メニュー
Structured debugging with competing hypotheses and escalation levels. Use when encountering unexpected behavior, multiple possible causes, or after 3+ failed fix attempts. Separates facts from theories and tracks evidence systematically.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Multi-hypothesis research campaign protocol. Use for open-ended optimization goals ("make X faster", "improve the metric") where several candidate ideas compete for a limited budget. Triages ideas cheaply, confirms survivors with full preregistered experiments, and keeps a committed ledger so dead ideas stay dead.
Controlled benchmark experiment protocol for solver, performance, routing, and optimization work. Use when testing a hypothesis against metrics, comparing implementations, or deciding whether to promote or abandon a benchmark-driven change.
Coordination-only protocol. You dispatch subagents, you do not execute code. Restricts tools to Read, Write (prompts/ and notes- only), Agent, Glob, Grep. Use when orchestrating multi-agent work.
Four-role sequential pipeline for complex, high-risk changes. Scout surveys the codebase, Coder implements with TDD, Analyst finds problems, Verifier gates the merge. Use for large refactors, architectural shifts, and interdependent multi-phase work.
Background knowledge for writing and launching subagent prompts. Covers prompt file conventions, worker identity declarations, batch dispatch patterns, and parallel swarm safety. Auto-invocable when dispatching agents.
Deletion-first fixed-point protocol for refactors, cleanup, migrations, and architectural convergence. Use when replacing old surfaces, removing helper debt, shrinking code, moving ownership boundaries, or auditing every symbol in a slice.
| name | investigation |
| description | Structured debugging with competing hypotheses and escalation levels. Use when encountering unexpected behavior, multiple possible causes, or after 3+ failed fix attempts. Separates facts from theories and tracks evidence systematically. |
| disable-model-invocation | false |
Use when: unexpected behavior, multiple possible causes, debugging > 3 failed attempts.
Start at L1. Escalate when stuck.
investigations/[topic].md in project rootSeparate FACTS (verified, have evidence) from THEORIES (plausible, untested).
Maintain several competing hypotheses. Chasing a single theory is confirmation bias with extra steps.
For each test:
Before each action: state hypothesis. After each action: record result.
# Investigation: [topic]
## Facts (verified)
- [thing] - evidence: [how verified]
## Theories (plausible)
1. [theory A] - would explain [X], predicts [Y]
2. [theory B] - would explain [X], predicts [Z]
3. [theory C] - would explain [W], predicts [Y]
## Tests Run
| Test | Hypothesis | Result | Rules Out | Supports |
|------|------------|--------|-----------|----------|
| | | | | |
## Current Best Theory
[Which theory survives the evidence, and why]
## Open Questions
-
## Next Action
[What to try next and what it would tell us]