Conciseness

The body is dense with unique domain knowledge (no filler, no explanations of concepts Claude already knows), but it is noticeably duplicated rather than lean: the 5-step pipeline is stated twice ('Operational flow (TL;DR)' and 'Pipeline — 5 steps'), the decision-gate probes appear in two sections, 'embedding every word is the common mistake' is repeated four times, and a 10-row DNA registry table is inlined alongside a dedicated dna/README.md. This fits anchor 3 ('mostly efficient but could be tightened') better than 4, where only minor trimming would be needed — merging the duplicated sections would cut substantial length. It is above anchor 2 because the verbosity is structural repetition of load-bearing rules, not padded or generic explanation.

Actionability

Fully executable throughout: copy-paste commands ('bash scripts/prepare.sh <project>', 'node scripts/preview-frames.cjs <project>', exact ffprobe/ffmpeg probe commands with flags, pinned 'npm install ... sharp@0.35.3 puppeteer@25.8.0 gsap@3.15.0'), exact numeric thresholds (luma > 150, ≥30% face uncovered per 0.3s, 80ms timing tolerance, 0.5s minimum on-screen), and pointers to exact schemas ('Schema: scripts/make-cinematic.cjs header'). Matches the anchor-5 pattern of copy-paste-ready commands covering the common cases; anchor 4's 'minor gaps' does not apply.

Workflow Clarity

The sequence is explicit and numbered (init → prepare → author JSON per mode → preview QA → render, with a 'Decision gate — RUN FIRST' before either mode), validation is explicit at every stage (compile-time verbatim gate, check-timing.cjs --strict, render gates for timing/occlusion/overflow/hand-off, a 5-point Visual QA failure checklist plus 5 positive checks), and there is a genuine feedback loop ('Apply fixes in plan.json / theme.json, recompile, re-preview ... Render once, when the previews pass', plus the fresh-eyes subagent review). This matches anchor 5's 'explicit validation steps; feedback loops for error recovery; checklists'; anchor 4 is ruled out because no checkpoint is missing.

Progressive Disclosure

Structure is genuinely good — a Shared knowledge table mapping 14 reference docs, per-need pointers ('skim by need', 'read before embedding') — but scored against the actual bundle: the three most load-bearing targets are missing. CATALOG.md (referenced ~10 times, called 'single source of truth for routing'), themes/README.md ('read FIRST — ... the exact theme.json schema'), and dna/README.md ('has the decision rule') do not exist in the bundle, so the routing and theme-schema steps dead-end; the inline 10-row DNA table also duplicates the (missing) registry file. This is worse than anchor 4's 'minor organization gaps' but better than anchors 1–2 (no deep nesting, no monolith; references are clearly signaled), so anchor 3 fits.

Specificity

Concrete actions are named — 'Add captions or subtitles to an existing single-subject talking-head video without editing the footage' plus 'runs locally end to end, including transcription and subject matting' — covering the domain with only minor gaps (compositing/preview/render steps are only implied by 'workflow'). It sits at anchor 4 rather than 5 because the action list is one core action plus supporting pipeline capabilities, slightly less comprehensive than the anchor-5 pattern of several distinct enumerated actions.

Completeness

Both questions answered explicitly: what — 'Add captions or subtitles to an existing single-subject talking-head video without editing the footage'; when — 'Use for plain verbatim captions, cinematic captions embedded behind the subject, VFX captions, 炸/特效/酷炫字幕, or a named identity from the 35-style catalog.' This matches the anchor-5 good example structure ('what' + explicit 'Use when/for' with concrete trigger phrases), and is above anchor 4 where the 'when' is merely present but less specific.

Trigger Term Quality

Natural terms are comprehensively covered with synonyms across languages: 'captions', 'subtitles', 'plain verbatim captions', 'cinematic captions embedded behind the subject', 'VFX captions', and the Chinese variants '炸/特效/酷炫字幕' that a user would actually say. Not below 5: no common synonym family for this task is missing; not applicable above 5.

Distinctiveness Conflict Risk

A clear niche — captioning existing single-subject talking-head footage with the footage untouched — with distinct triggers (verbatim/cinematic/VFX/Chinese VFX terms) and an explicit routing disambiguator ('Route by visual identity, not by backend engine'). Minimal overlap risk with generic video-editing or composition skills; anchor 5 fits, and anchor 4 (minor overlap with closely related skills) is weaker since the 'without editing the footage' boundary separates it from editing skills.