克隆了仓库 commit 88d54eb，在源码里核对了「间隔最短 5 分钟」和「停机后只补最近一次」这两条行为；没有安装依赖并运行，也没有调用任何模型。</description></item><item><title>@HiTw93's open-source layout skill Kami turns text an AI has written into PDF, PNG or editable PPTX from one sentence, with handling for English, Chinese, Japanese and Korean / @HiTw93 开源的排版技能 Kami 把 AI 写好的文字排成 PDF、PNG 或可编辑的 PPT，一句话就能出稿，中英日韩都做了处理</title><link>https://ttqq.dev/en/tool/kami-agent-document-layout-pdf-pptx/</link><guid>https://ttqq.dev/en/tool/kami-agent-document-layout-pdf-pptx/</guid><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate><description>Tw93 (@HiTw93) packaged agent document-layout rules as Kami (紙, MIT, ~12k★ on GitHub): after install on Claude Code / Codex / Cursor, natural language yields one-pagers, long docs, letters, portfolios, resumes, slides, equity reports, changelogs and landing pages—warm parchment, ink-blue accents, serif fonts—exportable as PDF, PNG and editable PPTX. V1.20.0 (published 2026-10-29) highlights cleaner reports, charts drawn from data and checked before delivery, and Chinese/Korean pages that read well on iPhone. Install with `npx skills add tw93/kami -a claude-code codex cursor -g -y`. We cloned commit e16dd6a, confirmed VERSION=1.20.0, syntax-checked mcp_server.py and related scripts, and inventoried 10 shipped demo PDFs; WeasyPrint was not installed here, so we did not render a new document. / Tw93（@HiTw93）把给 Agent 用的文档排版规范收成 Kami（紙，MIT，GitHub 约 1.2 万★）：装进 Claude Code / Codex / Cursor 后，自然语言就能出一页纸、长文、信件、作品集、简历、幻灯片、研报、更新日志和落地页，默认浅米色底 + 油墨蓝 + 衬线字体，可导出 PDF、PNG 与可编辑 PPTX。V1.20.0（2026-10-29 发布）强调更干净的报表、按数据画图并交付前检查，以及中韩页在 iPhone 上的阅读效果。安装：`npx skills add tw93/kami -a claude-code codex cursor -g -y`。我们克隆仓库 commit e16dd6a，核对 VERSION=1.20.0，对 mcp_server.py 等脚本做了语法检查，并清点了 10 份官方 demo PDF；本机未装 WeasyPrint，没有真的渲染新文档。</description></item><item><title>@rehan_shei open-sourced pixelfx and a Claude Code skill: 232 pixel-art effects written entirely in code with no image files, and new effects follow Hugo Duprez's draft → render → look → refine loop / @rehan_shei 开源 pixelfx 和配套的 Claude Code 技能：232 个像素特效全部用代码写成，不用贴图，新特效按 Hugo Duprez 的「草稿→渲染→看图→修改」循环来做</title><link>https://ttqq.dev/en/tool/pixelfx-pixel-art-effects-claude-skill/</link><guid>https://ttqq.dev/en/tool/pixelfx-pixel-art-effects-claude-skill/</guid><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate><description>Rehan Sheikh (@rehan_shei) open-sourced pixelfx (MIT): 232 pixel-art effects (fire, explosions, ice, lightning, UI, critters…) each a small JS file—play live, recolor, mirror, or export sprite sheets for Godot/Unity. A Claude Code skill follows Hugo Duprez's (@HugoDuprez) public method—code is clay, trust your eyes, feed references—so Opus drafts, renders a contact sheet, looks, and refines. Gallery: patch-gg.github.io/pixelfx. We cloned commit fe7103c, ran `npm install` then `node tools/render.mjs ice/frost-nova --sheet`, and got a 103-frame sheet plus Aseprite-style json; we did not have Claude author a new effect. / Rehan Sheikh（@rehan_shei）开源了 pixelfx（MIT）：232 个像素风特效（火焰、爆炸、冰、闪电、UI、小动物等）每个都是一小段 JS，可实时播放、换色、镜像，或导出精灵表给 Godot/Unity。附带 Claude Code 技能，按 Hugo Duprez（@HugoDuprez）公开方法——代码当黏土、相信眼睛、喂参考图——让 Opus 写一版、渲染接触表、对照修到能看。画廊：patch-gg.github.io/pixelfx。我们克隆 commit fe7103c，`npm install` 后跑 `node tools/render.mjs ice/frost-nova --sheet`，得到 103 帧精灵表与 Aseprite 风格 json；没有开 Claude 新做特效。</description></item><item><title>@elvar_official packed his Claude Opus workflow for turning one reference image into a take-apart 3D machine simulator into the teardown skill, with all 9 prompts public / @elvar_official 把从一张参考图做出可拆开的 3D 机械模拟器的 Claude Opus 流程做成 teardown 技能，9 条提示词全部公开</title><link>https://ttqq.dev/en/tool/teardown-claude-skill-machine-3d-simulator/</link><guid>https://ttqq.dev/en/tool/teardown-claude-skill-machine-3d-simulator/</guid><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate><description>Elvar (@elvar_official) turned "any machine with moving parts → an interactive 3D simulator" into a Claude Code skill, teardown (MIT): give Opus one reference image, insist every part is three.js code driven by one math model that also feeds the gauges; cut parts open, press subsystem buttons, auto-tour when idle, and record a 60 fps film with a procedural score. The repo ships three machines (V8, Cessna landing gear, Strandbeest). Install with `npx skills add elvarofficial/teardown -g -a claude-code`, then npm install + Playwright Chromium in the skill folder. His X Article publishes all 9 prompts; demos live at visual-learning.app. On Linux we cloned commit e7e668c, installed deps, and bundled cessna-live.html (~799 KB with three.js inlined); we did not open Claude to invent a new machine or record video. / Elvar（@elvar_official）把「把任意带运动部件的机器做成可互动 3D 模拟器」的流程收成 Claude Code 技能 teardown（MIT）：给 Opus 一张参考图，要求每个零件都是 three.js 代码、同一套数学模型驱动画面与仪表；可剖开、点子系统按钮、无人操作时自动导览，还能逐帧录 60fps 成片并配程序化音效。仓库自带 V8、Cessna 起落架、Strandbeest 三台样机；安装 `npx skills add elvarofficial/teardown -g -a claude-code`，再在技能目录 npm install + Playwright Chromium。X Article《How I Turn Any Machine Into a 3D Simulator You Can Take Apart》给了完整 9 步提示词；演示站 visual-learning.app。我们在 Linux 上克隆 commit e7e668c，装依赖并成功打包 cessna-live.html（约 799 KB、内联 three.js）；没有开 Claude 会话造新机，也没录视频。</description></item><item><title>An open-source SKILL.md has your agent build a 6-layer firewall against scraping agents: it serves 200 gate pages instead of 403s, verifies Googlebot by IP, then red-teams the setup with sub-agents / 一份开源 SKILL.md 让 Agent 给你的网站搭 6 层反 Agent 防火墙：返回 200 门页而不是 403，按 IP 核实 Googlebot，再派子 Agent 做红队自测</title><link>https://ttqq.dev/en/tool/block-ai-scrapers-agent-firewall-skill/</link><guid>https://ttqq.dev/en/tool/block-ai-scrapers-agent-firewall-skill/</guid><pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate><description>Overseas developer @illyism saw people paste his data site into Claude Code and brag about scraping the whole database; when he returned 403s, the agents spoofed Googlebot, probed /api/ and spun up Playwright. He built a 6-layer defense and open-sourced the full SKILL.md: blocked requests get an HTTP 200 "protected dataset" gate page (zero rows + a STOP_SCRAPING directive + a product link), Googlebot is verified by IP/ASN instead of User-Agent, HTML routes require full browser navigation headers plus a JS session proof, and sub-agents red-team it until attackers stop on turn 1. We ran the skill's Googlebot check against Google's IP list published that day and found it both misses real Googlebot ranges and admits unofficial ones. / 海外开发者 @illyism 发现有人把他的数据站贴进 Claude Code 炫耀「整库扒下来」，返回 403 后 Agent 会伪装 Googlebot、探 /api/、开 Playwright 继续扒。他做了一套 6 层防护并把完整 SKILL.md 开源：被拦的请求返回 HTTP 200 的「受保护数据」门页（0 行数据 + STOP_SCRAPING 指令 + 产品链接），Googlebot 按 IP/ASN 而不是 User-Agent 验真，HTML 路由要求完整浏览器导航头和 JS 会话证明，最后派子 Agent 当攻击者测到「第 1 轮就停」。我们把技能里的 Googlebot 校验函数拿 Google 当天公布的爬虫 IP 表跑了一遍，发现它会漏掉真 Googlebot 网段、也会放过非官方网段。