Your AI-powered coding chef crafting clean, secure, and maintainable code with precision
| Without chef | Chef 1.2 | Chef 1.3 | Chef 1.4 | |
|---|---|---|---|---|
| The reviewer accepts the first answer | 59% | 87% | 90% | 92% |
| Tasks sent back for fixes | 41% | 13% | 10% | 8% |
| Review rounds, as a share of tasks | 46% | 15% | 13% | 10% (β78%) |
| Lines the reviewer reads, median (lockfiles excluded) | 98 | 231 | 233 | 226 (+131%) |
| Agent time per task, first answer | 21 s | 42 s | 44 s | 46 s (+117%) |
| Agent time per task, fixes and new reviews | 29 s | 11 s | 7 s | 7 s (β76%) |
| Cost per task, review fixes included | $0.120 | $0.194 | $0.181 | $0.185 (+54%) |
13 tasks in 11 stacks, 3 runs each, Sonnet 5.5; percentages compare chef 1.4 with plain Claude. Chef 1.2, 1.3 and 1.4 differ by one or two reviews out of 39: read them as equal. The reviewer is a blind Opus session that never knows chef exists; when it would not merge, its required changes go back to the same setup, up to 2 rounds. The code works either way: every first answer passes its hidden tests. Runs cut short by an API or usage limit are refused, not counted; each result names the harness commit and reviewer that produced it, and every scorer is tested in CI on a known-good and a known-bad answer. How it was measured.
Chef 1.4's first answer costs 134% more than a plain one, because it already does what the reviewer would ask for: tests, strict types and linters, pinned versions, configs checked by their tools. Counting the fixes plain Claude needs afterwards, the gap falls to +54%, for 78% fewer review rounds. In agent time the two end level (73 s and 78 s per task until merged, +7%); the difference is on the reviewer's side: 81% fewer tasks come back for another review.
Across versions, on the Python, Docker and Prometheus tasks every version was run on:
Per token, chef is less cost-effective: $1 buys 8.2 accepted first answers without it and 5.4 with chef 1.4. What it saves is the reviewer's rounds: 1.10 reviews per task instead of 1.46. Once a review round costs more than $0.18 of the reviewer's time (about 11 seconds at $60/h), chef is the cheaper option; at $10 per round, $11.19 per task against $14.72 (β24%).
Reviews per task are the first review plus one after each fix. The model prices a round, not the lines in it: chef 1.4's code takes 2.3 times as many lines to read (median 226 against 98, lockfiles excluded).
Nine more tasks in the languages and tools developers use most, run on chef 1.4, 3 runs each, with the current reviewer:
| First answer accepted, of 3 | Node | React | Bash | SQL | PHP | PowerShell | C | C++ | Kotlin |
|---|---|---|---|---|---|---|---|---|---|
| Without chef | 0 | 3 | 3 | 2 | 1 | 0 | 0 | 0 | 3 |
| Chef 1.4 | 3 | 1 | 3 | 2 | 3 | 0 | 3 | 1 | 2 |
Chef is ahead on 4 stacks, behind on 2 and level on 3; with 3 runs per stack, read the total, not a single stack. Every chef first answer passes its hidden test; plain Claude fails 2 of 27 (PHP, a composer.json that is not valid JSON). How it was measured.
/plugin marketplace add Neotoxic-off/chef
/plugin install chef@chef
That's it: chef applies to every coding prompt from now on.
- Asks before guessing: one question when that settles it, the council when it does not. Remembers your answers in
.chef.md. - Your word wins: an explicit instruction in your prompt overrides chef's conventions and
.chef.md, never security or the project folder. - Tests every change, and sets up a formatter, a linter and a strict type checker. Warnings get fixed, never hidden.
- Types every signature, pins dependencies, images and actions.
- Knows your stack: 127 references (34 languages, frameworks, infra, CI) with the right tools and configs. The 11 stacks the benchmark covers (Python, TypeScript, Go, Rust, Java, C#, Docker, Helm, Terraform, GitHub Actions, Prometheus) are checked by it; the others say which versions they were written for, and that they are not benchmarked.
- Brings a council of agents for design decisions, and a reviewer for changes. The benchmark measures neither: its tasks are small and run without them.
- Talks less: short replies, agent reports capped at 20 lines, and the smallest complete change (
lean.md, adapted from ponytail, MIT).
| Command | |
|---|---|
/chef:level lite | full | max |
How much process: lite keeps every code rule but skips the council and reviewer. |
/chef:status |
Current level, project settings and detected stacks. |
/chef:transform <goal> |
Upgrade or migrate a codebase in small, tested, committed batches. |
/chef:level litedrops the council and the reviewer agent, the biggest cost on large tasks.- Turn chef off in one project, for you only:
claude plugin disable chef@chef --scope local. - Or skip the plugin: copy
export/AGENTS.md, chef's 12 core rules, into your project. It works with any agent (Codex, Cursor, Copilot) and gets part of the effect (in the 1.0 benchmark: 56% merged as is against 89% with chef, at $0.084 per task).
Edit skills/chef/SKILL.md or a stack file in skills/chef/references/, then run bash scripts/check.sh. Changes are listed in the CHANGELOG.
