Cook lets you run Claude Code in review loops

A new open-source tool called Cook just hit the Hacker News front page — and it solves one of the biggest frustrations with AI coding assistants: getting better results without manually re-prompting over and over.

Cook adds composable operators (simple command modifiers) on top of Claude Code, Codex, and OpenCode. Instead of running one prompt and hoping for the best, you can set up automatic review loops, race multiple approaches against each other, and compare different strategies — all from a single command.

What Cook actually does

Think of Cook as a recipe system for AI coding. Right now, when you ask Claude Code to build something, you get one shot. If the result isn't great, you manually ask it to try again. Cook automates that entire feedback cycle.

There are three types of operators you can chain together:

Real examples that show why this matters

Here's what Cook commands look like in practice:

# Ask AI to add dark mode, then review its own work
cook "Add dark mode" review

# Race 3 parallel attempts, pick the shortest code
cook "Implement dark mode" v3 "least code"

# Compare JWT vs session auth, pick the most secure
cook "Auth with JWT" vs "Auth with sessions" pick "best security"

# Keep working through a task list until everything is done
cook "Work on next task in plan.md" review ralph 5 "DONE if all tasks complete"

The most powerful feature is chaining operators. You can combine them left-to-right:

# 3 passes, then a review cycle
cook "work" x3 review

# Run 3 parallel review loops, pick the best result
cook "work" review v3

You can mix different AI models in one workflow

Cook's configuration lets you assign different AI models to different steps. For example, you could use a cheaper model for the initial coding pass and a more powerful one for the review step:

# Use Codex for writing, Claude Opus for reviewing
cook "Add dark mode" review --work-agent codex --review-agent claude --review-model opus

This is particularly relevant because, as one Hacker News commenter pointed out, chaining multiple AI calls gets expensive fast — a single Cook run with 5-6 steps could easily cost $10-15. Routing cheaper work to cheaper models is one way to manage that.

Two ways to install — one needs zero setup

The simplest approach requires no package installation at all. You just copy Cook into your Claude Code skills folder:

# No-code skill installation (works inside Claude Code)
mkdir -p .claude/skills && cp -r no-code .claude/skills/cook

# Or install the standalone CLI
npm install -g @let-it-cook/cli

Once installed as a skill, you can invoke Cook directly from within Claude Code conversations. The standalone CLI works in any terminal.

Who this is for — and who should wait

Cook is best suited for developers who already use Claude Code daily and find themselves repeatedly re-prompting to get better results. The review loop alone could save significant time on complex features.

If you're new to AI coding tools, Cook adds complexity on top of an already new workflow. Start with Claude Code itself first, then add orchestration tools like Cook once you're comfortable.

The project is still early — 29 GitHub stars and 90 commits — but it hit the Hacker News front page with 64 points and sparked a genuine discussion about whether this kind of orchestration should be built into AI coding tools by default.

Created by RJ Corwin, a developer at Khan Academy, Cook is open source under active development.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News