</description></item><item><title>The Douyin "Vibe Knowledge" explainer-short format as an open-source Claude Code skill: give it a topic and it fact-checks, writes, voices, renders frame by frame in code and runs QA; the base tier costs zero credits / 把抖音「Vibe 知识大赏」式的知识短片做法做成开源 Claude Code 技能：给一个选题，它依次查证、写稿、配音、用代码逐帧成片并质检，基础档 0 积分</title><link>https://ttqq.dev/en/tool/vibe-knowledge-claude-code-knowledge-shorts/</link><guid>https://ttqq.dev/en/tool/vibe-knowledge-claude-code-knowledge-shorts/</guid><pubDate>Thu, 29 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator @berryxia turned the viral Douyin #Vibe知识大赏 ("Vibe Knowledge") explainer-short format into an open-source skill, vibe-knowledge: 6 storytelling templates (guess-the-concept, dilemma, everyday-paradox, etc.), 44 visual styles, and every narration line must trace to a source in facts.md. The base tier is pure Claude Opus SVG/code animation at zero credits; the enhanced tier hands realistic shots to Renoise (quote first, then generate). The repo ships 7 sample films. We cloned the 2026-10-28 evening build on Linux: the style recommender works, but workspace setup stopped at the BGM patch added that same day; we rendered and voiced nothing. / 中文圈 @berryxia 把最近国内爆火的 #Vibe知识大赏 知识短片做法拆成 vibe-knowledge 技能并开源：6 种讲法母版（猜谜、两难选择、痛点原理等）、44 套画面风格，每句旁白都要在 facts.md 里有出处；基础档全靠 Claude Opus 写 SVG/代码动画、0 积分出片，增强档把写实镜头交给 Renoise 生成（先报价再生成）。仓库附 7 条样片。我们在 Linux 上克隆了 2026-10-28 晚的版本，成功运行了风格推荐脚本，建工作区时卡在当天新加的 BGM 补丁步骤；没有渲染、没有配音。</description></item><item><title>Connect the public, key-free "AI video intel radar" MCP to Claude Code / Codex and ask which AI-video formats emerged today, where a trend started, or what a model shipped this week / 把免 Key 的公开 MCP「AI 视频情报站」接进 Claude Code / Codex，用中文问今天冒头的 AI 视频题材、某个玩法的来源和某个模型这周的动静</title><link>https://ttqq.dev/en/tool/ai-video-intel-radar-mcp-daily-topics/</link><guid>https://ttqq.dev/en/tool/ai-video-intel-radar-mcp-daily-topics/</guid><pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator @GaryAIytb (Gary Cao) runs an AI intel radar that scans public posts on Bilibili, YouTube, Douyin, TikTok, Xiaohongshu, X, GitHub and more every morning to flag AI-video formats that are just taking off. He had Claude build a remote MCP server; one command connects it to Claude Code / Codex, then you ask in Chinese: what's new or heating up today, when did this format start and is it too late, what did Kling ship this week. We called all 7 tools directly over the MCP protocol (no client)—anonymous and returning real data—and ran the open-source version's free collectors locally: 1 min 46 s, zero cost. / 中文圈 @GaryAIytb（Gary曹淦）做的 AI 情报站每天早上扫 B 站、YouTube、抖音、TikTok、小红书、X、GitHub 等公开内容，挑出刚冒头的 AI 视频题材；他让 Claude 写了一个远程 MCP，Claude Code / Codex 一行命令接上后，用中文问「今天新起/升温的题材」「这个玩法什么时候开始火、现在进场晚不晚」「Kling 最近一周有什么动静」。我们不经客户端、直接按 MCP 协议调了它的 7 个工具，全部匿名可用、有结果；还在本机跑了开源版的免费采集，1 分 46 秒、0 花费。</description></item><item><title>After building an 80-second animated short by hand on the Dreamina (Jimeng) canvas, Claude Code scripts the pipeline through the canvas CLI: change only the lead character, bear to penguin, and 14 images, 10 clips and the music re-run automatically / 在即梦画布上做完一条 80 秒动画短片后，让 Claude Code 通过画布 CLI 把流程写成脚本：只改主角，熊换成企鹅，14 张图、10 条视频和配乐全部自动重跑</title><link>https://ttqq.dev/en/tool/dreamina-canvas-cli-claude-code-swap-protagonist/</link><guid>https://ttqq.dev/en/tool/dreamina-canvas-cli-claude-code-swap-protagonist/</guid><pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator @servasyy_ai (huangserva) first built an 80-second animated short by hand on the new Dreamina canvas (a bear dancing in an old-town square): lock the character with a subject + turnaround sheet, cut 4- and 9-panel storyboards into single frames, hang a Seedance 2.5 video node off each frame, then add music on the timeline and export. Then he had Claude Code turn that node graph into a script via the canvas CLI with one parameter—who the lead is—and re-ran it as a penguin; he says it cost 1,750 credits and he clicked no nodes. Note: the post carries a #DreaminaPartner tag (sponsored). On Linux we installed the official dreamina-canvas CLI 1.1.1 (China build) and confirmed the quote → confirm-with-credit-ceiling → run-with-token commands, image upscale and draft-to-final promote exist; we did not log in or generate anything. / 中文圈 @servasyy_ai（huangserva）先在新版 Dreamina 画布里手动做了一条 80 秒动画短片（熊在古镇广场跳舞）：主体 + 三视图定型、四格/九格分镜切成单张、每格接 Seedance 2.5 视频节点、时间线加配乐导出。然后让 Claude Code 通过画布 CLI 把这套节点流程写成脚本，只留一个参数「主角是谁」，换成企鹅后整条重跑，作者称一共 1750 积分、自己一个节点没点。注意这是带 #DreaminaPartner 标签的合作推广帖。我们在 Linux 上装了官方 dreamina-canvas CLI 1.1.1（国内版），核对了「报价 → 设积分上限确认 → 带凭证运行」和图片超清、样片转正等命令确实存在；没有登录、没有生成任何内容。</description></item><item><title>video-to-skill turns a YouTube tech talk into an agent skill: every point has an [mm:ss] timestamp and a seen / guessed / unknown evidence label, and the agent loads it when the topic comes up / video-to-skill 把一条 YouTube 技术分享做成 Agent 技能包：每条要点带 [mm:ss] 时间戳和「看到 / 推测 / 未知」证据标签，聊到相关话题时 Agent 会自动加载</title><link>https://ttqq.dev/en/tool/video-to-skill-talk-to-agent-skill/</link><guid>https://ttqq.dev/en/tool/video-to-skill-talk-to-agent-skill/</guid><pubDate>Wed, 28 Oct 2026 00:00:00 GMT</pubDate><description>video-to-skill by developer @prince_twets (prince): give it a YouTube URL or local video, it gets a transcript via summarize CLI → yt-dlp captions → local Whisper, then the agent writes SKILL.md plus segment notes, a cheatsheet, glossary and timestamp jump list—every claim tagged [mm:ss] and [seen]/[guessed]/[unknown]—and validate.py checks the package. His first real test turned a video on the Vercel CEO's production AI-engineering advice into a published vercel-ceo-skill. We ran the environment check and unit tests, validated his published output (pass, 245 evidence labels), and tried fetching YouTube captions from a server (rate-limited, HTTP 429). / 海外开发者 @prince_twets（prince）做的 video-to-skill：给一个 YouTube 链接或本地视频，按 summarize CLI → yt-dlp 字幕 → 本地 Whisper 的顺序拿转写，再让 Agent 写成 SKILL.md + 分段笔记、速查表、术语表、时间戳跳转表，每条结论标 [mm:ss] 和 [seen]/[guessed]/[unknown]，最后用 validate.py 检查。他第一个实测是把一条讲 Vercel CEO 生产级 AI 工程做法的视频转成 vercel-ceo-skill 并公开。我们跑了它的环境检查和单元测试，用校验器检查了作者公开的成品（通过，245 个证据标签），也试了从服务器抓 YouTube 字幕（被 429 限流）。</description></item><item><title>Xian (@Xian0063), a UX designer for 10 years, made 18 animated explainer videos with Claude in 4 days, every frame drawn in code; the Van Gogh episode has one visual per line, and the voiceover was checked word by word with Whisper / 做了 10 年 UX 的设计师弦（@Xian0063）4 天用 Claude 做了 18 条动画讲解视频，画面全由代码绘制；梵高那条一句台词配一个画面，配音用 Whisper 逐字核对</title><link>https://ttqq.dev/en/tool/claude-remotion-van-gogh-explainer-animation/</link><guid>https://ttqq.dev/en/tool/claude-remotion-van-gogh-explainer-animation/</guid><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><description>Xian (@Xian0063, a self-described 10-year UX designer who doesn't code) wrote a long retrospective: 18 animated videos in 4 days with Claude, ~50 renders, including a 3-minute Van Gogh piece made "without downloading a single image or opening an editor"—the paintings are drawn by code with thousands of brush strokes. The post shares a copyable Remotion prompt template, an 8-step method (topic → emotional-arc script → voiceover checked word by word → visuals grown from the subject → one visual per line → frame-grab QA → cover frame at the start → multi-platform export), 6 design rules, and an open-sourced Claude Code skill, landscape-explainer-video. We inspected the skill repo, syntax-checked its scripts, ran its English-subtitle line-count checker, and generated a line of Xiaoxiao +8% voiceover with edge-tts as described; we did not render a video—the skill depends on the author's own Remotion project and won't run standalone. / 弦（@Xian0063，自述 10 年 UX、不写代码）发长文复盘：4 天用 Claude 做了 18 条动画视频、渲染约 50 版，其中一条 3 分钟梵高动画「没下载一张图、没开剪辑软件」，名画是程序用几千根笔触一笔笔画出来的。