name: cook description: Runs the Cook Loop for ambitious visual builds that must match reference images, screens or videos through long autonomous iteration. Use it for 3D worlds made in code (three.js, WebGL, shaders), code-rendered films, launch and motion videos, games and polished UI that must hold up next to references, especially runs of many hours or days while the user is away. Trigger when the user types /cook or $cook, says "cook", "go cook", "let it cook", "iterate until it matches", "match the references", "99%", "99.9%", "9/10 minimum" or "at least 8h", or attaches reference images and asks for a recreation in code. It runs one intake, then loops build, stills, zero-context critics and fixes against a numeric bar, and holds the final render until the user approves.

Cook

Give your agent a target, a bar and time, then let it cook.

This skill runs the Cook Loop: one intake, then a long autonomous loop of build, render, compare, critique and fix until the work holds up next to the references. It is made for runs of 10 to 40 hours or more. The references carry the look. The brief sets the bar, the rules and the time.

Optional companions: /unlazy for effort and completion discipline (the Depth Tree), and any goal, loop or long-run mode the harness offers. Cook works without them.

State files

All run state lives in cook/ at the project root. These files are your memory across context resets and the handoff to any other agent.

Start or resume

• cook/BRIEF.md exists and the run is not shipped: you are resuming. Read BRIEF, SHOTS and the last two ROUNDS entries, then start the next round (step 6). Do not re-ask, restart or rebuild what works.
• After every context compaction, do the same before your next action.
• cook/ holds a shipped run: move it to cook-archive/<date>/ and start at step 1.
• No cook/ folder: start at step 1.

The Cook Loop

1. Intake (once)

Ask every open question in one message, each with a default, so the user can answer by number or reply "defaults". Skip what the prompt, the attachments or a filled templates/brief.md already answer. If nothing is open, say so and start. Questions and defaults: references/intake.md.

Collect: what to build, references, bar, time budget, hard rules, deliverable, hardware, usage limits, and what you may download or publish.

Then write cook/BRIEF.md and cook/ROUNDS.md and start. From now on, never ask a question. The user may be asleep. Decide, log the decision under Decisions in ROUNDS.md, and keep working.

2. Target

References are the minimum bar, not inspiration.

• Study every reference at full size before building anything.
• Write cook/SHOTS.md: one shot per reference image, with camera, framing, time of day and what to take from it (light, color, composition, materials, timing).
• Motion references: analyze them frame by frame into a list of effects, transitions and timings. Take from many, never copy one element 1:1.
• Getting references for 3D, UI and motion, and the analysis steps: references/target-and-bar.md.

3. Bar

Write a number per dimension into BRIEF.md: composition, light, color, materials, detail density (near and far), motion.

• Stills: a stretch target such as 99.9% similarity. Video: a score out of 10. References at about 8.5 mean a target of 9 or more.
• Models rate themselves too high. When the user gives an estimate, log the gap as the calibration offset and apply it to every later score.
• Keep measured numbers (a command produced them) apart from eyeballed ones (critic scores).

4. Rules

Defaults unless the brief says otherwise. Full rules: references/rules.md.

• Zero external assets: every mesh, texture, shader and effect in code, when the point is the flex.
• Sound: code-made only if it truly holds up. Otherwise licensed free libraries, credited.
• Launch videos: real logos and real photos beat drawn SVG people.
• Slop ban list: eyebrow labels, monospace uppercase text, random serif fonts, purple gradients, beige, 01/02/03 counters, glowing orbs, shockwave rings popping out of things, decorative particles, em dashes.
• Color matches the references: never grey or washed out, never oversaturated.
• Weak GPU: keep realtime cheap and get the quality from offline rendering (sub-frame accumulation for motion blur, soft shadows, depth of field, anti-aliasing).

5. Time

• Demos get a floor: at least 8h, and done after 2h means something is wrong. Big builds can take days. Small jobs get a hard cap.
• Log the start time. Read the clock at every round boundary.
• Pace usage: about 20% of a 5h window per hour, about 99% per window, never 100%.
• Depth over breadth: make one shot, scene or element exceptional before you expand.

6. Loop

Before round 1, build the harness: a fixed camera preset per shot, a still renderer and a color check. For side-by-side images use scripts/side-by-side.mjs from this skill (Node 18+ and ffmpeg): node <skill>/scripts/side-by-side.mjs --out cook/compare/r01/all.jpg cook/refs/R01.png cook/stills/r01/R01.png .... Every round reuses the harness.

Each round:

1. Focus: open user feedback first, else the weakest dimension in the weakest shot.
2. Fix it in depth. Split independent parts across parallel subagents.
3. Render stills from the shot presets, same framing and aspect as each reference. For motion, render contact sheets and short clips.
4. Make side-by-side images: reference left, render right.
5. Measure what can be measured: color per region, frame time, seconds per rendered frame.
6. Score with fresh-context critics: subagents with zero context that get only the side-by-side images and the rubric, optionally from another model. Your own score is logged, never decisive.
7. Log the round in cook/ROUNDS.md.
8. Send the stills to the user (Gate) and keep going.
9. Run the Stop check. If it fails, start the next round now.

