# autoimprove-cc — Auto-Improve SKILL.md Loop

> Run a Claude Code-native autoresearch loop to improve SKILL.md using binary assertions and git commit/reset; verified 63★ and exposes `/autoimprove`.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

```bash
git clone https://github.com/VoidLight00/autoimprove-cc.git
ln -s $(pwd)/autoimprove-cc/.claude/commands/autoimprove.md ~/.claude/commands/autoimprove.md
/autoimprove skills/my-skill --dry-run
```

## Intro

Run a Claude Code-native autoresearch loop to improve SKILL.md using binary assertions and git commit/reset; verified 63★ and exposes `/autoimprove`.

**Best for:** Iterating on skill quality with test-like assertions

**Works with:** Claude Code CLI · git repos · eval.json assertions

**Setup time:** 10–25 minutes

### Key facts (verified)

- GitHub: 63 stars · 5 forks · pushed 2026-04-10.
- License: MIT · owner avatar + repo URL verified via GitHub API.
- README-backed entrypoint: `/autoimprove skills/my-skill`.

## Main

- Use it to evolve SKILL.md overnight: the README describes a loop that scores assertions, edits, and commits only when score improves.

- Treat `eval.json` as your benchmark: make assertions binary (true/false) so the loop can optimize reliably.

- Use `--dry-run` to score without mutating git history, then enable commits once behavior is trusted.

- Keep the target skill inside a git repo; README notes rollback relies on git reset/commit behavior.

### Source-backed notes

- README frames the loop as a Claude Code-native adaptation of Karpathy autoresearch, using binary assertions for scoring.
- README provides quick start install steps that link Claude Code agents/commands and a `/autoimprove` command interface.

### FAQ

- **Will it rewrite my history?**: It can commit/reset; start with `--dry-run` and run in a branch if you’re unsure.
- **Do I need Python scripts?**: README says it runs with Claude Code agents + commands, no external Python runtime required.
- **What’s the metric?**: Binary assertion pass rate in eval.json; keep assertions precise and checkable.

## Source & Thanks

> Source: https://github.com/VoidLight00/autoimprove-cc
> License: MIT
> GitHub stars: 63 · forks: 5

---

<!-- ZH -->

## Quick Use

```bash
git clone https://github.com/VoidLight00/autoimprove-cc.git
ln -s $(pwd)/autoimprove-cc/.claude/commands/autoimprove.md ~/.claude/commands/autoimprove.md
/autoimprove skills/my-skill --dry-run
```

## Intro

用 Claude Code 原生 autoresearch 循环自动改进 SKILL.md：二值断言打分、git commit/reset；已验证 63★，通过 `/autoimprove` 一键运行与 dry-run。

**Best for:** 用“可验证断言”迭代提升 skills 质量的团队

**Works with:** Claude Code CLI · git 仓库 · eval.json 断言

**Setup time:** 10–25 minutes

### Key facts (verified)

- GitHub：63 stars · 5 forks；最近更新 2026-04-10。
- 许可证：MIT；作者头像与仓库链接均已通过 GitHub API 复核。
- README 中可对照的入口命令：`/autoimprove skills/my-skill`。

## Main

- 用于“过夜优化”：README 描述循环会打分断言、修改 SKILL.md，只有分数提升才 commit，否则 reset。

- 把 `eval.json` 当基准：用二值断言（真/假）降低主观性，才能稳定优化。

- 先用 `--dry-run` 只评分不改 git，确认行为可靠后再开启自动提交。

- 目标技能需在 git 仓库中；README 说明回滚依赖 git reset/commit 机制。

### Source-backed notes

- README 把它描述为 Karpathy autoresearch 的 Claude Code 改编版，并用二值断言作为评分机制。
- README 给出通过链接 agents/commands 的安装方式，并用 `/autoimprove` 作为统一入口。

### FAQ

- **会改写提交历史吗？**：可能会 commit/reset；不确定就先 `--dry-run`，或在分支里运行。
- **需要 Python 脚本吗？**：README 表示只用 Claude Code agents + commands 即可运行，无需外部 Python。
- **优化指标是什么？**：eval.json 的二值断言通过率；断言要精确、可检查。

## Source & Thanks

> Source: https://github.com/VoidLight00/autoimprove-cc
> License: MIT
> GitHub stars: 63 · forks: 5


---
Source: https://tokrepo.com/en/workflows/autoimprove-cc-auto-improve-skill-md-loop
Author: Skill Factory