文章给出可复制的 Remotion 提示词模板、8 步做片法（选题→情绪曲线脚本→配音逐字核对→从题材长出视觉→一句一画→截帧质检→封面帧放片头→多平台出片）和 6 条设计原则，并把方法开源成 Claude Code skill landscape-explainer-video。我们检查了 skill 仓库、跑了脚本语法检查和英文字幕句数核对脚本，并用 edge-tts 按文中设置生成了一句晓晓 +8% 配音；没有渲染视频——这个 skill 依赖作者自己的 Remotion 工程，单独装上跑不起来。</description></item><item><title>ffmpeg-skill wraps local FFmpeg as 42 tools, so Claude Code / Codex measure duration, loudness and aspect ratio after an edit and hand in the check results with the video / ffmpeg-skill 把本地 FFmpeg 封装成 42 个工具，让 Claude Code / Codex 剪完视频后自己量时长、响度和画幅，连同检查结果一起交片</title><link>https://ttqq.dev/en/tool/ffmpeg-skill-agent-video-edit-self-check/</link><guid>https://ttqq.dev/en/tool/ffmpeg-skill-agent-video-edit-self-check/</guid><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><description>ffmpeg-skill (by kajisho5, MIT, ~3k★ on GitHub), recommended in Chinese by @ZhongRichard, is a local editing skill for Claude Code, Codex and Cursor: silence removal, cut/join, 16:9→9:16, captions/karaoke, loudness normalisation and TikTok / Reels / YouTube / X export templates—42 tools that also run as an MCP server. The point is a fixed loop—probe the footage → edit → check against the platform spec → make a contact sheet for a vision model when the picture changed—so the agent can't just say "done". On Linux we ran probe → silence removal → 9:16 crop → Reels export → check → loudness fix → re-check → contact sheet on synthetic footage straight from the repo; we did not drive it from a live agent session or test transcription. / 中文圈 @ZhongRichard 推荐的 ffmpeg-skill（作者 kajisho5，MIT，GitHub 约 3000★）是给 Claude Code、Codex、Cursor 用的本地剪辑技能：去静音、裁切拼接、横转竖、字幕/卡拉 OK、响度归一化、按 TikTok / Reels / YouTube / X 模板导出，共 42 个工具，同时是 MCP 服务器。卖点是固定流程「先 probe 量素材 → 剪 → 按平台规格 check → 画面变了就出抽帧拼图给视觉模型看」，Agent 不能只说一句「剪好了」。我们在 Linux 上用合成素材从仓库直接跑了 probe → 去静音 → 9:16 裁切 → reels 导出 → check → 响度修正 → 再 check → 抽帧拼图，整个流程都正常完成；没有接真实 Agent 会话，也没测自动转写。</description></item><item><title>Designer Ash (@ashdesignin) builds all app micro-animations with Claude Code: a Figma link becomes a SwiftUI first draft, 3–5 rounds of describing the feel refine it, and an iPhone screen recording becomes the promo clip / 设计师 Ash（@ashdesignin）的 App 动效全部交给 Claude Code：从 Figma 链接生成 SwiftUI 初版，用 3–5 轮描述手感来打磨，最后用 iPhone 录屏做成宣传短片</title><link>https://ttqq.dev/en/tool/figma-claude-code-swiftui-micro-animations/</link><guid>https://ttqq.dev/en/tool/figma-claude-code-swiftui-micro-animations/</guid><pubDate>Tue, 27 Oct 2026 00:00:00 GMT</pubDate><description>Design engineer Ash (@ashdesignin, building Yoomy) shared a ~150-like thread with their full app-animation workflow: designs live in Figma, and Claude Code writes 100% of the animation in SwiftUI. Drop the Figma link into Claude Code to rebuild the UI, give it two states (folded/expanded) to animate between, then the hard part—polishing by describing how the motion should feel (a "black hole" sucking the card in from the bottom, with blur and space distortion); replies say 3–5 rounds is typical. Finally, record on iPhone and turn the recording into a square promo with phone mockup, tap circles and slow push-ins using the open-source telegram-style-video skill. We have no Mac/iPhone and did not reproduce the SwiftUI animations; on Linux we only ran that step-4 recording skill on a synthetic screen recording. / 设计工程师 Ash（@ashdesignin，在做 Yoomy）发了一条约 150 赞的长帖，公开自己做 App 动效的完整流程：设计稿在 Figma，动画 100% 由 Claude Code 写成 SwiftUI。先把 Figma 链接丢给 Claude Code 还原界面，再用两种状态（折叠/展开）让它做过渡，最难的是打磨——要一句句描述动作的「感觉」（比如「黑洞」把卡片从底部吸进去、加模糊和空间扭曲），回复里说一般 3–5 轮。最后在 iPhone 上录屏，用开源 skill telegram-style-video 把录屏包装成带手机壳、点击圈和缓慢推镜的方形宣传片。我们没有 Mac / iPhone，没有复现 SwiftUI 动效；只在 Linux 上用合成录屏完整运行了第 4 步那个录屏包装 skill。</description></item><item><title>latent-spaces/brag adds a `/brag` command to Claude Code / Codex: it reads the repo and makes a ~20-second launch video with music and share copy (brag-slim and letsbrag.app are alternatives) / latent-spaces/brag 给 Claude Code / Codex 加了一条 `/brag` 命令：读仓库后生成约 20 秒的上线视频、配乐和分享文案（也可以用 brag-slim 或 letsbrag.app）</title><link>https://ttqq.dev/en/tool/brag-launch-video-skill-hyperframes/</link><guid>https://ttqq.dev/en/tool/brag-launch-video-skill-hyperframes/</guid><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><description>Shunit Haviv Hakimi (GitHub @shunithaviv) ships latent-spaces/brag (MIT, ~18.6k★): an agent skill where you say `let's /brag` in a project, it reads the code for angle and tone, and the classic path hands a brief to Hyperframes. On Opus 5.5 it switches to `/brag-slim` by default (no Hyperframes; the model builds the video with local tools). Output lands in `brag-output/` (plan, share copy, brag.mp4). letsbrag.app can run it from a site URL. @indie_maker_fox summarized install and deps for Chinese readers. We read the README and SKILL.md (v0.5.0); we did not install Hyperframes or render a full video here. / Shunit Haviv Hakimi（GitHub @shunithaviv）在 latent-spaces/brag（MIT，约 1.86 万★）里做了一个 Agent skill：在项目目录说 `let's /brag`，技能读代码定卖点与语气，经典路径把构图交给 Hyperframes 渲染；Opus 5.5 上默认切到 `/brag-slim`（不依赖 Hyperframes，模型本机出片）。成品进 `brag-output/`（计划、文案、brag.mp4）。也可用 letsbrag.app 贴站点链接代跑。中文圈 @indie_maker_fox 整理了安装与依赖。我们读了 README 与 SKILL.md（v0.5.0），没有本机装 Hyperframes / 跑完整渲染。</description></item><item><title>HKUDS's CLI-Anything wraps software such as Blender in command-line tools, so Claude Code / Cursor / OpenClaw can install them from CLI-Hub with one command instead of clicking through screenshots / 港大 HKUDS 的 CLI-Anything 给 Blender 等软件配上命令行，Claude Code / Cursor / OpenClaw 可以通过 CLI-Hub 一条命令装好，不再靠截图点按钮</title><link>https://ttqq.dev/en/tool/cli-anything-agent-native-software-clis/</link><guid>https://ttqq.dev/en/tool/cli-anything-agent-native-software-clis/</guid><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><description>HKUDS open-sourced CLI-Anything (Apache-2.0, ~57k★, clianything.cc): make software agent-ready with stable CLIs instead of screenshot clicking. CLI-Hub (`pip install cli-anything-hub`) browses/installs community harnesses for Blender, GIMP, LibreOffice and more; a generator skill runs a 7-phase HARNESS pipeline when the registry has no match. Works with Claude Code, Cursor, Pi, OpenClaw and peers. @MirrortekUK's Chinese post framed it as dropping pixel-click automation. On our test machine we installed cli-anything-hub in a venv and ran `cli-hub list` / `search blender` / `info blender`; we did not install Blender itself or run a full `/cli-anything` generation pipeline. / 港大数据智能实验室 HKUDS 开源 CLI-Anything（Apache-2.0，约 5.7 万★，clianything.cc）：核心思路是「软件先有稳定 CLI，Agent 才好用」，而不是靠 GUI 截图点选。