Hook Generator

- **Fixed** the structure grader rejecting capitalized command openers such as "Try this"; case-insensitive matching is scoped to the opening verb so the uppercase emphasis detector stays case-sensitive.
- **Fixed** "For writers" identity hooks and "Bring" command hooks being unclassified. Near misses remain unclassified.
- Kept regression tests for all three failures outside the published bundle. Registry, recipes, media and all other detectors unchanged (DAS2-34).

Marketplace card media only (propagation of 1.5.0, DAS-213). No behaviour change.

- **Changed** `media.hero_image_url` to the SAY / WRITE / SHOW chalk diagram, so the card shows the 3-second
  video hook that 1.5.0 added. Same image as the blog cover and the `/skills-hook-generator` hero (one render,
  three surfaces). The previous INVENT / INSTALL hero stays in use inline on the blog.
- **Kept** all 3 screenshots unchanged (same URLs, same order).
- `skills/hook-generator/` content is byte-identical to 1.5.0 apart from the version line in SKILL.md and
  schema.yaml.
- Staged from the signed live 1.5.0 bundle (sha256 `318aa83f…ef71e`, verified), not from drafts.

The 3-second VIDEO hook (DAS-213) — deferred since 1.0.0 for "no research backing yet". This is that research,
and the recipe it earns:

- **Added** `recipes/video-hook.md` + `references/video-hooks.md` — the first 3 seconds of a TikTok, Reel or
  Short, written as three channels: **say** (spoken, 0–3 s), **write** (text on screen, with its hold time) and
  **show** (frame 0 → 3 s). One hook maps to one row of copy-engine's `video-script` beat table and to
  video-director's shot 1.
- **Added** 11 registry items `vid-*` (53 → 64). Each is a creator's top short at ≥ 3x their own median
  (5.2x–22.5x), their own post (no collab, no paid tag), read in frames. Metric = platform-reported views +
  the measured ratio. Install the shape; the numbers belong to the creator.
- **Added** `scripts/verify-video-hook.mjs` — text present · spoken hook line within budget (EN ≤ 10 words,
  WARN ≤ 13) · on-screen text ≤ 3 words/s of hold · no greeting, filler or meta opener · frame 0 never opens
  on a logo, title card, black or fade, and names its subject (a miss only WARNs) · a set gives ≥ 3 variants,
  ≤ 1 question · labels
  video-analyst can read back. PASS/FAIL with numbers. `--single` scores one hook; `--emit-texts` feeds
  `lint-anti-patterns` and `lint-build-diary`.
- **Added** 4 learnings: the thresholds and their sources · the promise decides, polish is table stakes ·
  model from a creator's best against their worst · collab posts inflate view ratios.
- **Evidence:** 24 videos — 12 creators × their top and low, 4 each on TikTok / Reels / Shorts — read on
  2026-09-26 in words and frames with `@di-atomic/video-analyst`. A self-centred first line: tops 1/12, lows
  8/12. Text on screen by 3 s: 11/12 vs 12/12 — table stakes, not the edge. The checker FAILs 0 of 11 tops and
  6 of 12 lows (in-sample; the next corpus refresh is the out-of-sample test).
- **Peer added (required):** `@di-atomic/video-analyst >=0.1.1` — reads a reference video's opening before
  writing, judges the rendered opening after.
- **Wired:** SKILL.md routes short video to the new recipe; the line that said the video hook was "NOT
  shipped" now says thumbnails belong to `@di-atomic/thumbnail`; triggers, gate block, registry count and
  the composition chain updated. AGENTS.md + DEPLOYMENT.md carry the video route, output contract and
  consumers. schema.yaml's coordination mirror matches the new SKILL.md wording.
- **Fixed — stale facts found by a whole-bundle sweep:** README said v1.2.0, 52 exemplars and "video
  deferred", and pointed at a `source/` folder that never ships → rewritten; attribution now lives in a root
  `PROVENANCE.md` (ideas only, licences reproduced). package.yaml's header claimed it was generated by
  `_build/build-package.mjs` → corrected (there is no generator; DAS-155). SKILL.md frontmatter version (was
  1.3.0), schema.yaml version (was 1.4.0), the "v1.0.1" non-EN gate labels, DEPLOYMENT.md's "3 scripts" and
  "v1.1 backlog" lines, and platform-algorithm's "deferred recipe" row.
- **Vendored:** `vendor-scripts.mjs` (step 0-pre) installed the canonical self-resolving coordination wrapper.
- Built from the signed live 1.4.2 bundle (sha256 `984d9532…`, verified), published from a clean staged copy.

The rest of the stale card copy, which 1.4.1 should have caught in the same pass:

- **Fixed** the package `description` — it still said **52** exemplars and **three** gate
  categories, and it still described the bundle as the flat `skills:[]` shape removed back in
  1.2.1. Live has been nested `skills:[{id,schema}]` since then, so that line was actively
  wrong about the artifact a customer installs.
- **Removed** the trailing `v1.2.0:` release note from the description. Version history belongs
  in this file, not in the field that renders on the marketplace card.
- **Added** the occasion-then-platform routing rule to the description, matching SKILL.md.
- No code, recipe, registry or script changes. `skills/hook-generator/` remains byte-identical
  to 1.4.0.

