{
  "markdown": "# auto-research\n\nAutonomous skill improvement for [Claude Code](https://docs.anthropic.com/en/docs/claude-code), inspired by Karpathy's [autoresearch](https://github.com/karpathy/autoresearch).\n\nInstead of optimizing a neural network overnight, it optimizes your Claude Code skills through parallel research agents and an iterative keep/discard improvement loop.\n\n## How it works\n\n```\n/auto-research my-skill-name\n```\n\nThe command runs through 5 phases:\n\n### Phase 1: Discovery\nReads your skill file, extracts metadata, and creates a backup.\n\n### Phase 2: Parallel Research (5 agents)\nSpawns 5 research agents simultaneously, each with a focused mission:\n\n| Agent | Role | Method |\n|-------|------|--------|\n| Domain Expert | Finds best practices and conventions | Web search for guides, docs, style guides |\n| Quality Auditor | Scores the skill on 7 dimensions | Structural analysis against quality rubric |\n| Competitive Analyst | Finds how others solve similar problems | Searches cursor rules, AI prompts, cheatsheets |\n| Gap Analyst | Identifies missing scenarios | User journey and edge case analysis |\n| Tech Scout | Checks for outdated or deprecated content | Searches changelogs, breaking changes |\n\n### Phase 3: Synthesis\nCombines all findings, de-duplicates, and ranks improvement proposals using:\n\n```\nPriority Score = Impact x Confidence / Complexity\n```\n\n### Phase 4: Iterative Improvement Loop\nApplies improvements one at a time (like autoresearch experiments):\n\n1. **Apply** a single, focused change\n2. **Evaluate** it (accuracy, clarity, value-add, simplicity)\n3. **Keep** if it improves the skill, **Discard** if it doesn't\n4. **Repeat** for the next proposal\n\n### Phase 5: Results Report\nShows before/after quality scores, lists all kept/discarded experiments.\n\n## The autoresearch parallel\n\n| autoresearch (ML) | auto-research (Skills) |\n|---|---|\n| Modify `train.py` | Modify `SKILL.md` |\n| Train for 5 min on GPU | Research with 5 parallel agents |\n| Measure `val_bpb` (lower = better) | Score on 7 quality dimensions (higher = better) |\n| Keep if metric improves | Keep if quality improves |\n| `git reset` if worse | Revert edit if worse |\n| `results.tsv` tracking | Experiment log with keep/discard |\n| Simplicity criterion | Same -- complex additions with marginal value get discarded |\n\n## Quality dimensions\n\nThe auditor agent scores skills on:\n\n1. **Actionability** -- Can Claude immediately act on the instructions?\n2. **Clarity** -- Is the language unambiguous?\n3. **Completeness** -- Does it cover the full workflow?\n4. **Examples** -- Are there enough concrete examples?\n5. **Edge Cases** -- Does it handle failure modes?\n6. **Conciseness** -- Every line earns its place?\n7. **Trigger Accuracy** -- Does the description match when it should activate?\n\n## Installation\n\n### Option 1: Plugin Directory (recommended)\n\nInstall directly from the Claude Code plugin directory:\n\n```\n/plugins\n```\n\nSearch for **auto-research** and install it.\n\n### Option 2: Install plugin from GitHub\n\n```bash\n# In any Claude Code session:\n/install-plugin https://github.com/gyoz-ai/auto-research\n```\n\n### Option 3: Manual install\n\nClone and symlink the skill:\n\n```bash\ngit clone https://github.com/gyoz-ai/auto-research.git ~/auto-research\nmkdir -p ~/.claude/skills/auto-research\nln -s ~/auto-research/skills/auto-research/SKILL.md ~/.claude/skills/auto-research/SKILL.md\n```\n\nOr copy directly:\n\n```bash\nmkdir -p ~/.claude/skills/auto-research\ncurl -o ~/.claude/skills/auto-research/SKILL.md \\\n  https://raw.githubusercontent.com/gyoz-ai/auto-research/main/skills/auto-research/SKILL.md\n```\n\n## Usage\n\n```bash\n# In any Claude Code session:\n\n# By skill name (looks in ~/.claude/skills/<name>/SKILL.md)\n/auto-research my-skill\n\n# By full path\n/auto-research ~/.claude/skills/my-skill/SKILL.md\n\n# By project-level path\n/auto-research .claude/skills/my-skill/SKILL.md\n```\n\nThe command runs autonomously -- it won't ask for permission between experiments. When it's done, you can:\n\n1. **Review** the changes in detail\n2. **Run another cycle** for deeper improvements\n3. **Revert** all changes if you don't like them\n\n## Requirements\n\n- [Claude Code](https://docs.anthropic.com/en/docs/claude-code) CLI, desktop app, or IDE extension\n- An existing skill to improve (in `~/.claude/skills/` or `.claude/skills/`)\n\n## Plugin structure\n\n```\nauto-research/\n├── .claude-plugin/\n│   └── plugin.json\n├── skills/\n│   └── auto-research/\n│       └── SKILL.md\n├── LICENSE\n└── README.md\n```\n\n## Credits\n\nMethodology inspired by [autoresearch](https://github.com/karpathy/autoresearch) by Andrej Karpathy -- the concept of autonomous, iterative experimentation with a keep/discard loop and simplicity criterion.\n\n## License\n\nMIT\n",
  "bytes": 4718,
  "sha": "38cd265d5d136f487dd27fc322c20b269b0af5019a55e17fab8d0fcd4c046cd7",
  "repo_slug": "gyoz-ai/auto-research",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_gyoz_ai_auto_research_auto_research_5d15b1a2/readme"
}