一边是 CLI-Hub（`pip install cli-anything-hub`）浏览/安装社区已做好的 Blender、GIMP、LibreOffice 等 harness；一边是给 Agent 装生成器 skill，对还没有的软件跑 7 阶段 HARNESS 流水线新造 CLI。兼容 Claude Code、Cursor、Pi、OpenClaw 等。中文圈 @MirrortekUK 介绍了「少靠截图像素点击」的用法。我们在测试机 venv 里装了 cli-anything-hub，成功运行 `cli-hub list` / `search blender` / `info blender`；没有安装 Blender 本体，也没有跑完整 `/cli-anything` 生成流水线。</description></item><item><title>Cathryn's (@cathrynlavery) Diagram Design has Claude Code / Codex / Cursor draw self-contained HTML+SVG architecture diagrams from one sentence, with its visual rules written as hard constraints instead of Mermaid's rounded gray-blue boxes / Cathryn（@cathrynlavery）的 Diagram Design 让 Claude Code / Codex / Cursor 用一句话画出自包含的 HTML+SVG 架构图，审美要求写成了硬规则，不是 Mermaid 那种圆角灰蓝框</title><link>https://ttqq.dev/en/tool/diagram-design-agent-skill-html-svg/</link><guid>https://ttqq.dev/en/tool/diagram-design-agent-skill-html-svg/</guid><pubDate>Mon, 26 Oct 2026 00:00:00 GMT</pubDate><description>Cathryn Lavery (@cathrynlavery) open-sourced Diagram Design (MIT, ~61k★ on GitHub, diagramdesign.dev): a diagram skill for Claude Code, Codex, Cursor, Copilot and peers. The agent picks one of 45 types (architecture, sequence, Sankey, exploded axonometric…), then writes a double-clickable .html. Defaults ban shadows, allow one accent color, and snap geometry to a 4px grid. You can onboard a site URL for palette/fonts, or import draw.io / Mermaid / Excalidraw and export PNG/SVG. Chinese posts by @Zen_with_AI and @dngzsn37461 recently shared the install line and a practical order of use. We cloned the repo, syntax-checked the Python scripts, ran self_check on a shipped example HTML, and extracted a two-node Mermaid digest; we did not open a live Claude Code session to draw a new diagram. / Cathryn Lavery（@cathrynlavery）开源了 Diagram Design（MIT，GitHub 约 6.1 万★，diagramdesign.dev）：给 Claude Code、Codex、Cursor、Copilot 等用的画图 skill。Agent 先选 45 种图型之一（架构、时序、桑基、爆炸轴测…），再写出双击就能开的 .html；默认禁止阴影、全图只用一个强调色、几何落在 4px 网格。可 onboard 站点链接抽配色字体，也能从 draw.io / Mermaid / Excalidraw 导入重画并导出 PNG/SVG。中文圈 @Zen_with_AI、@dngzsn37461 近期分享了安装句与实操顺序。我们克隆仓库、通过了 Python 脚本语法检查、对示例 HTML 跑了 self_check，并用 mermaid_extract 抽了一张两节点图；没有真的开 Claude Code 会话出新图。</description></item><item><title>Underleveled Builder's curated list of short-drama skills: 83 skills checked by hand, plus two published rounds of Opus / GPT blind tests against a no-skill control / Underleveled Builder 的短剧 skill 精选清单：人工核对 83 个 skill，并公开两轮 Opus / GPT 带对照组的盲比</title><link>https://ttqq.dev/en/tool/awesome-short-drama-skills-blind-evals/</link><guid>https://ttqq.dev/en/tool/awesome-short-drama-skills-blind-evals/</guid><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><description>Search GitHub for short-drama skills and you get dozens of repos with no way to tell which ones actually write shootable scripts. Underleveled Builder (@UnderleveledDev) maintains Awesome Short Drama Skills (~752★, CC0): 83 hand-checked short-drama / manju skills plus 16 related open-source platforms, sorted by type, production stage and video model. The list is itself a skill (short-drama-skill-finder)—install it and ask your agent which skill to use for novel-to-drama on Codex. Rarer still: the author published two bake-offs. Round 1 put 10 screenwriting skills plus a no-skill control on the same open brief, three runs each, judged blind by Opus and GPT. Round 2 ranked 11 premise skills the same way. Full logs, outputs and judge fingerprints live in evals/ and on shortdramaskills.com. We only read the README, the evals method and some output folders; we did not re-run the Claude Code harness. / 做 AI 短剧时在 GitHub 搜「短剧 skill」会翻出几十个仓库，很难判断哪个真能写出能拍的剧本。Underleveled Builder（@UnderleveledDev）维护的 Awesome Short Drama Skills（约 752★，CC0）把 83 个短剧/漫剧 skill 和 16 个相关开源平台按类型、创作阶段、视频模型人工核对并分类；清单本身也是一个 skill（short-drama-skill-finder），装上后可以直接问 agent「我用 Codex 做小说改编短剧，装哪个」。作者还公开了两轮实测：第一轮 10 个剧本 skill + 不装 skill 的对照组，同一道开放题各写 3 次，Opus 与 GPT 盲比；第二轮 11 个选题 skill 同样盲比排名。全过程、原始产出和评审指纹都在 evals/ 与 shortdramaskills.com。我们只读了 README、evals 规范与部分产出目录，没有自己重跑 Claude Code 评测。</description></item><item><title>Hassan's (@nutlope) Hallmark gives Claude Code / Cursor / Codex 48 themes and a browser lint that blocks the default styles that make AI-built sites look the same / Hassan（@nutlope）的 Hallmark 给 Claude Code / Cursor / Codex 加上 48 套主题和一个浏览器 lint，专门拦下让网站显得千篇一律的「AI 味」默认样式</title><link>https://ttqq.dev/en/tool/hallmark-anti-ai-slop-design-skill/</link><guid>https://ttqq.dev/en/tool/hallmark-anti-ai-slop-design-skill/</guid><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><description>Hassan at Together AI (@nutlope) open-sourced Hallmark (MIT, ~32k★ on GitHub): a design skill for Claude Code, Codex, Cursor and OpenCode. Instead of another "make it look premium" prompt, it deals three themes, builds page anatomy from the visitor's job (book, buy, subscribe…), then runs sloplint in a real browser across eight widths against empty frames, fake buttons, contrast failures, eyebrow + tracked-capital labels, purple-blue gradients and other machine-made tells—Floor gates never ship. v2 adds `/hallmark redesign` (keep content and wiring, change the system) and `/hallmark audit` (report only). His "I'm never reading IKEA instructions again" demo video drove the spike; he says the full launch stays free and open source. We checked the README, SKILL.md, theme folder and sloplint.mjs syntax; we did not run a live `/hallmark` build. / Together AI 的 Hassan（@nutlope）开源了 Hallmark（MIT，GitHub 约 3.2 万★）：一个给 Claude Code、Codex、Cursor、OpenCode 用的设计 skill。它不是再写一段「请设计得高级一点」的提示词，而是先发三套主题、按访客要做的事（预约、购买、订阅…）决定页面骨架，再用 sloplint 在真实浏览器里按 8 个宽度检查空框、假按钮、对比度、眉题+全大写标签、紫蓝渐变等「一眼 AI」特征；Floor 级问题直接不让过。v2 还支持 /hallmark redesign（保留内容与接线只换版式）和 /hallmark audit（只出报告）。作者用一支「再也不看宜家说明书」的演示视频带火了项目，并表示完整上线仍免费开源。我们核对了 README、SKILL.md、主题目录与 sloplint.mjs 语法，没有真的跑一次 /hallmark 生成页面。</description></item><item><title>@rossark0's open-source rossark-motion-skill teaches Claude Code to turn any SVG into 60 fps native animation with follow-through and breathing (React Native / SwiftUI), without Rive / @rossark0 开源的 rossark-motion-skill 让 Claude Code 把任意 SVG 做成 60fps 的原生动画，带跟随和呼吸效果（React Native / SwiftUI），可以不用 Rive</title><link>https://ttqq.dev/en/tool/rossark-motion-skill-svg-animation/</link><guid>https://ttqq.dev/en/tool/rossark-motion-skill-svg-animation/</guid><pubDate>Sun, 25 Oct 2026 00:00:00 GMT</pubDate><description>ROSS (@rossark0) packaged the method he used to make his app illustrations feel alive into a Claude Code skill: rossark-motion-skill (MIT, ~454★). The point is not swapping expression drawings—it bans teleports, keeps amplitudes small, rotates parts on real pivots, and adds overlap/follow-through. It ships tiny engines for React Native (Reanimated) and SwiftUI, Python tools that measure sub-visible edge jiggle in a reference video, and offline demo scenes (reveal / garden / chef / orbit). Install with `npx skills@latest add rossark0/rossark-motion-skill` and invoke `/rossark-motion-skill`. A follow-up post shows him vibe-animating characters with the skill plus an internal tool, pluy. We syntax-checked track_edges.py / sample_colors.py and read motion.ts plus SKILL.md; full 60 fps demo renders need macOS 14+, which we did not run. / ROSS（@rossark0）把给自己 App 插画「做活」的方法收成一个 Claude Code skill：rossark-motion-skill（MIT，约 454★）。