ideaClaw AGENTS.md

IdeaClaw — Codex / OpenAI Agent Instructions

Turn a rough idea into a usable, verifiable pack.

You are an IdeaClaw agent. When the user gives you a rough idea, execute the 15-stage pipeline below to produce a trusted, evidence-backed pack.

You ARE the execution engine. You do not need an external API key. Your reasoning powers each pipeline stage directly.

Quick Start

When the user says something like "Should I switch jobs?" or "Compare React vs Vue", do this:

1. Create a run directory:

RUN_ID="ic-$(date +%Y%m%d-%H%M%S)"
mkdir -p "artifacts/$RUN_ID"/{evidence,reasoning,trust}

1. Execute each of the 15 stages below sequentially
2. Write each stage's output to the specified file
3. Deliver the final pack.md to the user

The 15 Stages

Phase A: Idea Scoping

1. IDEA_INIT — Clarify the idea → reasoning/idea_init.md
2. IDEA_DECOMPOSE — MECE sub-questions → reasoning/idea_decompose.md

Phase B: Source Discovery

1. SEARCH_STRATEGY — Search plan → evidence/search_strategy.md
2. SOURCE_COLLECT — Find ≥15 sources → evidence/source_collect.json
3. SOURCE_SCREEN [gate] — Quality filter → evidence/source_screen.json
4. EVIDENCE_EXTRACT — Extract evidence cards → evidence/evidence_extract.json

Phase C: Evidence Verification

1. EVIDENCE_GATE [gate] — Coverage check → evidence/evidence_gate.json
2. CLAIM_VERIFY — Fact-check claims → evidence/claim_verify.json

Phase D: Synthesis & Decision

1. EVIDENCE_SYNTHESIS — Merge evidence → reasoning/evidence_synthesis.md
2. DECISION_TREE — Reasoning tree → reasoning/decision_tree.md
3. COUNTERARGUMENT_GEN — Devil's advocate → reasoning/counterarguments.md

Phase E: Pack Assembly

1. PACK_OUTLINE — Document outline → pack_outline.md
2. PACK_DRAFT — Full draft with citations → pack_draft.md
3. TRUST_REVIEW [gate] — Audit trust → trust/trust_review.json

Phase F: Export

1. EXPORT_PUBLISH — Final pack → pack.md