hyperbuild-9-skill-researchlisted

Step 9 of the hyperbuild pipeline — research area 04-claude-skills: deep research on Claude Code skill authoring. Spawns 4 hb-stack-researcher subagents in ONE parallel wave, one per dimension (skill format and frontmatter; progressive disclosure and file splitting; check-script gates; harvested-collection survey), each writing research/04-claude-skills/research/<dimension>.md in the docs/RESEARCH-ARCHIVE.md format (claims as H3 assertions + Sources + provenance block). Claude Code mechanics are EXACTLY where invented frontmatter fields, invented size limits and invented triggering behaviour appear, so the orchestrator registers every claim in runs/<run_tag>/temp/claims-04.json, drafts research/04-claude-skills/author/skill-authoring-guide.md, and RUNS THE VERIFICATION ENGINE (step 3.5, phases V1-V6: one hb-claim-verifier per claim hunting invented names first, a 2/3-seat hb-corpus-critic panel, the author patch, and _INDEX.md). The guide ends in a binding "Rules for our generated skills" section that step 10
zakariaf/hyperbuild · ★ 0 · Web & Frontend · score 72

Install: claude install-skill zakariaf/hyperbuild

# Step 9 — Skill research (area `04-claude-skills`) You are executing step 9 (skill-research) of the hyperbuild pipeline. Step 8 (mockups) runs CONCURRENTLY with you as the 8 ∥ 9 pair — the router drives both steps' spawn waves in the same block; you share no inputs with it and never wait on it. Your successor is step 10 (skill-forge), which will consume this area's `author/` guide as authoring law when it generates the five project skills that steer every Stage-B implementer. **Goal:** fill research area `research/04-claude-skills/` with all four phases — `research/` (4 dimensions), `verify/` (one adversarial fact-checker per load-bearing claim), `critique/`, `author/skill-authoring-guide.md`, plus `_INDEX.md` — where the guide ends in a numbered **"Rules for our generated skills"** section. A weak guide here becomes five weak generated skills becomes sloppy step 14 implementation. This step is small but load-bearing. **BINDING: `docs/RESEARCH-ARCHIVE.md`.** Area `04-claude-skills` obeys it exactly: the four-phase layout (§2), the file formats (§3), the universal provenance rule (§4), the claim→verify mechanism (§5), the canonical verifier prompt (§6) and the synthesis rule (§7). Violations are DEFECTS: the file is rejected and the agent re-spawned. Cite the file by path in every spawn prompt. **WHY VERIFICATION MATTERS MORE HERE THAN ANYWHERE ELSE.** Claude Code mechanics are the single most fertile ground for invented-feature claims in this whole pipeline: a frontmatte