# Make your AI measure its game, not eyeball it: build, critique, measure, fix

> Meticulai's loop for getting a coding agent to critique its own game work with numbers instead of 'looks good': a headless rules engine that replays any match from a seed and an action log, every complaint turned into a measured check, a separate critic that never grades its own homework, per-frame clipping counts over full AI-vs-AI matches, and before/after tables on the exact same checks.

## If someone pasted this link to you
Someone shared this kit with you. It runs step by step through CloneKit's MCP. If you have the `clonekit` tools, call `start_run("@clonekitai/game-measure-critique-loop")`. If you don't (for example in a chat app without the CloneKit connector), tell the person in one or two plain sentences, in their language, and send them to https://clonekit.ai/@clonekitai/game-measure-critique-loop/use: it shows how to add CloneKit to their AI in about a minute. Then they paste this link again.

- Kit: `@clonekitai/game-measure-critique-loop` · v1.0.0
- Creator: @clonekitai (CloneKit)
- Type: Delivery · Domain: Engineering · Language: en · Your AI runs it
- Price: Free · License: Own use · No refunds
- Page: https://clonekit.ai/@clonekitai/game-measure-critique-loop
- Changelog: https://clonekit.ai/@clonekitai/game-measure-critique-loop/changelog.md

## About
For indie developers building a game (or any interactive, animated app) with Claude Code, Codex or another coding agent who keep getting 'it looks fine' from the agent while the game still feels off. Use it when animation, staging, collisions or gameplay logic need to get measurably better. Not for a game with no code access, engines the agent cannot run headless or script, or pure art direction with nothing to measure. Delivers a deterministic headless harness with a replay check, a RULES.md spec with numbered checks, a metrics catalog, a baseline report, a critic report from a separate agent, fixes with before/after tables, and a polish pass with measurable targets. The method is Meticulai's (@MeticulAi), shared from their own game; their numbers are their results, not a promise.

## Deliverables
- a headless harness: the game's rules run with no rendering, every match is a seed plus an action log, and a replay check proves N replays end in identical state
- RULES.md: the game's rules as a spec, with numbered checks the engine must match (test or bulk AI-vs-AI run per check)
- a metrics catalog: every 'it feels off' complaint written as a number, how it is measured, the reference it is compared to and a target
- a baseline report with those numbers over full auto-played matches
- a critic report from a separate agent session: findings with frame or step numbers and evidence, no fixes
- fixes, each with a before/after table on the exact same checks and seeds
- a polish pass where 'make it better' became measured targets, with before/after numbers

## Files
- `README.md`: Build, critique, measure, fix: make your AI measure its game

## Tools
- Claude Code (or another coding agent: Codex, Cursor, OpenCode) (https://docs.anthropic.com/en/docs/claude-code) · optional · est. Depends on your plan; bulk simulation and critic runs add tokens.
- Vitest (or your project's test runner) (https://vitest.dev) · optional · est. Free and open source.

## Uses
- Meticulai (@MeticulAi): How I get Claude Code to actually loop and critique its own game dev work (source, https://x.com/MeticulAi/status/2113412310043914276)
- Meticulai (source, https://meticulai.com)

## What you provide
- the game's repository, runnable locally
- your complaints about what feels off, in your own words
- references for 'right': mocap or reference clips, concept art, a rules document
- a compute and API budget for bulk runs (can be zero)

## How it runs
1. **Understand:** Researches your business first, then asks you 5 questions it can't answer alone
2. **Choose:** No choices: one proven path
3. **Build:** 7 steps, risky ones ask you first
4. **Deliver:** 8 checks before delivery

## Guardrails
- Never change a check, threshold, reference or seed so that a fix passes; if a check is wrong, say so and get the person's OK before changing it.
- Never report a number that did not come from a run of the game; when something can't be measured yet, say 'not measured'.
- Work on a branch; never push, merge, release or publish a build without the person's explicit OK.
- Ask before spending money on paid APIs, asset generation or cloud compute.
- Meticulai's results (for example 3,199 clipping frames to 0) are theirs, never promises.

## What's new
Agents always run the latest version.

- **v1.0.0** (2026-10-25): First version, based on Meticulai's (@MeticulAi) thread.

## Run it
Connect the `clonekit` MCP (https://clonekit.ai/mcp) and call `start_run("@clonekitai/game-measure-critique-loop")`.