domains / research / auto-research / auto-research

auto-research

Skill card

combined

The three copies use the same name but describe three different procedures. dzhng/skills defines a local optimization loop that maps tunable parameters against a benchmark. huytieu/COG-second-brain is a thin pointer that delegates to a different playbook and subagents. sickn33/agentic-awesome-skills defines an external research workflow that requires explicit user consent before web search or ChatGPT consultation. None of the copies is a reproduction of the others, so crowning one would hide real alternatives.

Repositories

Other copies

Combined skill

# Auto Research

This skill covers three distinct workflows. Select the one that matches the user's request.

## Workflow A: Local Parameter Optimization (dzhng/skills)

1. Recover the research state: objective, incumbent artifact, regression set, hypothesis queue, and remaining budget.
2. Clarify success criteria: evaluator, metric, baseline, and required improvement. Define the smallest useful loop with one cheap task.
3. Establish the active task's baseline. Run the unchanged artifact once or reuse a compatible recorded result.
4. Plan a small hypothesis batch. List distinct explanations of the bottleneck and the parameters that could test them.
5. Test one hypothesis on the active task. Change one conceptual factor, record the prediction, and compare observation with prediction.
6. Combine and promote compatible winners. Check interactions and remove one change at a time when attribution is unclear.

## Workflow B: Delegated Strategic Research (huytieu/COG-second-brain)

1. Read .claude/skills/auto-research/SKILL.md and execute it exactly as written.
2. Follow .agents/rules/cog.md.
3. Where the playbook delegates to a .claude/agents/<name> worker, invoke .agents/agents/<name>.md via invoke_subagent instead.

## Workflow C: External Research with Explicit Consent (sickn33/agentic-awesome-skills)

1. Propose the research boundary: state the source, exact query or redacted prompt, whether any local text would leave the machine, and the likely cost. Wait for user approval.
2. Research: after approval, use web search or a browser session the user has explicitly selected and authorized. Do not install packages automatically or access browser cookies, other tabs, saved passwords, or sessions.
3. Present: distill findings into concise options with sources.
4. Await Approval: do not write code until the user says "go ahead" or picks an option.
5. Implement: once approved, execute with confidence.

### Conflicts

- dzhng/skills::skills/engineering/auto-research/SKILL.md runs local experiments without requiring external web search or third-party consultation.
- huytieu/COG-second-brain::.agents/skills/auto-research/SKILL.md does not define the procedure itself; it delegates to .claude/skills/auto-research/SKILL.md.
- sickn33/agentic-awesome-skills::plugins/agentic-awesome-skills-claude/skills/auto-research/SKILL.md requires explicit user approval before any web search or ChatGPT submission and forbids sending context, files, or browser state without consent.