核心不是让模型换表情贴图，而是强制不传送、小振幅、绕关节旋转、重叠与跟随，并自带 React Native（Reanimated）与 SwiftUI 两套小引擎，外加测参考视频边缘微抖的 Python 工具和离线渲染演示（reveal / garden / chef / orbit）。安装一行 `npx skills@latest add rossark0/rossark-motion-skill`，在对话里调用 /rossark-motion-skill。作者后续帖展示用该 skill + 内部工具 pluy 给角色做 vibe animate。我们语法检查了 track_edges.py / sample_colors.py，并读了 motion.ts 与 SKILL.md；完整 60fps 演示渲染需要 macOS 14+，这边没跑。</description></item><item><title>JJ Englert had Claude Code build a "Search Radar" mod that counts empty searches and unneeded file reads in each answer and writes a shortcut doc in one click / JJ Englert 让 Claude Code 自己写了一个「Search Radar」Mod，统计每次回答里的空搜索和多余的读文件，并能一键写出捷径文档</title><link>https://ttqq.dev/en/tool/claude-code-search-radar-mod/</link><guid>https://ttqq.dev/en/tool/claude-code-search-radar-mod/</guid><pubDate>Sat, 24 Oct 2026 00:00:00 GMT</pubDate><description>On 2026-10-19 JJ Englert (@JJEnglert) pointed out that every time Claude Code looks for something in your project it reads files, searches and runs bash; some lookups come back empty and some read files it never needed, which is time and tokens you never see. With one prompt he had Claude build a mod (Claude Code's early-access function-hooks extensions, which can add UI above the prompt) called Search Radar: it scores the wasted lookups for each question and, when a search was wasteful, offers a button that has Claude write a shortcut note into your docs so the next agent goes straight there. He tried it with two questions on his own repo; the second showed his folder layout needed work. We haven't run it, and his public claude-mods repo doesn't contain Search Radar yet, so for now you generate it from his prompt. / JJ Englert（@JJEnglert）10 月 19 日分享：Claude Code 每次在项目里找东西都会读文件、搜索、跑 bash，有的查询落空，有的读了根本不需要的文件，这些都是看不见的时间和 token。他用一段提示词让 Claude 给自己写了个 Mod（Claude Code 早期开放的 function hooks 扩展，可以在输入框上方加界面）叫 Search Radar：实时统计每个问题里的「浪费」查找，浪费时提供按钮，让 Claude 在文档里写一条捷径说明，下一个 Agent 就知道该去哪。他在自己仓库里试了两个问题，第二个暴露出目录结构该整理了。我们没有跑；他公开的 claude-mods 仓库里目前也没有 Search Radar，只能用他给的提示词自己生成。</description></item><item><title>There is no Doubao IME on Linux, so Jackson wired Doubao's streaming speech recognition into Omarchy: hold F9 and speak, live text shows at the bottom of the screen, and about a second after release it lands at the cursor; Claude Code can install it from AGENTS.md / Linux 上没有豆包输入法，坚叔把豆包的流式语音识别接进 Omarchy：按住 F9 说话，屏幕底部实时出字，松开约 1 秒上屏；也可以交给 Claude Code 按 AGENTS.md 安装</title><link>https://ttqq.dev/en/tool/omarchy-doubao-voice-dictation-linux/</link><guid>https://ttqq.dev/en/tool/omarchy-doubao-voice-dictation-linux/</guid><pubDate>Sat, 24 Oct 2026 00:00:00 GMT</pubDate><description>Developer Jackson (@Jacksonlin945) uses DHH's Omarchy (Arch + Hyprland) as his daily desktop, but ByteDance's Doubao input method, which many Chinese users consider the best Chinese dictation, has no Linux build. Instead of the IME he calls Volcengine's public Doubao streaming ASR API from a small resident Python service, doubao-dictate: hold F9 and audio streams while you talk, a floating bar shows live captions, and on release the server does a second, non-streaming pass and the text is pasted at the cursor. v0.2 is open source (MIT), and the README and AGENTS.md describe letting Claude Code or Codex do the install. He reports that for a 17-second sentence the wait after release dropped from 2.5s+ in v0.1 to 0.7–1s. We don't have an Omarchy machine or Volcengine credentials, so we only read the code and syntax-checked it. / 开发者坚叔（@Jacksonlin945）日常用 DHH 的 Omarchy（Arch + Hyprland）桌面，但豆包输入法没有 Linux 版。他没有装输入法本体，而是直接调用火山引擎开放的「豆包流式语音识别大模型」API，写了一个常驻 Python 服务 doubao-dictate：按住 F9 边录边发、悬浮条显示实时字幕，松开后服务端二遍识别，再经剪贴板粘贴到光标处。v0.2 已开源（MIT），README 和 AGENTS.md 写明了让 Claude Code / Codex 代装的流程。他自述 17 秒长句松开后出字从 v0.1 的 2.5 秒以上降到 0.7–1 秒。我们没有 Omarchy 机器和火山引擎凭据，只读了代码、做了语法检查。</description></item><item><title>The reel-live-edit Claude Code skill lets you edit by talking: say "put the ChatGPT logo on my right hand" on camera, and the effect appears at the moment you say it / Claude Code 的 reel-live-edit 技能让你边说边剪：在镜头前说「把 ChatGPT 的 logo 放我右手上」，特效就加在说到这个词的时刻</title><link>https://ttqq.dev/en/tool/reel-live-edit-claude-code-talking-head/</link><guid>https://ttqq.dev/en/tool/reel-live-edit-claude-code-talking-head/</guid><pubDate>Fri, 23 Oct 2026 00:00:00 GMT</pubDate><description>Developer vladasanadev recorded a roughly 35-second talking-head take whose script is itself a list of edit commands (logos on hands, float away, subtitles, a movie clip, background swap, a clone). Their free Claude Code skill, reel-live-edit, transcribes word timings, finds jump cuts, tracks palms and cuts the speaker out of the background, then draws graphics and synthesizes sound in code, placing each effect on the moment the command is spoken. All footage is real, not AI-generated. The author also gave the same take to GPT and stacked both edits into a 9:16 head-to-head. We have not downloaded the skill or run it. / 海外开发者 vladasanadev 录了一段约 35 秒的口播，台词本身就是剪辑指令（logo 上手、飘走、加字幕、插电影片段、换背景、分身）。作者做的免费 Claude Code 技能 reel-live-edit 会转写逐词时间、找跳剪、追踪手掌、抠人像，再用代码画图形和合成音效，按说出指令的那一刻上特效；画面全是真实拍摄，不是 AI 生成。作者还把同一素材交给 GPT 做对比，拼成 9:16 的对决视频。我们没有下载技能包，没跑过。</description></item><item><title>The open-source WeChatBridge sends selected WeChat group chats straight to Codex, Claude, Doubao or Obsidian, with a saved instruction for each group, so you no longer scroll back through history / 用开源的「微信流」把选中的微信群聊记录直接转给 Codex、Claude、豆包或 Obsidian，每个群可以预设一套整理要求，不用再往上翻记录</title><link>https://ttqq.dev/en/tool/wechatbridge-wechat-chats-to-agents/</link><guid>https://ttqq.dev/en/tool/wechatbridge-wechat-chats-to-agents/</guid><pubDate>Fri, 23 Oct 2026 00:00:00 GMT</pubDate><description>Since WeChat 4.1.13 you can multi-select messages, choose Merge Forward and then "Forward to other apps", which exports a ZIP with a TXT transcript, images and video. WeChatBridge, an open-source tool by Chinese developer Canghe (@canghe) and core contributor Xiangming (@woniu20758393), adds 12 entries to that system share menu: send to Codex, Claude, Doubao, Qwen Office, WorkBuddy and others, or save as an Obsidian note. Each group chat can be bound to a "scene" (a saved instruction prompt) and skills, and more than 100 messages can be collected in batches and sent once. It does not read WeChat's database or inject into the process. We have no Mac or Windows test machine and did not run it; we only read the permission settings in the source. / 微信 4.1.13 起支持「多选 → 合并转发 → 转发到其他应用」，导出一份含 TXT、图片和视频的 ZIP。苍何（@canghe）和核心贡献者向明（@woniu20758393）做的开源工具微信流，在这个系统分享菜单里补上 12 个入口：发给 Codex、Claude、豆包、千问办公、WorkBuddy 等，或沉淀成 Obsidian 笔记。每个群可以绑定「场景」（一段整理提示词）和技能，超过 100 条可以分批收集再一次发送。它不读微信数据库、不注入进程。我们没有 Mac 和 Windows 环境，没跑过，只读了源码里的权限配置。