---
name: autoresearch
description: Auto-improve any OpenClaw skill using iterative prompt optimization. Inspired by Karpathy's autoresearch method. Run when asked to improve, optimize, or auto-tune any skill. Runs a loop — score → change → test → keep/revert — until target quality is reached. Fully autonomous once started.
---

# Autoresearch Skill

Auto-improve any skill file in this workspace using iterative, scored prompt optimization.

## The Method (Karpathy's Autoresearch)

1. Define what "good" looks like as a binary checklist (yes/no questions)
2. Run the skill on test inputs
3. Score each output against the checklist
4. Propose ONE targeted change to the skill prompt
5. Test again — keep if score improved, revert if not
6. Repeat until 95%+ or max rounds hit

## Quick Start

```bash
cd /Users/jacopo/.openclaw/workspace-bertuccio/skills/autoresearch

python3 scripts/autoresearch.py \
  --skill /path/to/SKILL.md \
  --checklist inputs/my-checklist.json \
  --inputs inputs/my-inputs.json \
  --max-rounds 30 \
  --target 0.95 \
  --dashboard \
  --out /path/to/improved-SKILL.md
```

## Input Formats

### checklist.json
```json
[
  {"question": "Does the output include a specific number or result? (not vague promises)"},
  {"question": "Is the output free of buzzwords like 'revolutionary', 'cutting-edge', 'synergy'?"},
  {"question": "Does the output include a clear, specific call to action?"},
  {"question": "Is the output under 200 words?"},
  {"question": "Does the first line address a specific pain point?"}
]
```

3-6 questions is the sweet spot. Too many and the skill overfits.

### inputs.json
```json
[
  "Write landing page copy for an AI productivity tool for freelancers",
  "Write landing page copy for a B2B SaaS that automates invoicing",
  "Write landing page copy for a fitness app targeting busy parents"
]
```

## Dashboard

Pass `--dashboard` to get a live HTML dashboard at:
`skills/autoresearch/autoresearch-dashboard/index.html`

Auto-refreshes every 8 seconds. Shows:
- Current score + history chart
- Per-check pass/fail breakdown
- Change log with keep/revert status

## Outputs

| File | Description |
|------|-------------|
| `improved-SKILL.md` | Best version found (saves as you go) |
| `backup-SKILL.md` | Original skill, untouched |
| `autoresearch-log.json` | Full history: every round, score, change, keep/revert |
| `autoresearch-dashboard/` | Live HTML dashboard |

## Recommended Checklist Questions

Good checklist questions are:
- **Binary** — pass or fail, not "sort of"
- **Specific** — "mentions a number" not "is concrete"
- **Independent** — each checks one thing

Bad questions:
- "Is the output high quality?" (too vague)
- "Is the tone right?" (too subjective)
- "Is it good?" (useless)

## Requirements

- `ANTHROPIC_API_KEY` set in environment
- Python 3.9+
- `pip3 install anthropic`

## Notes

- The improved skill is saved separately — original is always backed up
- The agent uses Claude itself to both evaluate outputs and propose changes
- Changes that hurt the score are automatically reverted
- The changelog in `autoresearch-log.json` is the most durable artifact — hand it to a future model to continue where this one left off