verify-before-claimlisted

# Verify Before Claim Closes the single most expensive friction pattern in agent-assisted work: **treating a memory as a fact without a live check**. Implemented as one skill plus an always-loaded floor rule in CLAUDE.md (a passive CLAUDE.md rule beats a pile of active skills — the skill is the reference gate, the rule is the floor). ## Iron Law **Do NOT assert a fact whose primary source you did not open IN THIS SAME response.** A memory from 5 prompts ago, a summary doc, or "I think" is NOT a source. Re-read is mandatory. ## Gate Function (5 steps before any claim) 1. **What type of claim is it?** — existence ("doesn't exist / there is no such thing") · status ("blocked / not deployed / already done") · measured value (number/metric/%) · API behavior (enum/payload/response code) · vendored capability (whether a vendored UI library can do X). 2. **What is the primary source** for this type? (table below) 3. **Did I open it in THIS response?** — is there a tool call (Read/Grep/curl/WebFetch) above in this same turn? If NOT → STOP, open it, then phrase the claim. 4. **Was the search broad enough?** — not a single directory: broad `find`/`grep` + git history + variant names + the full primary source (not just a summary). "I looked in one place and didn't find it" ≠ "it doesn't exist". 5. **Phrase with provenance** — `[from: Read file:line]` / `[from: curl <endpoint>]` / `[from: inference]`. If it cannot be verified — mark it "requires verification", do NOT present it as fa