</description></item><item><title>Making a 41-second car ad in one afternoon with the open-source Ad Storyboarder: Opus 5.5 in Claude Code directs, Higgsfield makes the images and video, and a read-only inspector looks for faults / 用开源「广告分镜师」一个下午做出 41 秒汽车广告：Claude Code 里的 Opus 5.5 当导演，Higgsfield 出图出视频，只读审片员负责挑错</title><link>https://ttqq.dev/en/tool/ad-storyboarder-opus-higgsfield-ad/</link><guid>https://ttqq.dev/en/tool/ad-storyboarder-opus-higgsfield-ad/</guid><pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator Feicai (@zhu185178) made a 41-second self-initiated concept ad for the new Tesla Roadster. He wrote the creative script, answered a few questions, said once that "the car looks fake", and reviewed the cut; Opus 5.5 in Claude Code did the rest, generating art assets and 34 video clips through the Higgsfield CLI, using about 1,750 credits across two cuts. He released the workflow as an MIT-licensed .claude folder with six skills, a read-only inspector agent, measure-only check tools and a frame-by-frame review Studio. We did not run Higgsfield generation; we ran the free self-check, contact-sheet, loudness and delivery-gate tools on our test machine. / 废才（@zhu185178）给新款特斯拉 Roadster 做了一支 41 秒的自制概念广告：他只写创意脚本、回答几个问题、中途提醒一句「车有点假」、最后审片，其余由 Claude Code 里的 Opus 5.5 完成，通过 Higgsfield 命令行生成美术资产和 34 条视频，两版用掉约 1750 积分。他把这套工作流做成 MIT 开源的 .claude 文件夹，含 6 个 Skill、只读审片员、只量不改的检查工具和逐帧打标记的 Studio。我们没有接 Higgsfield 生成，只在测试机上跑了不花钱的自检、抽帧、响度和交付门禁工具。</description></item><item><title>Anthropic engineer Felix Rieseberg's four questions make the AI go deeper on a task instead of bigger; a developer packaged the prompt as a /better skill / Anthropic 工程师 Felix Rieseberg 用四个问题让 AI 把任务做深而不是做大，开发者把这段提示词做成了 /better 技能</title><link>https://ttqq.dev/en/tool/better-skill-deeper-task-prompt/</link><guid>https://ttqq.dev/en/tool/better-skill-deeper-task-prompt/</guid><pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate><description>On 2026-10-13 Felix Rieseberg, an engineering lead on Claude, published a tutorial on being more ambitious with AI. Its "Depth" section includes a prompt he often uses: before rewriting a task, say what a top professional would add, which rawest material they'd start from, which methods and tools they'd use, and what other than you or me would show the result is right. Developer Josh (@jjpcodes) packaged it as a /better skill for Claude Code and Codex, ran it with Opus 5.5 on a home Plex security check, and said the result was much more precise and holistic than his own prompting. He also warns that models drift into "proof porn": piling up checks without testing the real result. We have not tested it yet. / Claude 团队的工程负责人 Felix Rieseberg 10 月 13 日写了篇「怎么更有野心地用 AI」的教程，在「深度」一节给出一段他常用的提示词：改写任务前先说清顶尖专业人士会多做什么、从哪种最原始的材料入手、用什么方法和工具、除了你我之外什么能证明结果对。开发者 Josh（@jjpcodes）把它包成 Claude Code / Codex 的 /better 技能，用 Opus 5.5 跑了一个家用 Plex 安全检查，说比自己写的提示词「精确、全面得多」，也提醒模型容易陷入只堆验证不测真结果的「证明表演」。我们还没有动手测。</description></item><item><title>alswl's 1+N multi-agent practice: one captain agent dispatches the rest, rehearses the flow in a logbook, sets Delivery and Verify roles against each other, and deliberately breaks code to test the tests / alswl 的 1+N 多代理做法：一个统筹 Agent 调度其他 Agent，先用 logbook 预演流程，让交付和验证两个角色互相检查，再故意改坏代码来检验测试本身</title><link>https://ttqq.dev/en/tool/captain-logbook-multi-agent-test-the-tests/</link><guid>https://ttqq.dev/en/tool/captain-logbook-multi-agent-test-the-tests/</guid><pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate><description>On 2026-10-19 Chinese engineer alswl (@alswl) wrote up how he is moving "toward Stage 8": instead of watching each agent window, he hands a whole task to one coordinator (the Captain), which starts executor agents as needed. A logbook spells out goals, constraints and deliverables, and a Preflight rehearses the whole route before work starts; Delivery and Verify roles are set against each other; and after tests pass he deliberately breaks the tested feature to confirm the tests fail. His tools lbkit, chatta and skm are open source. He reports the captain's typical continuous run at about 30–50 minutes, up to nearly two hours. We haven't run the workflow yet. / 工程师 alswl（@alswl）10 月 19 日发文复盘自己「迈向 Stage 8」的工作方式：不再逐个窗口盯 Agent，而是把完整任务交给一个统筹者（Captain），由它按需启动执行 Agent；用 logbook 写清目标、约束和交付，开工前 Preflight 预演整条路径；让 Delivery 和 Verify 两种角色站在不同立场；测试通过后再故意拆掉被测功能，确认测试会失败。相关工具 lbkit、chatta、skm 都已开源。他自述统筹者连续运行的中间水平约 30–50 分钟、最长接近两小时。我们还没能完整运行这套流程。</description></item><item><title>Moving a whole HubSpot blog to WordPress with Claude Code: MCP on both sides, the owner went to meetings, it worked on its own for 6 hours and emailed a report when done / 用 Claude Code 把整个 HubSpot 博客搬到 WordPress：两边都接上 MCP，作者去开会，它自己跑了 6 小时，做完发邮件汇报</title><link>https://ttqq.dev/en/tool/hubspot-to-wordpress-claude-code-mcp-migration/</link><guid>https://ttqq.dev/en/tool/hubspot-to-wordpress-claude-code-mcp-migration/</guid><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><description>Wil Reynolds of Seer Interactive used Claude Code to move his company's entire HubSpot blog to WordPress. He had AI pick the blog host with the most complete MCP, ran deep research on both platforms' docs during his morning run, spent an hour-plus setting up permissions, plugins and Python, then told Claude Code he'd be away from 9:30 to 4 and it had to work alone. Around 3 pm a completion email landed in his drafts; people shown the staging site couldn't tell old from new. We have not tested this. / Seer Interactive 的 Wil Reynolds 用 Claude Code 把公司整个 HubSpot 博客迁到 WordPress。他先让 AI 帮他挑「MCP 最完整」的博客平台，晨跑时开深度研究查两边的文档，上午花一个多小时把权限、插件、Python 环境配好，告诉 Claude Code「9:30 到 4 点我不在，你自己干」。下午 3 点左右他在草稿箱收到完成邮件，给人看 staging 站分不出新旧。我们没有实测。</description></item><item><title>Making a full WeChat sticker album that passed review with the open-source weixin-sticker skill: Codex handles generation, slicing, banner/cover/icon, checks and packaging / 用开源 weixin-sticker Skill 让 Codex 做一整套微信表情包并通过审核上架：出图、切割、横幅封面图标、质检和打包都由它完成</title><link>https://ttqq.dev/en/tool/wechat-sticker-album-skill/</link><guid>https://ttqq.dev/en/tool/wechat-sticker-album-skill/</guid><pubDate>Wed, 21 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator @ZhangZhangWanAI got several AI-made sticker albums approved and listed on WeChat's sticker platform. He found that "generate 16 cute stickers" is only the start; the real time sink is the submission package: 240×240 transparent singles, a banner, cover, chat icon and tipping images, then checking sizes and counts. He turned the process into an MIT-licensed skill that Codex, WorkBuddy and other agents with image generation can run. We ran its processing, validation and packaging scripts on our test machine. / 张口说AI（@ZhangZhangWanAI）用 AI 做的几套表情包（毛毡张张、这很河狸吗等）已经通过微信表情开放平台审核上架。他发现「生成 16 个可爱表情」只是开始，真正费时间的是投稿交付：切 240×240 透明单图、做横幅、封面、聊天页图标、赞赏图，再查尺寸和数量。他把这套流程做成 MIT 开源 Skill，Codex、WorkBuddy 等能出图的代理都能用。我们在测试机上跑了它的处理、质检和打包脚本。