How do I install the Autoresearch Diagrams skill?

Run npx skillmds@latest add mhassan0000/autoresearch-diagrams in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Autoresearch Diagrams safe to use?

This skill has not completed SkillMD's automated safety review yet. Independent scanners report: SkillSpector: PASS, Skill Scanner: PASS. Capability flags: executes scripts, makes network calls. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Autoresearch Diagrams?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Autoresearch Diagrams free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Autoresearch Diagrams?

MHassan0000 (@mhassan0000) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Autoresearch Diagrams by mhassan0000

Name: Autoresearch Diagrams
Rating: 4.666666666666667 (3 reviews)
Author: MHassan0000

Autoresearch Diagrams

Self-improving diagram prompt optimization using the Karpathy autoresearch pattern. Generates batches of diagrams, evaluates via Claude vision, mutates the prompt, keeps winners. Includes a live web dashboard.

MHassan0000 Updated Sep 8, 2026 1 repo stars

File contents

Autoresearch Diagrams — Self-Improving Prompt Optimization

What It Does

Applies the Karpathy autoresearch pattern to diagram generation prompts. Every 2 minutes:

Generates 10 diagrams with the current prompt (Gemini image gen)
Evaluates each against 4 criteria via Claude Sonnet vision (score out of 40)
Keeps the prompt if it beats the best score, discards otherwise
Mutates the best prompt to try to improve further
Logs everything to JSONL for tracking

Eval Criteria (4 per image, 40 max per batch)

Legible & grammatical — all text readable, correctly spelled
Pastel colors — soft pastel fills only, no saturated/dark colors
Linear layout — strictly left-to-right or top-to-bottom
No numbers — zero digits, ordinals, or step numbers

Quick Start

# Run continuous loop (every 2 min)
python3 .claude/skills/autoresearch-diagrams/autoresearch.py

# Single cycle (test)
python3 .claude/skills/autoresearch-diagrams/autoresearch.py --once

# Run N cycles
python3 .claude/skills/autoresearch-diagrams/autoresearch.py --cycles 10

# Start the live dashboard
python3 .claude/skills/autoresearch-diagrams/dashboard.py --port 8501
# Then open http://localhost:8501

Environment

Requires in .env:

NANO_BANANA_API_KEY=your_gemini_api_key
ANTHROPIC_API_KEY=your_anthropic_api_key

Dependencies: google-genai, anthropic, python-dotenv

File Structure

.claude/skills/autoresearch-diagrams/
  SKILL.md              # This file
  autoresearch.py       # Main generate -> eval -> mutate loop
  dashboard.py          # Live web dashboard (Chart.js)
  data/
    prompt.txt          # Current prompt being optimized
    best_prompt.txt     # Best prompt found so far
    state.json          # Loop state (run number, best score)
    results.jsonl       # Append-only experiment log
    diagrams/
      run_001/          # 10 diagrams per run
      run_002/
      ...

Models

Generation: gemini-2.5-flash-image (Gemini native image gen)
Evaluation: claude-sonnet-4-6 (vision + structured JSON output)
Mutation: claude-sonnet-4-6 (prompt rewriting based on failure analysis)

Dashboard

Serves at http://localhost:8501 with:

4 stat cards (current best, baseline, improvement %, runs/kept)
Score-over-time chart with keep/discard dot coloring
Per-criterion breakdown charts (legible, pastel, linear, no numbers)
Run history table
Current best prompt display
Auto-refreshes every 15s

Cost

~$0.03-0.05 per diagram generation (Gemini Flash)
~$0.01 per eval (Sonnet vision, small image + short response)
~$0.01 per mutation (Sonnet text)
Total: ~$0.40-0.60 per cycle (10 diagrams)
At 2-min intervals: ~$12-18/hour

---name: autoresearch-diagramsdescription: Self-improving diagram prompt optimization using the Karpathy autoresearch pattern. Generates batches of diagrams, evaluates via Claude vision, mutates the prompt, keeps winners. Includes a live web dashboard.---# Autoresearch Diagrams — Self-Improving Prompt Optimization## What It DoesApplies the Karpathy autoresearch pattern to diagram generation prompts. Every 2 minutes:1. Generates 10 diagrams with the current prompt (Gemini image gen)2. Evaluates each against 4 criteria via Claude Sonnet vision (score out of 40)3. Keeps the prompt if it beats the best score, discards otherwise4. Mutates the best prompt to try to improve further5. Logs everything to JSONL for tracking## Eval Criteria (4 per image, 40 max per batch)1. **Legible & grammatical** — all text readable, correctly spelled2. **Pastel colors** — soft pastel fills only, no saturated/dark colors3. **Linear layout** — strictly left-to-right or top-to-bottom4. **No numbers** — zero digits, ordinals, or step numbers## Quick Start```bash# Run continuous loop (every 2 min)python3 .claude/skills/autoresearch-diagrams/autoresearch.py# Single cycle (test)python3 .claude/skills/autoresearch-diagrams/autoresearch.py --once# Run N cyclespython3 .claude/skills/autoresearch-diagrams/autoresearch.py --cycles 10# Start the live dashboardpython3 .claude/skills/autoresearch-diagrams/dashboard.py --port 8501# Then open http://localhost:8501```## EnvironmentRequires in `.env`:```NANO_BANANA_API_KEY=your_gemini_api_keyANTHROPIC_API_KEY=your_anthropic_api_key```Dependencies: `google-genai`, `anthropic`, `python-dotenv`## File Structure```.claude/skills/autoresearch-diagrams/ SKILL.md # This file autoresearch.py # Main generate -> eval -> mutate loop dashboard.py # Live web dashboard (Chart.js) data/ prompt.txt # Current prompt being optimized best_prompt.txt # Best prompt found so far state.json # Loop state (run number, best score) results.jsonl # Append-only experiment log diagrams/ run_001/ # 10 diagrams per run run_002/ ...```## Models- **Generation**: `gemini-2.5-flash-image` (Gemini native image gen)- **Evaluation**: `claude-sonnet-4-6` (vision + structured JSON output)- **Mutation**: `claude-sonnet-4-6` (prompt rewriting based on failure analysis)## DashboardServes at `http://localhost:8501` with:- 4 stat cards (current best, baseline, improvement %, runs/kept)- Score-over-time chart with keep/discard dot coloring- Per-criterion breakdown charts (legible, pastel, linear, no numbers)- Run history table- Current best prompt display- Auto-refreshes every 15s## Cost- ~$0.03-0.05 per diagram generation (Gemini Flash)- ~$0.01 per eval (Sonnet vision, small image + short response)- ~$0.01 per mutation (Sonnet text)- **Total: ~$0.40-0.60 per cycle (10 diagrams)**- At 2-min intervals: ~$12-18/hour

Frequently asked questions