chain-of-verification

> **AI-consumed reference.** Optimized for Claude to read during execution. > Human-readable explanation: see [docs/architecture/HIERARCHICAL_PLANNING.md](../../../docs/architecture/HIERARCHICAL_PLANNING.md) > or [docs/getting-started/](../../../docs/getting-started/) depending on topic. # Chain-of-Verification Catch hallucinations by forcing the model to verify its own claims via tools. **Governed by:** `rules/workflow/chain-of-verification.md` (when / why — MANDATORY in Phase 4) --- ## When NOT to Use - Simple file-list reports with no factual claims - Quick fixes that don't produce any claim - User uses `just do:` (but still verify destructive claims) - Already ran CoVe on this exact content (cache in memory) --- ## The 4-Step Protocol ### Step 1 — Draft Generate the answer/report as normal. Do NOT optimize for correctness yet. ### Step 2 — Plan verifications List **3–5 focused questions** that would disprove the draft. Each question must be: - **Answerable via tool** (Read/Bash/Grep/Glob) — not from memory - **Independent** — doesn't assume the draft is correct - **Specific** — targets a concrete claim, not a vibe ### Step 3 — Execute verifications For each question, run the tool and record the **actual result**. Don't reason about what "should" be — record what IS. ### Step 4 — Revise Compare each claim in the draft against verification results: - Claim verified → keep - Claim contradicted → fix - Claim unverifiable → soften ("appears to…" or remove) --...