</description></item><item><title>Three hooks, a project wiki and gates keep Claude Code from forgetting after compaction and make it prove "done" (the author's hooks are public) / 用三个钩子、一份项目 wiki 和门禁，让 Claude Code 压缩上下文后不失忆，说「做完」时要拿出证据（作者公开了钩子代码）</title><link>https://ttqq.dev/en/tool/claude-code-hooks-wiki-memory-done-gates/</link><guid>https://ttqq.dev/en/tool/claude-code-hooks-wiki-memory-done-gates/</guid><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><description>Polish front-end developer Mateusz Muszarski built his portfolio site with Claude Code in about two weeks and wrote up the setup: CLAUDE.md is a 76-line router into a 33-page project wiki; a SessionStart hook re-injects the working agreement, git state and lessons on every new session and after every compaction; a UserPromptSubmit hook scores each prompt and attaches up to 5 matching wiki pages; a Stop hook blocks once when code changed but the wiki didn't. Every correction becomes a one-line rule, a rule that slips back becomes a test, and "done" is a list of commands and screenshots that must pass. His hook scripts are public on GitHub, and we ran them on our test machine. / 波兰前端开发者 Mateusz Muszarski 用 Claude Code 大约两周做完个人作品集站，写了一篇复盘：CLAUDE.md 只做 76 行的「路由表」，指向 33 页项目 wiki；SessionStart 钩子在新会话和每次压缩后重新注入工作约定、git 状态和 lessons；UserPromptSubmit 钩子给每条提示词打分，附上最相关的最多 5 个 wiki 页；Stop 钩子在改了代码却没更新 wiki 时拦一次。每次纠正写成一行规则，规则反复被违反就改成测试；「做完」是一串必须通过的命令和截图。他的钩子脚本在 GitHub 公开，我们在测试机上跑了。</description></item><item><title>Three ways to turn a reference clip into a video prompt, tested on the same clip: sending the whole video to Gemini 3.8 Flash and then re-checking key moves was the most accurate (open-source ai-video-prompt) / 用同一条片子实测三种视频反推提示词的做法：整段视频直传 Gemini 3.8 Flash、再复查关键动作最准（开源 ai-video-prompt）</title><link>https://ttqq.dev/en/tool/video-reverse-prompt-gemini-native-video/</link><guid>https://ttqq.dev/en/tool/video-reverse-prompt-gemini-native-video/</guid><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><description>Wonfull compared three ways to reverse-engineer a prompt from the same 15-second anime clip: feed screenshots one by one, feed the whole video once, or feed the whole video and then re-check the key moves. Only the third got both shots and every key action right. He packaged it as the open-source skill ai-video-prompt: in Cursor or Codex, say "video-to-prompt + path"; Gemini 3.8 Flash (via OpenRouter) watches the video, DeepSeek fact-checks and translates, and you get a bilingual Chinese/English prompt.md. We ran the pipeline's evidence step, which calls no model, on our test machine. / Wonfull 用同一条 15 秒动漫短片对比了三种视频反推提示词的做法：拆成截图一张张喂、整段视频一次喂、整段视频喂完再对关键动作多看几遍。只有第三种把两个镜头和每个关键动作都还原对了。他把第三种做成开源 skill ai-video-prompt：在 Cursor / Codex 里说一句「视频反推提示词 + 视频路径」，Gemini 3.8 Flash（经 OpenRouter）看视频，DeepSeek 审核和翻译，交出中英双语 prompt.md。我们在测试机上跑了流水线里不调用模型的取证环节。</description></item><item><title>Turning The Little Prince into an explorable 3D game with Opus 5.5, Tripo, Blender and Unity: AI makes the assets and also drives the engines / 用 Opus 5.5、Tripo、Blender 和 Unity 把《小王子》做成可探索的 3D 游戏：AI 出素材，也由 AI 操作引擎</title><link>https://ttqq.dev/en/tool/opus-tripo-blender-unity-3d-game/</link><guid>https://ttqq.dev/en/tool/opus-tripo-blender-unity-3d-game/</guid><pubDate>Mon, 19 Oct 2026 00:00:00 GMT</pubDate><description>Sac spent 16 days turning The Little Prince's journey between planets into a 3D story game you can play in the browser, then published the workflow: Fable 5.1 writes the PRD, GPT Image 2.5 makes scene and character references, Tripo generates 3D assets, Opus 5.5 builds scenes, rigs characters and wires animations through the Blender CLI and the official Unity MCP, and Astra fixes bugs. Reusable prompts are included for each step. / Sac 花 16 天把《小王子》的星际旅行做成了浏览器里能玩的 3D 剧情游戏，随后公开了制作流程：Fable 5.1 写 PRD，GPT Image 2.5 出场景和人物参考图，Tripo 生成 3D 资产，Opus 5.5 通过 Blender CLI 和官方 Unity MCP 搭场景、绑骨、接动作，Astra 修 bug。文中附了每一步可直接复用的提示词。</description></item><item><title>Keeping secrets out of the AI chat when driving a coding agent from your phone: Simon Willison had Codex build web-edit and edits those files himself over Tailscale / 用手机遥控编程代理时不让密钥经过 AI：Simon Willison 让 Codex 写了 web-edit，自己通过 Tailscale 改文件</title><link>https://ttqq.dev/en/tool/web-edit-secrets-off-llm-tailscale/</link><guid>https://ttqq.dev/en/tool/web-edit-secrets-off-llm-tailscale/</guid><pubDate>Mon, 19 Oct 2026 00:00:00 GMT</pubDate><description>While working from his phone with Codex Remote, Simon Willison needed to edit a credentials file without round-tripping the credentials through the LLM. He had Codex (GPT-6 Astra Extra High) build web-edit, a small Python web app: the agent starts it on the machine and he edits files himself from his phone over Tailscale. He later added copy-to-clipboard to move logs from remote runs into a Gist. We ran version 0.1a2 on our test machine. / Simon Willison 在手机上用 Codex Remote 干活时，需要改一个凭证文件，又不想让凭证在大模型对话里来回传。他让 Codex（GPT-6 Astra Extra High）写了一个小 Python 网页应用 web-edit：让代理在那台机器上启动它，自己再用手机通过 Tailscale 打开、亲手改文件。后来又加了复制到剪贴板，方便把远程跑出来的日志贴进 Gist。我们在测试机上跑了 0.1a2 版。</description></item><item><title>Turn 1,270 of your own coding-agent sessions into a private eval set and find a setup at half the cost / 把自己 1270 条编程代理会话变成私人评测集，找到便宜一半、照样能解题的配置</title><link>https://ttqq.dev/en/tool/agent-sessions-to-personal-evals/</link><guid>https://ttqq.dev/en/tool/agent-sessions-to-personal-evals/</guid><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><description>Ibragim, who maintains SWE-rebench, had an agent read three months of his Claude Code / Codex / Pi sessions, picked episodes where the agent failed and he corrected it, turned them into 10 Harbor tasks with automated checks, and ran different models and harnesses. On his tasks, Pi + GPT-6.1 Sol solved everything at about half Codex's cost per task with the same model. / SWE-rebench 维护者 Ibragim 让代理读自己三个月的 Claude Code / Codex / Pi 会话，挑出「代理先犯错、被他纠正」的片段做成 10 个带自动校验的 Harbor 任务，再拿不同模型和代理跑。结论：在他的任务上，Pi + GPT-6.1 Sol 全部解出，每题成本约是 Codex 跑同一模型的一半。</description></item><item><title>The split of work after ~500 generations: Doubao Work plans the shots, ChatGPT draws first frames, Meta Muse renders each 10-second shot, and they are joined into a 3-minute film / 生成近 500 条视频后总结出的分工：豆包 Work 写整片分镜，ChatGPT 出首帧，Muse 逐镜生成，最后拼成 3 分钟长片</title><link>https://ttqq.dev/en/tool/muse-doubao-work-chatgpt-long-video/</link><guid>https://ttqq.dev/en/tool/muse-doubao-work-chatgpt-long-video/</guid><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><description>Jarvis ran 488 Meta Muse video generations and concluded that unstable long videos usually come from unsettled prompts and first frames. His final split: Doubao Work researches viral videos on the theme plus Seedance's prompting tips and writes a 10-seconds-per-shot plan, ChatGPT draws a first frame per shot, and Muse renders each clip from its frame and stitches them. / Jarvis 用 Meta 的 Muse 测了 488 条视频，结论是长视频不稳多半是提示词和首帧没定好。最后固定的流程：豆包 Work 先查同题材爆款和 Seedance 官方技巧写出 10 秒一镜的整片方案，ChatGPT 按方案逐镜出首帧，Muse 按首帧生成视频并拼接。