Copy Engine

Video scripts. copy-engine now writes the spoken words of a video as a timed beat table: one row per beat, with the voiceover, a word budget and the on-screen text, and one table per language. It routes to craft that already exists instead of re-teaching it.

- **`recipes/video-script.md`** routes by video type. A selling video (VSL, demo, ad) takes its structure from `@di-atomic/webinars`, a 15 to 60 second short takes the short-form shape, and a tutorial takes `@di-atomic/video-director`'s chaptering. The story comes from this skill's own registry, and the first 3 seconds come from `@di-atomic/hook-generator` `recipes/video-hook.md`. The output is the beat table, which goes to video-director's `verify-vo-fit`. FAIL beats come back with `max_words` per language, and only those beats get rewritten.
- **`references/spoken-register.md`**: one breath per sentence, budget before you write, never read a URL or a package name aloud, say what a number means, pauses as explicit markup per TTS engine (never `...`), budget the close first, names are heard, and localize rather than translate.
- **`scripts/lint-spoken-register.mjs`** (the 8th script) checks what the writer controls and a clock cannot see: a URL, package name or code read aloud, `...` used as a pause, and digits. It emits the voiceover text, the beats JSON and the lines for the clock and the TTS stopwatch. **It sets no budget.** The word budget and the fit are `verify-vo-fit`'s, from video-director's one speech-rate table, counted in spoken tokens. That method catches 11 of 11 known overflows, where this script's unreleased first draft, with its own words-per-window budget, caught 10.
- **Worked example, gated on the real clock** (video-director 0.2.1, edge-tts audio): section-designer's own 60-second demo script went from 5 of 7 beats over and 67.53 s of speech to **0 of 7 in English (35.19 s) and Russian (37.22 s)**, words only. The hook beat now opens on the product rather than a title card. The Russian is an in-session translation awaiting native review.
- Added `learnings/2026-09-26-total-fits-beats-fail/` (17 learnings, was 16). A script can fit its runtime and still fail beat by beat.
- **Fixed: `skills/copy-engine/schema.yaml` was invalid YAML** in every 0.6.x release. Its first three lines began with `>` instead of `#`, and a strict parser rejected it at 1:3, while the package still scored 100/100 HEALTHY. The skill-level version inside it moves to 0.7.0.

Two verifier false-positive classes fixed. No change to what the skill teaches or produces — this release only stops its own gates from refusing correct copy.

- **`lint-cliches` refused correct technical nouns.** `harness` was hard-banned as a bare token, so a shipping LinkedIn article FAILed on two noun uses (`A skill is not a prompt. It is a harness.`) with zero verb uses in the file. `harness` and `leverage` now match the marketing VERB construction instead of the token: `harness the power of` and `we leverage cutting-edge synergy` still FAIL, while `a skill is a harness` and `the highest-leverage answer` pass. The other 24 hard-ban tokens are untouched.
- **Both word linters scored their own teaching material.** An article that quotes a spam example to warn against it, or names the banned vocabulary in a list, was counted as using it. A shared use-vs-mention pre-filter now excludes fenced code, inline code, blockquotes and `(a, b, c)` vocabulary lists, and `<!-- lint-specimen -->` marks what no rule can safely detect. Every run reports exactly what it excluded; `--no-strip` re-scores the raw text.
- **The filter cannot be used to hide copy from its gate.** When too little prose survives to judge (under 50 words), both linters re-score the RAW text and report that verdict, loudly. Spam pasted wholesale as a blockquote scores 19 and FAILs.
- **Rejected: treating quoted spans as mention.** Measured, it took real spam from 9 (FAIL) to 1 (PASS) without tripping the guard, because scare-quotes around salesy terms are ordinary marketing usage. The rejected design is recorded in the helper and pinned by a regression fixture so it is not re-added.
- Added `learnings/2026-08-24-use-vs-mention-breaks-word-linters/` (16 learnings, was 15) and an 11-case regression suite at `tests/fixtures-das-144/`.
- The skill now passes its own two word-matching gates across **73/73** markdown artifacts, up from 63/73 (cliches) and 72/73 (spam).

⚠ The originating bug report attributed the second defect to fenced code blocks. It did not reproduce: the file it cited contains zero fences and fence-stripping changes its score by 0. The real carriers were a bold bad-example line and a parenthetical vocabulary list.

Packaging fix — the Layer-7 coordination gate now actually runs from the shipped bundle.

- **`scripts/verify-coordination.mjs` crashed on import in every prior release.** It imported
  `'../../../scripts/_verify-coordination-shared.mjs'` — a path that resolves outside the bundle —
  while the helper was never staged beside it. The gate died with `ERR_MODULE_NOT_FOUND` on every
  run and therefore verified nothing, silently, at 100/100 HEALTHY.
- **The wrapper is now self-resolving** and the shared helper is vendored into the bundle. No
  behaviour change to the skill itself — this release only makes its own quality gate functional.