Marketplace card copy caught up with 1.4.0 (no behaviour change):

- **Fixed** `marketing.long_description` and the "Fails your copy before your audience does"
  feature, both of which still said **three** gate scripts. There are five as of 1.4.0
  (`lint-build-diary.mjs` and `verify-coordination.mjs` were the additions). Understated rather
  than false, but a card that undercounts its own gates undersells the thing it is selling.
- **Added** a card feature for the launch occasion recipe — 1.4.0's headline capability was not
  represented on the marketplace card at all.
- No code, recipe, registry or script changes. `skills/hook-generator/` is byte-identical to 1.4.0.
- Built by pulling the signed live 1.4.0 bundle (sha256-verified), patching only `package.yaml`
  and this file, and publishing from that clean staged copy — per the 1.3.0/1.3.1/1.4.0 precedent.

The launch occasion (DAS-152) — the skill can now tell a launch headline from a build diary:

- **Added** `recipes/launch-hook.md` — the skill's **first occasion recipe**. Every other recipe
  answers *"which surface?"*; this one answers *"what is the moment?"*. All six existing recipes
  are platforms (LinkedIn · X · Instagram · email · multilingual · regulated), so nothing told the
  skill that a launch plays by different rules. Loaded **in addition to** the platform recipe: the
  platform sets the length ceiling, the occasion sets the subject.
- **Added** `scripts/lint-build-diary.mjs` — the enforceable half. Fails a launch headline that
  sells how the thing was built instead of what the buyer can now do. Three-level ladder:
  `1·build-diary` ❌ ("I verified 158 citations. 42 failed.") · `2·symptom` ⚠ ("Why your webinar
  gets great feedback and nobody buys") · `3·launch` ✅ ("Create webinars like Russell Brunson").
  Builder-voice ("I verified…", "how I built…", "so your agent doesn't have to") is a hard,
  non-overridable FAIL. Build vocabulary (`citations`, `pipeline`, `corpus`, `schema`, `registry`…)
  fails only when the line carries no buyer signal, so "Cold email scripts that book meetings"
  survives on its outcome.
- **Level 2 is the trap, and it is now gated.** It passes every other verifier in this skill —
  length, slop, structure diversity all say fine — because it is a *good* headline. It just
  announces nothing: a stranger reading it never learns anything shipped. It FAILS under
  `--occasion launch` and PASSES under `--occasion content`. The previous fix for this defect
  stopped at level 2, which is why the defect was reported twice on the same skill.
- **Added** first-launch vs upgrade branching, resolved by **registry lookup, not judgement** —
  `tool_help({name})`: no earlier published version → first-launch; one exists → upgrade. Carries
  the subtlety that the canonical blog slug is version-less, so an upgrade **rewrites the launch
  post in place** rather than writing a new story.
- **Wired** into SKILL.md: the decision tree is now *occasion first, then platform*, and the gate
  block runs the lint per surface — blog title, social openers, marketplace card headline, listing
  hero. The angle propagates, and last time the defect fired all four social drafts inherited it.
- Rule source: `TACTIC-founder-epiphany-story.md` §2 — "explain the vehicle, do not tease it".
- **Fixed (found while wiring the gate, outside DAS-152's scope):** `scripts/verify-coordination.mjs`
  imported `'../../../scripts/_verify-coordination-shared.mjs'`, which resolves to the bundle ROOT in
  the nested layout — but the helper is vendored **beside** the script. So the file crashed with
  `ERR_MODULE_NOT_FOUND` in every shipped nested bundle: one of the gate scripts had never actually
  run for a customer. The file's own comment already documented the correct value
  (`'./_verify-coordination-shared.mjs'`); line 23 contradicted it. Corrected to match.
- **Fixed:** with the import repaired the check ran for the first time and failed —
  `skills/hook-generator/schema.yaml` (the 6-line guidance-only stub) carried no `guidance.uses:`
  mirror of the `## Coordination` block. Added it, values copied **verbatim** from SKILL.md, nothing
  invented. Also synced the stub's stale `version: "1.2.1"` to `1.4.0`.
- Registry (53 items) and references UNCHANGED. **All 5 gate scripts now PASS — and all 5 run.**

Grader defect fix (DAS-148) — number-led titles no longer fall through to UNCLASSIFIED:

