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
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.
Visit site /better skill (Claude Code / Codex, any chat model) ↗- Pricing
- Free
- Region
- Global
- Where it's from
- Overseas
- Added
- Oct 22, 2026
- Version we used
- Prompt from Felix Rieseberg's post (2026-10-13); /better SKILL.md gist by joshp123 (created 2026-10-19, updated 2026-10-21)
- Categories
- ProductivityCoding
- Our test
- Not tested by us
- Cost / time
- Free. It's just a prompt / one SKILL.md; the extra cost is the tokens and time the model spends doing the deeper version.
The task
Get a coding agent or chat model to stop at more than a surface-level job: work the way a field's best practitioners would, start from primary material, and define success criteria that don't rely on the model grading itself.
Steps
- Simplest use: paste the prompt below after a task you're about to give an AI (or as its own message). It answers the four questions first, then rewrites your prompt into a deeper version; approve that, then have it run the new version.
- As a skill: download SKILL.md from Josh's gist (gist.github.com/joshp123/a1443d30fa1b7c577e74882895cea61b) into ~/.claude/skills/better/ or ~/.agents/skills/better/. It only triggers when you explicitly type /better (or $better).
- The skill handles two cases. A) You gave it a prompt you haven't started: it answers (1)–(4) and returns the rewritten prompt without running it. B) Work is already in progress (/better on its own, or with feedback like "not good enough" or "too shallow"): it answers the four questions and rewrites the brief as the original request plus your feedback, deepened.
- In case B, if the new brief changes the goal, widens the scope, or needs a decision, money or access only you can give, it shows you and waits; otherwise it redoes the work: starting from the raw material in answer (2), using the methods and tools in (3), and running the check in (4) before reporting what it showed.
- Josh's example: checking whether his home Plex server was secure. Most of the improvement after /better came from the "what would a top security engineer do" question.
- Felix's post pairs it with other tactics: insist on the instruments the best practitioners use (databases and statistical models, not markdown tables); give the model raw data and have it adversarially test your assumptions; insist on outside validation rather than letting the model grade itself, ideally a tool it can run for immediate feedback.
Prompt
I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I'm not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that.
What the author says
In his 2026-10-13 post Felix calls it "a simple prompt that often helps me", notes that when prompted to act like a professional, models often switch vocabulary without switching methods, and, as of October 2026, implores readers to insist on outside validation instead of letting the model judge itself. On 10-19 Josh posted that it's "the most useful prompt I've found in months" and works "to great effect" as a /better skill with Opus 5.5: the top-professional question reminds the model to use best practices it knows but doesn't apply; "rawest material" stops it from leaning on proxies (e.g. Claude sampling logs instead of reading them); forcing independent success criteria gives stronger results. His before/after on a Plex security check says the /better version was "much more precise and holistic". The post had about 980 likes and 1,790 bookmarks.
Our test
We have not yet installed /better in Claude Code or Codex for a with/without comparison, so "more precise and holistic" is the authors' claim. We checked that the prompt text matches between Felix's post and the gist, and read the SKILL.md trigger and pause rules.
How to get in
Felix's blog and Josh's GitHub gist are public; the X posts need a login. The prompt isn't tied to any model.
Who it suits
People already using Claude Code, Codex or a chat model for real work whose results feel usable but shallow, especially for tasks outside their own expertise (security checks, data analysis, business reviews).
Pitfalls
- "Proof porn": Josh warns that once forced to define success criteria, the model may endlessly hash and compare without testing the real result. He says all frontier models as of October 2026 love doing this, so keep an eye on it.
- Deeper, not bigger, but the rewritten task can still get much heavier and costlier. The skill pauses when the goal or scope changes or money/access is needed; if you paste the raw prompt, that judgment is on you.
- The rawest material may not fit in context: a reply asked what happens when a full log won't fit, and neither the post nor the skill says whether to read in chunks or flag a partial read. Add that instruction yourself.
- Felix's approach assumes the model has a computer that can install software and reach the relevant data (MCP connectors etc.). In a plain chat window, answers (2) and (3) often stay on paper.
- Attribution: the prompt text comes from Felix's post; Josh's gist just says "Core prompt supplied by Josh, verbatim", and notes that the "How to apply it" section was written by Opus 5.5.
- The evidence is the authors' own accounts and one screenshot, with no public controlled comparison, and we haven't tested it yet.
More in this category
- Keep a long-running AI-built project from turning into a messCodingClaudeCodexoil-delegateNeeds Claude and Codex plans◐ Partly testedChina
- Have fresh-context AI reviewers poke holes before you codeCodingEGMCodexClaude CodePlugin free; reviews use your model quota◐ Partly testedOverseas
- 1,050 memory recalls audited: what makes agent memory pay offCodingNowledge MemCodexClaude CodeFree tier: 50 memories; review uses model quota○ Not tested by usChina
- Isolate an always-on agent on its own machine, then red-team itAutomation & AgentsOpenClawClaude CodeOllama$20/mo Ollama Cloud + a Claude plan○ Not tested by usOverseas
- A to-do list beside your agents, one AI chat per taskProductivityToday BoardcmuxClaude CodeFree, open source; needs macOS + cmux◐ Partly testedChina