Drop a trigger phrase into Claude Code. AutoResearch sets up your project, generates program.md, runs the baseline — and starts iterating while you sleep.
One-call AutoResearch onboarding for any ML project. Answer 5 questions — Claude handles everything: git branch, program.md, baseline run, and a copy-paste agent prompt you fire before bed.
autoresearch/[date] and initializes results.tsvprogram.md from framework-specific templatesSkills in /Users/meetdeshani/Desktop/skills/ are auto-loaded by Claude Code on every session — no extra config needed. The skill is already available if you're on Meet's machine.
# In any Claude Code session, just type: autoresearch setup # Or with a hint: set up autoresearch for my LoRA fine-tuning project # Skill file location (already exists): /Users/meetdeshani/Desktop/skills/autoresearch-setup/SKILL.md
Copy or symlink the skill folder into any project so it's available when Claude Code runs in that directory.
# Option A — hard copy (static snapshot) cp -r /Users/meetdeshani/Desktop/skills/autoresearch-setup/ \ /path/to/your/project/skills/autoresearch-setup/ # Option B — symlink (stays in sync with source) mkdir -p /path/to/your/project/skills ln -s /Users/meetdeshani/Desktop/skills/autoresearch-setup/ \ /path/to/your/project/skills/autoresearch-setup # Then open Claude Code in the project dir — skill is live
Deploy to a VPS running OpenClaw or any remote Claude agent.
# Upload to OpenClaw skills directory on VPS scp -r /Users/meetdeshani/Desktop/skills/autoresearch-setup/ \ root@your-vps:/opt/services/openclaw/data/agents/main/agent/skills/ # Restart OpenClaw to pick up new skill ssh root@your-vps "docker restart openclaw" # For any Claude API agent — paste SKILL.md into system prompt cat /Users/meetdeshani/Desktop/skills/autoresearch-setup/SKILL.md
Applies Karpathy's AutoResearch loop to SKILL.md files. Instead of training a neural network, it improves prompt quality — one surgical edit at a time, keeping gains and reverting regressions, using SQS (Skill Quality Score) as the metric.
.autoresearch/results.tsv with hypothesis, axis, expected gain, and actual deltaoptimize higgsfield-prompt) or all at once (optimize my skills)Already installed on Meet's machine. Open any Claude Code session and use a trigger phrase.
# In any Claude Code session: optimize my skills # For a specific skill only: autoresearch-skill-optimizer meet-documents # Report only — no changes: autoresearch-skill-optimizer report # Skill file location: /Users/meetdeshani/Desktop/skills/autoresearch-skill-optimizer/SKILL.md
Copy or symlink to use in any project with its own skills directory.
# Hard copy cp -r /Users/meetdeshani/Desktop/skills/autoresearch-skill-optimizer/ \ /path/to/your/project/skills/autoresearch-skill-optimizer/ # Or symlink (stays in sync) mkdir -p /path/to/your/project/skills ln -s /Users/meetdeshani/Desktop/skills/autoresearch-skill-optimizer/ \ /path/to/your/project/skills/autoresearch-skill-optimizer
Deploy to VPS running OpenClaw or paste SKILL.md into any Claude API system prompt.
# Upload skill to OpenClaw on VPS scp -r /Users/meetdeshani/Desktop/skills/autoresearch-skill-optimizer/ \ root@your-vps:/opt/services/openclaw/data/agents/main/agent/skills/ # Restart to pick up ssh root@your-vps "docker restart openclaw" # For Claude API — read SKILL.md and paste into system prompt cat /Users/meetdeshani/Desktop/skills/autoresearch-skill-optimizer/SKILL.md
They work independently but compose well — set up AutoResearch on a project, then optimize the skills that power it.
| Question | autoresearch-setup | autoresearch-skill-optimizer |
|---|---|---|
| I want to run AutoResearch on my ML project | ✓ This one | — |
| I want to improve my Claude Code SKILL.md files | — | ✓ This one |
| Works on HuggingFace Trainer, PyTorch, Axolotl | ✓ All frameworks | — |
| Works on Apple Silicon (no NVIDIA GPU) | ✓ Mode D → MLX | ✓ No GPU needed |
| Creates git branch + baseline run | ✓ | — |
| Runs hypothesis → edit → score → keep/revert loop | ✓ (on model code) | ✓ (on SKILL.md) |
| Produces a copy-paste overnight agent prompt | ✓ | — |
| Logs results to TSV file | ✓ results.tsv | ✓ .autoresearch/results.tsv |
Works with any Claude Code session. No config files, no environment variables, no YAML.
Skills are in /Desktop/skills/ — auto-loaded by Claude Code. No extra setup if you're on Meet's machine. Otherwise copy the SKILL.md folder into your project or Claude Code skills path.
# Open Claude Code and type: what skills do you have?
Navigate to your ML project in Claude Code, then trigger the setup skill. Claude will ask 5 questions and do the rest.
set up autoresearch on my project # Claude will ask: 1. What is your training script? 2. What hardware? (H100 / A100 / 4090 / M2) 3. What is your validation metric? 4. What should the agent explore? 5. Which AI agent? (Claude Code / API / Codex)
After setup, Claude prints an exact agent launch prompt. Copy it. Paste it into Claude Code (or your chosen agent) and hit Enter.
═══════════════════════════════════ COPY THIS INTO CLAUDE CODE ═══════════════════════════════════ Read program.md and train.py. Run: uv run python train.py Parse val_loss from output. If improvement → git commit "exp-N" Else → git checkout train.py Repeat 100x. Log to results.tsv. ═══════════════════════════════════
The agent runs 100 experiments overnight. In the morning, check what improved.
# See what was kept grep "KEPT" results.tsv # Review all changes git diff main..autoresearch/$(date +%b%d | tr A-Z a-z) # Update program.md with learnings update program.md based on results.tsv
Both skills are ready. The loop starts tonight.