- **Fixed** `scripts/grade-against-examples.py`: the `number-led` detector gated on a fixed
  eight-noun vocabulary (`things | ways | tips | patterns | reasons | lessons | hooks | types`),
  so a well-formed number-led title whose noun was off-list returned **UNCLASSIFIED**. Live
  repro: *"The 7 checks I run on every cold email"* — a textbook listicle title the classifier
  could not see. The same hole hit `checks · files · scripts · steps · rules · gates · signals ·
  questions · traps · mistakes · reviews`, i.e. most of the house register, which is built on
  countable engineering artifacts.
- **Changed** the gate from vocabulary to **shape**: `<number> [modifier…] <plural-noun> <clause>`,
  scanned across a 4-token window so adjectives don't break the match ("5 simple ways"). A
  hardcoded list will always lag the writing; the shape does not.
- **Kept** `PLURAL_NOUNS` as a *confidence signal* rather than the gate — it now promotes
  irregular plurals (`people`, `children`, `criteria`) and acronym plurals (`CFOs`, `KPIs`)
  that an `-s` heuristic cannot reach. Added `NON_NOUN_S`, a stoplist that keeps
  *"3 is enough"* from reading as a listicle.
- **Added** UNCLASSIFIED diagnostics — a failing hook now prints the token trail it scanned
  (`"3" → scanned ['is','enough','for','anyone'] — no token read as a plural noun`), so the
  next vocabulary gap is a one-minute read instead of a debugging session.
- **Added** `tests/_fixtures-number-led-nouns.json` — an 8-hook regression set carrying all four
  DAS-148 fixtures plus enough structural variety to exercise the set-level gates. Exits 0.
- Detector ORDER unchanged, so narrative still beats a bare figure ("Yesterday I lost 3 clients"
  → `admission`) and "5 tips for 2026" still reads `number-led`, not `future-shock`. The
  original `tests/_fixtures-cfo-hooks.json` run is **byte-identical** to its pre-fix baseline.
- Content, registry (53 items), references, recipes and the other three gate scripts UNCHANGED.
  All 4 gate scripts PASS.
- Built by pulling the live 1.3.0 nested bundle from the marketplace (sha256-verified,
  signature-verified) and applying the single script fix on top — the local drafts lineage was
  NOT used, per the 1.3.0 precedent.

Stacked two-line hook — new composition technique (field-observed enrichment):

- **Added** `references/stacked-hooks.md` — the two-line opener (line 1 promise + open loop, line 2 escalating proof, cut at the see-more fold). A *composition* over any base angle, not a seventh angle, with the escalation rule + 4 anti-patterns + a grading note.
- **Added** registry item `li-stacked-double-hook` (registry now **53** items) — source: Lara Acosta (@laraacostar, 349k LinkedIn followers), field-observed 2026-07 via scrapecreators-api. New tag namespace `composition:stacked-hook`. Metric honestly labelled: the "$1M" figure is copy-internal (author's claim), NOT a source-reported engagement metric.
- **Wired** into SKILL.md universal loop (step 2 loads `stacked-hooks.md`; step 4 offers ≥1 stacked variant on see-more surfaces) + `psychology-of-hooks.md` (new Composition section). Corrected the registry count in SKILL.md (was 58, actually 53).
- Grader UNCHANGED — stacking is applied, not detected; each line still classifies under an existing structure. All 3 gate scripts still pass.
- Built by pulling the live 1.2.1 nested bundle from the marketplace (sha256-verified) and applying on top — the local drafts flat lineage was NOT used.

Guidance-only conversion (per OPVS "DEFINITIVE: does the skill DO something?" rule, 2026-06-17):

- **Removed** `schema.yaml` + all method declarations + the phantom HTTP backend. hook-generator's "methods" (generateHooks / scoreHook / auditHook) were always LLM reasoning, never callable APIs — declaring `base_url` + `auth` + methods against a non-existent backend caused upstream 404s when agents tried to call them. The old `client/schema.yaml` is preserved at `source/legacy/client-v1.1.0/` for provenance.
- **Changed** to the canonical guidance-only shape: `skills: []` + `peer_dependencies` (4 upstream OPVS coordination skills + voice-builder). Access pattern: `skill_search` → `skill_read` (reads SKILL.md directly out of the .tgz). Nothing to 404.
- **Added** `README.md` (marketplace listing) + this `CHANGELOG.md` at bundle root (security-review required-docs check). `docs.skill_guide` → `SKILL.md`, `docs.readme` → `README.md`, `docs.changelog` → this file.
- **Reframed** SKILL.md `## Methods (Tier 3)` → `## Operations (guidance-only, no callable API)`.
- Content unchanged: SKILL.md router, 4 references, 5 recipes, 52-item registry, 9 learnings, 4 scripts.

- **Added** v2.2 Coordination cross-references (`## Coordination` in SKILL.md + `scripts/verify-coordination.mjs`): agentdocs + agentmemory + opvs-protocol declared USE; agentboard n/a.

- **Added** SpiderPublish-surface awareness — `references/publishing-to-spiderpublish.md` (block-fit, `dynamic_landing` per-lead LP heroes, deploy gate) + composition chains terminating at SpiderPublish + a hosting-surface learning.

Initial release (v2.1 7-layer). SKILL.md router + schema.yaml (3 methods) + 4 references + 5 recipes + 52 EN registry items (6 multilingual gated) + 8 learnings + 3 verifier scripts. Synthesis of charlie947 + coreyhaines31 MIT upstreams; wave-1 research pack (52 source URLs). Shipped 100/100.