← ClaudeAtlas

hyperbuild-3-5-research-auditlisted

Step 3.5 of the hyperbuild pipeline — the VERIFY + CRITIQUE + INDEX phases of research area 01-product-and-market, and the home of the REUSABLE VERIFICATION ENGINE that steps 5, 6 and 9 re-run over their own areas. Runs after steps 2 AND 3 (the 2 ∥ 3 pair) both complete. Reads the area claim registry runs/<run_tag>/temp/claims-01.json, spawns ONE hb-claim-verifier PER CLAIM in parallel batches — each told to REFUTE its single claim against primary sources and to write research/01-product-and-market/verify/<dimension>--<claim-slug>.md with a closed-vocabulary verdict (CONFIRMED | PARTIALLY_TRUE | REFUTED | UNVERIFIABLE) — then runs the CRITIQUE PANEL (2 standard / 3 premier, distinct lenses) over the whole area corpus. The orchestrator then PATCHES the area's author/ syntheses per verdict: REFUTED claims leave the argument and are recorded under "## Refuted by verification", PARTIALLY_TRUE claims carry their correction inline, nothing is ever silently deleted — and closes the area with _INDEX.md. Invoked by th
zakariaf/hyperbuild · ★ 0 · Web & Frontend · score 72
Install: claude install-skill zakariaf/hyperbuild
# Step 3.5 — Area 01 verification (verify → critique → patch → index) You are executing step 3.5 (research-audit) of the hyperbuild pipeline. Steps 2 and 3 ran concurrently as the 2 ∥ 3 pair and filled `research/01-product-and-market/research/` with competitor dossiers and per-platform sentiment files, then drafted the two `author/` syntheses. Those findings are UNVERIFIED BY CONSTRUCTION. Step 4, your successor, builds the PRD directly on the syntheses and must never cite a claim you refuted. **BINDING FORMAT CONTRACT: `docs/RESEARCH-ARCHIVE.md`.** Read it before you spawn anything, and cite it BY PATH in every spawn prompt. Its violations are DEFECTS, not style disagreements — a file that breaks it is rejected and the agent re-spawned. **THIS STEP IS ALSO THE ENGINE.** Steps 5, 6 and 9 do NOT re-derive this procedure — each cites "the verification engine described in step 3.5, applied to area `<NN>`", binds the parameter block below, and runs phases V1–V6 unchanged. Anything you change in the engine section changes all four research areas. **Goal:** every load-bearing claim in area 01 has a `verify/` file whose verdict was reached by an agent trying to REFUTE it; a critique panel has read the whole area corpus; both `author/` syntheses have been patched so they rest only on claims that survived; and `research/01-product-and-market/_INDEX.md` lists every agent, grouped by phase, with file sizes. **Why this step exists:** a top-5 "pain point" that is really one viral thr