Marketplace card only. No change to the skill's content, shape, methods or peers.

- **Added the SPEC-3 asset kit.** Top-level `media:` block with a hero and three panels, rendered in the house dark-board chalk style: six passing checks against one unanswered question, the em-dash count 17 against 0, the Epiphany Bridge's five phases, and Q13 achievement against Q14 transformation.
- **Added `marketing.long_description`** (181 words, first-person, 25 numbers) and **five feature bullets**, all verb-led and under 90 characters. Both were absent, so the listing had no body copy and no images.
- `demo_video_url` is deliberately omitted rather than empty; an empty string fails publish validation.

**Adds the story-led register, and migrates the bundle onto the house nested shape.** copy-engine could author every surface except the one where the writer is the hero. The content change is additive — no existing recipe, registry item, reference, learning or script was modified. The packaging change is not: this is the first release off the flat `skills: []` meta-package shape.

- **Added `recipes/story-led-article.md`** — the founder's-epiphany register, distinct from `blog-article-skeleton.md` (thesis → evidence → reframe). Collects the 14 answers, gates on four verifiers, closes on transformation rather than achievement.
- **Added 3 registry items** — `epiphany-bridge-script` (Brunson, *Expert Secrets* Secret #9 pp213-222: the 5 phases / 14 questions) · `heros-two-journeys` (Hauge + Vogler via Secret #8 pp197-211: plot triad, 5 rapport devices, 4 desires, achievement vs transformation) · `narrative-voice-bars` (the measurable register targets).
- **Added `scripts/verify-narrative-voice.mjs`** (7th script) — four bars: em-dashes · 1st-person /1k · 2nd-person /1k · avg sentence, with `--mode story|explainer`. Strips fenced code before measuring.
- **Corrected the scope note.** Story-led long-form is now supported; only the *explainer* long-form remains skeleton-only.
- Counts (measured on disk, not copied from SKILL.md): recipes 7→8 · registry 40→43 · scripts 6→7 · references unchanged · learnings unchanged.
- **Migrated to the house NESTED bundle shape.** Was the flat `skills: []` meta-package (0.4.0-0.5.1 lineage); now ships `skills:[{id: copy-engine, schema: ./skills/copy-engine/schema.yaml, visibility: public}]` with a methodless `schema.yaml` carrying the skill-level `guidance:` block (use / not / chains / limits / pitfalls / case_study), matching every other @di-atomic guidance skill. `skill_read` now resolves SKILL.md from the schema's own directory; SKILL.md ships byte-identical at bundle root and at `skills/copy-engine/`. Also adds `type: skill`, `visibility: public`, `pricing`, `support` and a `marketing.author` block — the last was absent, so the listing had no author.
- **Added two story triggers to SKILL.md frontmatter** — `tell my founder story` and `write the launch article`. The story route existed in the decision tree and alias list but nothing in `triggers:` would surface it, so the new register was unreachable from a natural ask.
- **Fixed pre-existing count drift in SKILL.md.** It claimed 6 recipes and 5 scripts while disk already held 7 and 6 — `example-shipping-blog-with-inline-images.md` and `verify-coordination.mjs` were both unlisted. Three further stale claims were caught at ship time: the frontmatter `version` still read 0.5.1, the frontmatter description still said "40 exemplars + 5 scripts", and the layers block still said 13 learnings against 15 on disk. Same defect class as outbound-engine's verifier miscount: a skill that cannot count its own files.

**Why it exists.** A live Di-Atomic long-form draft passed every existing gate and still read cold. Measured after the fact: 17 em-dashes, 16.8 first-person/1k, 11.1 second-person/1k against a house target of 53.9, 13.4-word sentences. None of that was visible by reading it. `verify-narrative-voice` now FAILs that draft on all four bars and PASSes the story-led rewrite of the same material.

**Known gap it does not fix.** `lint-cliches` hard-bans `harness` as a marketing verb and has no part-of-speech awareness, so it FAILs correct technical usage of the noun ("a skill is a harness"). Documented in the recipe's gotchas; needs a fix in the linter.

Fixes the botched **v0.5.0** publish. Same content as the v0.5.0 entry below; no other changes.

- **v0.5.0 shipped as an empty 133-byte tarball** — the publish CLI was pointed at `.claude/skills/di-atomic-copy-engine/`, which is a **symlink** to this real package directory. The CLI did not follow the symlink and packaged nothing. Consumers of `@di-atomic/copy-engine@0.5.0` should upgrade to 0.5.1 (or stay on 0.4.1). Root cause + upstream fix documented in the AgentBoard card comment on `2d4cfd8e-f2bb-4db8-8e73-ddc5b13930c6`.
- **Publish path lesson learned** — always publish from the real source directory (`docs/di-atomic-skills/drafts/di-atomic-copy-engine-v0.2.0/`), never from the `.claude/skills/*` symlink.

