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
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.
Visit site Claude + Codex (oil-delegate) + Git Worktree ↗- Pricing
- Paid
- Region
- Global
- Where it's from
- Chinese community
- Added
- Nov 14, 2026
- Tested on
- Nov 14, 2026
- Version we used
- oil-oil/oil-delegate a4e42de; author used Claude Opus 5.5 + Codex (GPT)
- Categories
- CodingProductivity
- Original author
- 欧呦 (oil-oil) @I_am_oil_oil ↗
- Our test
- Partly tested
- Cost / time
- The process is free, but you need both Claude and ChatGPT/Codex plans; he burns over a billion Codex tokens a day. oil-delegate and oil-ui are free and open source. The second half of the article is paid Oil Book content.
The task
Keep developing a feature-heavy app with AI over months, with a repo that stays maintainable and new features that don't get harder to add.
Steps
- Before coding, have AI research online: what existing open-source projects use and their licenses (MIT or Apache allow commercial forks; avoid GPL and AGPL), how cross-platform frameworks compare with native code on interaction, performance and maintenance, and how well it knows the language. This is why he chose native Swift over Electron or Tauri.
- Have several strong models cross-check the plan, then let a strong planner (he used Claude Opus 5.5) split the app into milestones that can each be verified on their own.
- Daily split: Claude leads the conversation and writes human-readable plans, keeping the main context clean; oil-delegate sends critique, coding and edge-case review to Codex.
- Develop several features in parallel with Git Worktree; separate logic tests from UI tests, wrap frequent debugging actions in CLI scripts that can run silently in the background, and periodically delete padding tests.
- Write recurring problems into AGENTS.md and Review.md, and use oil-ui to keep interaction and styling consistent.
- Periodically have AI check the docs against the code and delete or update stale ones.
- For stubborn bugs, have two strong models investigate separately and state their confidence; replace features that are low-confidence and costly to maintain. Gate releases with scripted checks on tests and version numbers, and add a crash-triggered safe mode that guides users to update.
Prompt
I want to build a Mac screen recording and editing app, focused on interaction quality, the ceiling on rendering performance, and long-term maintainability. Please research current mainstream approaches and open-source projects online: 1. List the tech stacks and licenses of existing open-source projects, assess whether any MIT-licensed project is suitable to fork, and compare the effort of adapting it with building from scratch; 2. Compare cross-platform frameworks and native languages on interaction design, rendering performance and long-term maintenance, and tell me how well you know each language; 3. Explain the trade-offs in plain language so I can decide.
What the author says
Oil @I_am_oil_oil (X video post, 12 Nov 2026; the long article "How to develop a complex project with AI over the long term" is on Oil Book, with the second half for paying members): built Oil Studio over one to two months in Swift he didn't know. Codex, called through oil-delegate, writes and reviews code, "burning over 1 billion GPT tokens a day while saving Claude quota". Product design and task assignment go to Claude Opus 5.5. Native Swift beats cross-platform frameworks on Mac for interaction and rendering speed. Wrong docs are worse than no docs.
Our test
On 14 Nov 2026 on Linux (Python 3.13.5) we cloned oil-oil/oil-delegate (a4e42de). python3 -m unittest discover -s tests: 57 tests, 56 passed, 1 error (test_sessions_claude_skips_symlinks_cloud_directories_and_placeholders raised a TypeError inside pathlib's is_symlink; we didn't work out whether the test's mocking is incompatible with this Python version or the code is at fault). scripts/delegate.py --home /tmp/odh onboard returned NEEDS_ONBOARDING after probing Claude Code, WorkBuddy, Codex and others; none are installed on our box and it reported each as not installed, without calling any model. We did not test Oil Studio itself, the Claude/Codex split, parallel worktrees, doc audits or the release process.
What we found when we tried it:oil-delegate's onboarding only detects tools, without calling models or signing in, as its README says. The long-term workflow itself is the author's experience; we didn't run it.
How to get in
The original is a Chinese X post with a video, plus a Chinese article on Oil Book (https://book.oiloil.org/articles/agent-long-project-management), whose second half is for members. oil-delegate's README is in Chinese: https://github.com/oil-oil/oil-delegate
Who it suits
Solo developers building an app with AI that will evolve for months, possibly in a language they don't know well, especially those already using both Claude and Codex.
Pitfalls
- This is one developer's experience on one project with no comparison data; verify it on yours.
- A billion tokens a day needs high-tier plans; a basic plan can't sustain this split.
- The second half of the article (testing, governance and release details) is paywalled.
- oil-delegate only forwards tasks to agent CLIs already installed and signed in on your machine; it provides no model itself.
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- AI backend code that guards against races, leaks and bad migrationsCodingbackendskillsClaude CodePostgreSQLFree; about 2× tokens per task◐ Partly testedOverseas
- An afternoon of AI work, a 3D action game you play in the browserCodingClaude CodeTripoJimengSkill free; Claude quota plus asset tools◐ Partly testedChina
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas
- Scan a skill before installing, then read the flagged linesCodingSkillSpectoruvFree; static scan needs no model● Tested by usChina