Never stop when it first works. Never claim a check that was not run. If something was not measured, write "not measured". Round details, critic prompts and pacing: references/loop-and-critics.md.

7. Gate

• After every round, send the stills with a short note: what changed, scores, next focus. Do not pause, and do not end your turn to wait for a reply.
• User feedback beats critic priorities. Log the exact words and address them in the next round.
• Never start the final render, or any other long and costly export, before the user approves the stills in chat. While you wait, keep refining and test the render pipeline on a few short frames.

8. Ship

After approval, render the final deliverable and check the file. For a video on X: the first frame is the thumbnail and must be impressive, best scene first, a new scene every 6 to 7 seconds, no text on screen, licensed music credited, 1080p 30fps is enough. For a public repo: README, credits, the real prompt and honest numbers. Push or publish only if BRIEF.md allows it. Details: references/ship.md.

Stop check

Stop only when all of these are true:

• Every dimension in every shot meets the bar, by critic median after calibration.
• The time floor is reached (capped jobs have no floor).
• The user approved the stills, and the final deliverable is rendered and its file checked.

Done before the floor means the bar was read too low. Reread the references at full size, check far distance, color and the weakest shot, and start another round. If a hard cap or a usage stop comes first, finish the round, log it and report plainly what is still below the bar.

Never

• Ask a question after intake.
• Start the final render before the user approves the stills.
• Copy one reference element 1:1, or name third-party reference clips in public files.
• Let your own score decide a stop.
• Report a number you did not measure or a check you did not run.
• Use 100% of a usage window.
• Put API keys, tokens or private information in anything you publish.

Files in this skill

cook

Give your agent a target, a bar and time, then let it cook.

An agent skill for ambitious visual builds that have to hold up next to references: 3D worlds in code, films, launch videos, games and UI. It runs the Cook Loop, one intake and then a long autonomous loop of build, stills, fresh-context critics and fixes until the work matches.

Install | Quick start | The Cook Loop | Made with this

Install

Use the skills CLI for supported agents:

npx skills add Leonxlnx/cook

Manual locations:

Claude Code:  ~/.claude/skills/cook
Codex CLI:    ~/.codex/skills/cook

Invoke it as /cook in Claude Code or $cook in Codex. It also triggers on phrases like "go cook", "iterate until it matches" or "match the references 99%".

The side-by-side script needs Node 18+ and ffmpeg on your PATH. Everything else is plain Markdown.

Quick start

Attach your reference images and send:

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

The agent asks every open question once, with defaults, then works on its own. You get stills after every round. The final video is only rendered after you approve them.

Want full control? Fill in templates/brief.md and send that instead.

The Cook Loop

1. Intake. All questions up front, each with a default. After that the agent never asks again, so it can run through the night.
2. Target. The references are the minimum bar, not inspiration. Every reference becomes a shot with notes on what to take from it. Motion references get analyzed frame by frame.
3. Bar. A number per dimension: composition, light, color, materials, detail near and far, motion. For example 99.9% for stills or at least 9/10 for a video. Models rate themselves too high, so your own estimate calibrates every score.
4. Rules. Zero external assets by default, honest sound and asset policy, a slop ban list, colors that are neither grey nor oversaturated, offline rendering on weak GPUs.
5. Time. A floor for demos (at least 8h, done after 2h means something is wrong), a hard cap for small jobs, and usage pacing so the run never hits a limit and stalls.
6. Loop. Build, render stills at the reference shots, compare side by side, score with zero-context critics, fix the weakest part, log the round in cook/ROUNDS.md, repeat.
7. Gate. Stills go to you after every round without pausing. No final render before you approve.
8. Ship. First frame is the thumbnail, best scene first, no text on screen, licensed music credited, public repo with honest numbers if you want one.

All run state lives in a cook/ folder in your project, so the agent can resume after a context reset or hand the run to another agent.

What's inside

SKILL.md                      the loop the agent follows
references/intake.md          questions, defaults, the brief
references/target-and-bar.md  references, shot list, rubric, calibration
references/loop-and-critics.md rounds, critic prompts, feedback, pacing
references/rules.md           assets, sound, ban list, color, rendering
references/ship.md            final render, video spec, repo, numbers
templates/brief.md            fill-in prompt for your own runs
templates/ROUNDS.md           the round ledger
scripts/side-by-side.mjs      reference left, render right, one row per shot

Build a comparison image yourself:

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

Made with this

Athenaion, an ancient Greek world built by Claude Opus 5.5 to match 9 AI-generated reference images. No external assets.

Post | Repo

More demos made the same way:

• Alpine lake: post, repo
• 10 Chinese landscapes: post, repo
• Golden temple: post, repo
• Afterglow pianist film: post, repo

All demos: leonlin.de/demos

Works well with

• unlazy for effort and completion discipline. Cook sets the target and the bar, unlazy makes sure the work behind it gets done.
• Any goal, loop or long-run mode your agent offers. Cook works without them.

License

MIT. Made by Leon Lin (@LexnLin).