voidlight00/autoimprove-cc
Developer toolsThis page was translated automatically. Technical details may differ from the original. View the original
Claude Code version of skill-autoimprove — auto-improve any CLAUDE.md using Karpathy autoresearch loop
Overview
Karpathy autoresearch loop for SKILL.md — Claude Code native implementation. A Claude Code native system that automatically improves SKILL.md overnight. /autoimprove Run a single line and wake up in the morning with better skills. Apply Karpathy's autoresearch loop to SKILL.md: Key difference: Karpathy uses numerical metrics for train.py, but we use pass rates from (true/false tests) as metrics. 1. — Presence of required sections, heading layer, file structure 2. — Slash commands, activation conditions, argument formats 3. — Step-by-step processes, branching conditions, output formats 4. — Prohibited items, constraints, Do NOT items 5. — Examples, code blocks, edge cases, dependencies If there is no eval.json, /autoimprove analyzes the SKILL.md and generates it automatically. Review after creation and make corrections if necessary. Andrej Karpathy's autoresearch is an autonomous research loop where AI models autonomously modify code, run benchmarks, and commit improvements as needed. Key insight: If skill quality can also be quantified, automatic optimization is possible. Binary assessments are the method for that quantification.
Overview (machine translated)
Karpathy autoresearch loop for SKILL.md — Claude Code native implementation.
Skill Evolution (PDF)
(Original PDF: assets/skill-evolution.pdf)
The Claude Code native system automatically improves SKILL.md overnight.
/autoimprove Running in one line wakes you up in the morning with a better skill.
It works with only Claude Code agents + commands without Python scripts.
How It Works
Apply Karpathy’s autoresearch loop to SKILL.md:
Key difference: Karpathy uses the numerical metric train.py,
We use the pass rate from binary assertions as a metric.
Quick Start
Global Installation (Used in All Claude Code Projects)
Project-Specific Installation
Execution
How to Write eval.json
eval.json is a collection of binary assertions that measure the quality of SKILL.md.
Structure
Field Description
| Field | Description |
|---|---|
id |
test-001 Unique ID |
description |
Test Group Description |
prompt |
Prompt scenarios for user input |
expected_behavior |
Expected Behavior for Prompts |
assertions |
Assertion arrays that can be identified as true/false |
Tips for Writing Good Assertions
Recommended Categories (5)
- Structure — Presence of required sections, heading layers, file structure
- Triggers — Slash commands, activation conditions, argument formats
- Workflow — Step-by-step process, branching conditions, output formats
- Rules — Prohibitions, Restrictions, Do NOT Items
- Completeness — Examples, code blocks, edge cases, dependencies
If there is no eval.json, /autoimprove analyzes the SKILL.md and generates it automatically.
Review and modify if necessary after creation.
Architecture
This page was translated automatically. Technical details may differ from the original.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add voidlight00/autoimprove-cc