Experience wave. Additive only — no scope change, no restructure, no removed content.

- **Added** `learnings/2026-07-19-spiderpublish-body-shape-differs-by-surface/` (KI shape). The blog `content_create_post` endpoint auto-derives a Tiptap AST from `{markdown, _authored: "markdown"}` on write; `changelog_create` and `content_create_doc` do NOT — sending only markdown publishes with a blank body div on the live page. No `changelog_update` MCP tool exists and re-creating with the same version returns 409, so the shape must be right on first publish. Learning documents the root cause + the pre-built Tiptap doc shape that works.
- **Added** `learnings/2026-07-19-dual-register-within-one-piece/` (KI shape). Long-form Di-Atomic launch blogs mix two named registers cleanly separated by beat: Mission-lean for story beats (openers, turning points, reflection close), Precision for mechanism beats (versions, thresholds, benchmarks). Enforcing one register across the whole piece flattens the read. Documents the beat-level split used in the shipped 2026-07-17 section-designer v0.3.0 launch blog (Flesch 69.5, 0 em-dashes, 3-verifier PASS).
- **Added** `recipes/example-shipping-blog-with-inline-images.md` — verbatim worked example from the shipped section-designer v0.3.0 launch. Supersedes the "skeleton + intro only" scope on `blog-article-skeleton.md` for the specific case of Di-Atomic skill launch blogs. Full-length drafting (800-1200 words) with 3-5 inline illustrations is now a proven, dogfoodable shape. Live proof at https://di-atomic.com/blog/shipping-di-atomic-section-designer-v0-3-0.
- **Added** `references/spiderpublish-body-shape-matrix.md` — surface × body-shape × behavior lookup table + copy-paste-ready Tiptap AST fragments (paragraph, heading, image, code block, inline code marks). Written to close the silent-failure trap where a markdown-authored body publishes blank on `/changelog` and `/docs/*` while the same shape works on `/blog/*`.
- **Appended** to `references/spiderpublish-handoff.md` — new "Body-shape gotcha per surface" section at the bottom. Fills the docs/changelog gap flagged in the v0.2.1 handoff contract with a concrete pre-built Tiptap AST recipe per surface.

SKILL.md decision tree, `client/schema.yaml` (absent — guidance-only), `package.yaml` peer ranges, tier (`beta`), and provider_config placeholder unchanged. Scripts, existing learnings (13), existing recipes (6), existing references (6), registry (40) all unchanged.

OPVS pre-review fixes (see vendor reply 2026-06-16 "DEFINITIVE: guidance-only skills are first-class"):

- **Fixed** unsatisfiable peer version ranges. `@opvs-ai/agentdocs` and `@opvs-ai/agentboard` were declared `>=3.0.0` but the live marketplace versions are `1.1.0` and `1.2.0` respectively. Corrected to `>=1.1.0` / `>=1.2.0`. Without this the meta-package would fail peer resolution at install.
- **Fixed** `manifest.docs` field shape — `docs.readme` now points at `README.md` (the marketplace listing), `docs.skill_guide` points at `SKILL.md` (what `skill_read` returns), `docs.changelog` points at this file.
- **Added** `CHANGELOG.md` at bundle root (this file) to satisfy the security-review required-docs check (README.md + SKILL.md + CHANGELOG.md must all be present at bundle root).

No content or behavior changes vs 0.4.0. SKILL.md, references/, recipes/, registry/, learnings/, scripts/ unchanged.

Restructured as guidance-only meta-package per OPVS reply 2026-06-16:

- **Removed** `schema.yaml` and all method declarations. copy-engine's "methods" were always LLM reasoning, never callable APIs. Per OPVS gateway architecture, declaring methods + `base_url` + `auth` against a non-existent backend was causing upstream 404s when agents called the methods.
- **Added** `skills: []` + 5 `peer_dependencies` (4 upstream OPVS coordination skills + voice-builder) — the canonical guidance-only shape.
- **Access pattern:** agents now reach copy-engine via `skill_search` → `skill_read` (substrate tool that reads SKILL.md directly out of the bundle .tgz). No HTTP call. Nothing to 404.
- v2.2 Coordination cross-references in SKILL.md mirror the `peer_dependencies` list.

- **Added** v2.2 Coordination cross-references section to SKILL.md (locked 2026-06-16).
- **Added** `scripts/verify-coordination.mjs` Layer 7 enforcement (4 surfaces declared).
- Other 7-layer content unchanged from 0.2.x.

- 7-layer v2.1 retrofit: added `recipes/`, `instructions/`, `scripts/` (mandatory).
- Added SpiderMedia + SpiderPublish handoff contracts.

Initial release. 4 methods (draftCopy / editCopy / scoreCopy / auditCompliance) with bundled exemplar registry.