← ClaudeAtlas

hyperbuild-5-stack-researchlisted

Step 5 of the hyperbuild pipeline — builds the engineering research area `research/02-engineering/` to the docs/RESEARCH-ARCHIVE.md contract. The orchestrator probes the real toolchain, runs a bounded harvest DISCOVER sweep, then DERIVES 6–8 (standard) / 10–14 (premier) engineering dimensions for the platform chosen in step 1 from a documented base set; spawns ONE hb-stack-researcher per dimension (harvest-first) → research/; extracts every load-bearing claim into temp/claims-02.json; runs the verification engine (one hb-claim-verifier per claim, each told to REFUTE it) → verify/; runs the hb-corpus-critic panel → critique/; writes _INDEX.md; and only THEN authors `research/02-engineering/author/stack-guide.md` on the claims that survived. Steps 10, 13, 14 build by that guide and the step-15 code critic enforces it. Invoked by the hyperbuild router via Skill(); not run directly by users.
zakariaf/hyperbuild · ★ 0 · Data & Documents · score 72
Install: claude install-skill zakariaf/hyperbuild
# Step 5 — Engineering research (area `02-engineering`: research → verify → critique → author) You are executing step 5 (stack-research) of the hyperbuild pipeline. Steps 4 and 4.5 fixed WHAT to build (the PRD + feature specs); this step fixes HOW to build it — step 10 forges the project skills from the stack guide, step 13 scaffolds by it, step 14 implements by it, and step 15's `hb-code-critic` treats deviations from it as findings. **Goal:** a complete research AREA at `research/02-engineering/` — all four phases — ending in `research/02-engineering/author/stack-guide.md`, a committed "we will do X" guide whose every decision rests on a claim that survived adversarial verification. **Why four phases and not four topic docs.** The old shape of this step spawned four researchers and shipped their merge. A surveying researcher optimizes for coverage: it will confidently repeat a 2023 blog post, a package archived last year, and an API name that never existed — and step 14 will then write six files against that API name. In the exemplar corpus the verification pass caught a discontinued golden-test toolkit still being recommended, a state library whose entire test idiom had changed, a lint package that cannot be used on the installed SDK, and a widely-cited test-pyramid ratio that appears to be fabricated. **A survey is a hypothesis; only `verify/` turns it into a finding.** **BINDING FORMAT: `docs/RESEARCH-ARCHIVE.md`.** Every file this step produces obeys it — the area l