auto-research

name	auto-research
description	Autoresearch loop using plan → propose → test → score → keep/discard → repeat. Use this skill for deep research style tasks where a systematic process is needed over large amounts of research. Do not use this for research only needing 1-3 queries.
version	1.0.0

Autoresearch (Karpathy FS)

Purpose

Run deep research using a strict iterative loop with persistent markdown state.

This is an optimization process, not note-taking.

---

When to Use

Use this skill ONLY if:

• the task requires multiple research iterations
• correctness matters more than speed
• the answer cannot be trusted from a single pass

Do NOT use for:

• simple questions
• quick lookups
• brainstorming

---

Core Loop

Repeat:

1. pick weakest topic
2. propose ONE improvement
3. apply change
4. score before and after
5. keep or discard
6. log iteration

---

Execution Rules

Each iteration MUST include:

• iteration number
• topic
• one change only
• score before
• score after
• keep/discard decision

If any are missing, the iteration is invalid.

---

Workspace

Create: research// program.md status.md final.md experiments.md topics/ research// program.md status.md final.md experiments.md topics/