RSS
Built with Claude

cook: an agent skill that iterates visual builds until they match the references

cook is an agent skill for visual builds that have to match reference images or videos. It gives the agent a target, a numeric bar and time, then loops through build, compare, critique and fix.

Person holding a photo of an elephant, illustrating the cook skill that matches builds to references
Photo by Thiago BF on Unsplash

Leonxlnx/cook gives an AI coding agent a clear target, a numeric bar and a time budget, then lets it work through long rounds of building, comparing its work with reference images or videos, and fixing the weakest part until the work matches.

What is cook?

cook is an agent skill for visual builds that have to match reference images or videos.

It runs the Cook Loop: one intake, then a long autonomous loop of build, stills, fresh-context critics and fixes until the work holds up next to the references. The README describes it for ambitious visual builds such as 3D worlds in code, films, launch videos, games and user interfaces.

References are the minimum bar, not inspiration. Every reference becomes a shot with notes on what to take from it, and motion references are analysed frame by frame. The agent keeps all run state in a cook/ folder in the project, so a run can resume after a context reset or be handed to another agent.

Who is cook for?

cook is for developers who use Claude Code or Codex CLI and want an agent to carry a long visual build while they are away.

It fits 3D worlds made in code with three.js, WebGL or shaders, code-rendered films, launch and motion videos, games and polished user interfaces that must hold up next to references. The skill is made for runs of 10 to 40 hours or more, and it also works with the unlazy skill for effort and completion discipline and with any goal, loop or long-run mode the agent offers, although cook works without them.

Is cook free and open source?

cook is free and open source under the MIT License.

  • MIT License, with copyright held by Leon Lin in 2026.
  • About 700 GitHub stars as of November 18, 2026.
  • GitHub lists JavaScript as the repository's main language.
  • As of November 18, 2026, the repository lists no releases.
  • The side-by-side script needs Node 18 or later and ffmpeg on the PATH; everything else is plain Markdown.

What stands out

The loop is built around a target, a measured bar and a gate the user controls.

  • One intake, then no more questions: the agent asks every open question once, each with a default, and then works on its own, so it can run through the night.
  • A numeric bar per dimension: composition, light, color, materials, detail near and far, and motion, with 99.9% as a stretch target for stills and at least 9 out of 10 for video.
  • Fresh-context critics: subagents with no history score the side-by-side images, the median score is used, and the agent's own score is logged but never decisive.
  • A gate after every round: stills go to the user after every round without pausing, and no final render starts before the user approves the stills in chat.
  • Resumable state: the brief, shot list, round ledger, references, stills and comparisons live in a cook/ folder, so a run can continue after a context reset or move to another agent.

How do I install cook?

Install cook with the skills CLI, then invoke it as /cook in Claude Code or $cook in Codex.

From the project's README at https://github.com/Leonxlnx/cook/blob/main/README.md, use the skills CLI for supported agents:

npx skills add Leonxlnx/cook

For a manual install, the README lists ~/.claude/skills/cook for Claude Code and ~/.codex/skills/cook for Codex CLI. Invoke it as /cook in Claude Code or $cook in Codex, and it also triggers on phrases such as go cook, iterate until it matches, or match the references 99%.

To start a run, attach reference images and send an example from the README:

/cook build this ancient greek world in three.js, match the 9 attached references

The README also gives a command to build a comparison image yourself, with the reference on the left and the render on the right, one row per shot. That script needs Node 18 or later and ffmpeg on the PATH.

node scripts/side-by-side.mjs --out compare.jpg ref1.png render1.png ref2.png render2.png

How it was built with Claude

The repository's commit that added the skill carries a Co-Authored-By trailer naming Claude Opus 5.5.

The repository's file tree contains no CLAUDE.md, AGENTS.md or .claude folder.

Try it with your AI. Try cook on a small visual build first, using only the project's own commands and interfaces.

Install cook with npx skills add Leonxlnx/cook, then open Claude Code or Codex CLI and
invoke /cook or $cook. Tell the agent: [what to build, in one sentence, and the stack,
for example three.js]. Attach [your reference images or videos] and set the bar as
[99.9% for stills or at least 9 out of 10 for video]. Set a time limit of [a floor of 8
hours for a demo, or a hard cap of 1.5 hours for a small job]. Use only the official
commands and interfaces named in the README. Show me any risky step first, including
downloads, pushes, publishing and final renders, and wait for my confirmation before you
run it. Do not install anything from unknown sources, do not touch credential files, and
do not delete anything. Ask every open question once, with defaults, then work on your
own and send me stills after every round. Do not start the final render before I approve
the stills.

What are cook's limits?

  • Demos have a time floor of at least 8 hours, and finishing after 2 hours means something is wrong; small jobs get a hard cap instead.
  • The agent must never use 100% of a usage window, and the README paces use at about 20% of a 5 hour window per hour and about 99% per window.
  • By default, no external assets are used: every mesh, texture, shader and effect is generated in code, unless the brief relaxes that rule, and anything external that is used must be listed in the credits.
  • References must not be copied element by element: the skill says to take from many references and never copy one element 1:1, and never to name third-party reference clips in public files.
  • Publishing is limited: only references the user made or owns may be published, a public repository is prepared only if the brief allows pushing, and the cook/ working folder is published only if the user wants it to be.
  • Before anything goes public, the skill calls for a privacy scan for API keys, tokens, private keys, personal paths, emails and private names, and says a key pasted in a prompt must be removed and the user told to rotate it.
  • Sound has its own limit: code-made sound is used only if it truly holds up, otherwise licensed free audio is used with credits, and if no listening skill is available the sound is stated plainly as not reviewed by ear.

Problems and feedback

Report problems and give feedback on the Issues page for Leonxlnx/cook at https://github.com/Leonxlnx/cook/issues.

Source: Leonxlnx/cook repository, README, SKILL.md, references/intake.md, references/target-and-bar.md, references/loop-and-critics.md, references/rules.md, references/ship.md, LICENSE, Issues, Releases, skills CLI

This story first appeared in the claudefox daily briefing for November 19, 2026.