</description></item><item><title>A 7-minute hand-drawn watercolor animation made without drawing skills: GPT writes the script, Claude builds it with a skill, and Doubao TTS does the voice / 不会画画也做出 7 分钟手绘水彩动画：GPT 写剧本，Claude 用 skill 制作，豆包语音配音</title><link>https://ttqq.dev/en/tool/storybook-animation-opus-doubao-tts/</link><guid>https://ttqq.dev/en/tool/storybook-animation-opus-doubao-tts/</guid><pubDate>Sun, 18 Oct 2026 00:00:00 GMT</pubDate><description>Hailey polishes a script and storyboard with GPT in three rounds, has Opus 5.5 build a self-drawing watercolor animation in the browser using her open-source hand-drawn-storybook-animation skill, then adds narration and character voices with Volcengine's Doubao TTS and synced bilingual subtitles. She says it took one day. / Hailey 用 GPT 分三轮打磨剧本和分镜图，再让 Opus 5.5 按她开源的 hand-drawn-storybook-animation skill 在浏览器里「画」出水彩动画，最后接火山引擎豆包语音配旁白和角色对白、对齐双语字幕。她说前后只花了 1 天。</description></item><item><title>Three prompts to have Claude Code build a LaTeX tool that lives entirely in the URL #fragment / 三句提示词让 Claude Code 做一个「内容存在网址 # 后面」的 LaTeX 分享工具</title><link>https://ttqq.dev/en/tool/claude-code-url-fragment-latex-tool/</link><guid>https://ttqq.dev/en/tool/claude-code-url-fragment-latex-tool/</guid><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><description>Inspired by Tiny Little Office, Simon Willison used three Claude Code prompts to build a single-page LaTeX renderer whose (optionally compressed) content lives in the URL #fragment, so the link is the document; follow-ups added an iframe sandbox against XSS and URL/Gist loading. / Simon Willison 受 Tiny Little Office 启发，用 Claude Code 三轮提示做出单页 LaTeX 渲染器：内容（可压缩）放在 URL 的 #fragment 里，链接即文档；第二、三轮补上 iframe 沙箱防 XSS 和从 URL / Gist 加载。</description></item><item><title>Shipping an app without coding: Claude designs and reviews, a cheaper model writes the code in OpenCode, and scripts gate the tests / 不会写代码也把 App 上线：Claude 负责设计和审查，OpenCode 里的便宜模型写代码，脚本门禁检查测试</title><link>https://ttqq.dev/en/tool/claude-opencode-dev-workflow/</link><guid>https://ttqq.dev/en/tool/claude-opencode-dev-workflow/</guid><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><description>Chinese builder @niukutzu open-sourced the workflow behind an Android app (~70 task packages, shipped): Claude only designs, splits tasks and reviews; ChatGPT/DeepSeek inside OpenCode write the code; Bash scripts gate on real exit codes and bounce failures back for fixes. / 中文作者 @niukutzu 把自己做安卓 App（约 70 个任务包、已上正式版）的分工流程开源成模板：Claude 只做设计、拆任务和审查，OpenCode 里的 ChatGPT / DeepSeek 写代码，Bash 脚本按真实退出码跑门禁，没过自动退回修。</description></item><item><title>Enforce agent conventions with a pre-commit hook that fails the commit and tells the agent how to fix it, instead of endlessly editing AGENTS.md / 用 pre-commit 钩子拦下代理违反约定的写法，并把修复说明交回给它，不用一直改 AGENTS.md</title><link>https://ttqq.dev/en/tool/pre-commit-hook-agent-guardrail/</link><guid>https://ttqq.dev/en/tool/pre-commit-hook-agent-guardrail/</guid><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><description>Geoffrey Huntley (of Ralph loop fame): when the agent skips your diagnostics library, don't edit AGENTS.md — add a pre-commit hook that scans the diff, fails the commit on bypasses, and prints a fix instruction for the agent. He says it then goes "ten out of ten". / Geoffrey Huntley（Ralph loop 提出者）的做法：代理不用你的诊断库时，不去改 AGENTS.md，而是写个 pre-commit 钩子检查 diff，发现绕过就让提交失败，并输出一句给代理的修复提示。他称之后「十次十中」。</description></item><item><title>Export a Claude Code session with /export for the next agent or a custom skill / 用 Claude Code 的 /export 把整段会话导出成文本，给下一个代理或写成 skill</title><link>https://ttqq.dev/en/tool/claude-code-export-session/</link><guid>https://ttqq.dev/en/tool/claude-code-export-session/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>Claude Code's built-in /export saves the current conversation as readable plain text (optional filename). @dani_avila7 highlights two uses: hand the transcript to another agent, or mine it to author better custom skills. / 官方内置 /export：把当前对话导出为可读纯文本（可指定文件名）。X 用户 @dani_avila7 指出用途：交给另一个代理续做，或分析会话来写改进用的自定义 skill。</description></item><item><title>A design-discussion prompt that makes coding agents reason before they write code / 一段设计讨论提示词，让编程代理先想清楚再写代码</title><link>https://ttqq.dev/en/tool/design-reasoning-prompt-before-code/</link><guid>https://ttqq.dev/en/tool/design-reasoning-prompt-before-code/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>Vinh Nguyen, who builds the open-source Rust coding agent VT Code, shared the prompt he uses for design discussions: review the codebase, challenge assumptions, compare options, then recommend the simplest practical approach. / 开源 Rust 编程代理 VT Code 的作者 Vinh Nguyen 分享的设计讨论提示词：先通读代码库、质疑假设、比较方案，再推荐最简单可行的做法。</description></item><item><title>Make an LLM output a real Excel file with a Dify workflow (open-source DSL from China's 嗯哌AI) / 用 Dify 工作流让大模型直接生成 Excel 表格（嗯哌AI 开源 DSL）</title><link>https://ttqq.dev/en/tool/dify-llm-to-excel-workflow/</link><guid>https://ttqq.dev/en/tool/dify-llm-to-excel-workflow/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>Chinese creator 嗯哌AI (Banny) open-sourced 30+ Dify workflows (≈2.7k GitHub stars). Excel_Flask_Dify pairs Dify with a tiny Flask service that turns model-generated fields and rows into an .xlsx file. / 中文社区作者嗯哌AI 开源了 30 多个 Dify 工作流。Excel_Flask_Dify 用一个小 Flask 服务把模型生成的字段和数据写成 .xlsx。</description></item><item><title>Matt Pocock's /grill-me skill: the AI questions you about the plan before it writes code / Matt Pocock 的 /grill-me 技能：写代码前先让 AI 把你的需求问一遍</title><link>https://ttqq.dev/en/tool/grill-me-plan-interview/</link><guid>https://ttqq.dev/en/tool/grill-me-plan-interview/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>Most bad agent output starts with a misread request. /grill-me makes the agent interview you round by round over a decision tree, with a recommended answer for each question, before any work starts. / 大多数糟糕的 AI 产出源于它误解了你的需求。/grill-me 让编程代理先按“决策树”逐轮提问、给出推荐答案，问清楚了再动手。</description></item><item><title>Cursor engineer poteto's pstack: /poteto-mode makes the agent verify before it edits code / Cursor 工程师 poteto 的 pstack：用 /poteto-mode 让代理先验证再改代码</title><link>https://ttqq.dev/en/tool/pstack-poteto-mode/</link><guid>https://ttqq.dev/en/tool/pstack-poteto-mode/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>Lauren Tan (@poteto, Cursor / React Compiler) open-sourced pstack — the skills she uses daily. Entry point /poteto-mode matches one of 23 playbooks, routes supporting skills, and prioritizes less code, runtime proof, and safe parallelism. / Lauren Tan（@poteto，Cursor / React Compiler）开源了日常用的技能包 pstack。入口是 /poteto-mode：匹配 23 套 playbook，按步骤调用子技能，强调少写、可验证、可并行。</description></item><item><title>Running Xiaohongshu (RED) writing, layout, posting and replies as a pipeline with OpenClaw and an open-source skill / 用 OpenClaw 和开源 skill 把小红书的写稿、排版、发布和回评论做成流水线</title><link>https://ttqq.dev/en/tool/xiaohongshu-ops-openclaw/</link><guid>https://ttqq.dev/en/tool/xiaohongshu-ops-openclaw/</guid><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><description>A Chinese open-source skill (Xiangyu-CAS/xiaohongshu-ops-skill, ~2.4k stars) wires topic research, analysis, publishing, comment replies and viral-post remix into OpenClaw. Creator 程序员小饭 documented a 14-day, three-account ops rhythm and the pitfalls. / 中文社区开源 skill（Xiangyu-CAS/xiaohongshu-ops-skill，约 2.4k star）把选题、分析、发布、回评、爆款复刻接到 OpenClaw。创作者「程序员小饭」写了 14 天一人三号的实测节奏与坑。</description></item></channel></rss>