Autodev — Testing skill for Claude Code

English · 简体中文

autodev

**Stop letting your agent grade its own homework.**

An agent can add a feature and call it done, or change some code and call it faster. autodev makes both claims checkable: agree on a benchmark first, stay inside an agreed list of files, and accept a change only when a script says it is better. Everything else gets rolled back.


Phase	What happens	What you do
Contract	Write down the goal, the criterion, the budget, and which files may change	Confirm once
Baseline	Write the tests or build the benchmark, run it once, and hash the files that must not change	Nothing
Loop	Edit the allowed files → run the benchmark → the script decides → commit or roll back → log the attempt	Wait for the budget you approved
Review	Re-run all tests on the final code, then refactor, then report what the loop gained	Nothing

   

npx skills add Momoyeyu/autodev -g

Works with Claude Code, Cursor, Codex CLI, OpenCode, and anything else that reads `SKILL.md`.

The problem

Ask an agent to make something faster and it will happily make a change, declare victory, and move on. Three things go wrong, and none of them are about laziness:

1. It grades its own homework. With no measurement taken before the edit, "faster" is an opinion. A test written after the code passes immediately and proves nothing. A benchmark number with no baseline proves just as little.
2. **It optimizes the measurement instead of