Conciseness

The body is dense with domain-specific, non-obvious rules (no time wasted explaining things Claude already knows), but it is noticeably padded by repetition: "embedding every word is the common mistake" appears three times (intro, Caption model, Non-negotiables), the rail-first default is restated in at least four places, and long nested parentheticals (e.g. the Theme-mode exception aside) force re-reading. Anchor 3 fits — mostly efficient but could be tightened by deduplication; not 2 because the padding is repetition of load-bearing rules rather than generic explanation.

Actionability

Highly concrete and executable: pinned install commands (sharp@0.35.3 puppeteer@25.8.0 gsap@3.15.0), exact probe commands (ffprobe, the 1fps contact-sheet ffmpeg line), full pipeline invocations (bash scripts/prepare.sh, node scripts/make-cinematic.cjs, render-theme.sh), and exact thresholds (80ms timing tolerance, luma 60/180/150, ≥0.5s captions). It misses anchor 5 because the single most-cited artifact — "pick ONE identity from CATALOG.md" (~8 references, "single source of truth for routing"), plus dna/README.md and themes/README.md — is not in the bundle, so the identity-pick step cannot be executed as written; not 3 because the commands present are copy-paste ready, not pseudocode.

Workflow Clarity

A clear 5-step pipeline with a decision gate that runs first, explicit validation at every stage (timing/occlusion/overflow gates, check-timing.cjs --strict, verbatim-completeness gate at compile time), a preview-before-render checkpoint, and explicit feedback loops: "Apply fixes in cinematic.json / theme.json, recompile, re-preview — each loop costs seconds. Render once, when the previews pass", plus recovery paths (WHISPER_MODEL=medium retry, else refuse). This matches anchor 5's sequence + validation + feedback loops + checklists (5 failure checks, 5 positive checks).

Progressive Disclosure

Structure is good in form — a Shared knowledge table of 14 reference docs (all present in references/), clearly signaled one-level-deep links, and "Skim by need" guidance. But scoring against the actual bundle: the three most load-bearing references — CATALOG.md, dna/README.md, and themes/README.md — are repeatedly cited yet absent from the bundle, and links also point outside it (../hyperframes/references/plugin-installation.md, brief-contract.md), so navigation breaks exactly where routing depends on it. That is beyond anchor 4's "minor organization gaps" and fits anchor 3's incomplete reference structure.

Specificity

Names several concrete capabilities: "Add captions or subtitles to an existing single-subject talking-head video without editing the footage", "cinematic captions embedded behind the subject", and "runs locally end to end, including transcription and subject matting". It falls short of anchor 5 because routing/policy sentences ("Route by visual identity, not by backend engine") take up space without naming additional concrete actions, and it stays above anchor 3 because more than 1–2 specific actions are listed.

Completeness

Clearly answers both: the "what" ("Add captions or subtitles to an existing single-subject talking-head video without editing the footage") and an explicit "Use for..." trigger clause enumerating concrete scenarios (plain verbatim captions, cinematic embedded captions, VFX captions, named catalog identities). This matches anchor 5 exactly; anchor 4 would require the 'when' to be less explicit than the enumerated 'Use for' list is.

Trigger Term Quality

Good natural-term coverage: "captions", "subtitles", "talking-head video", "verbatim captions", "cinematic captions", "VFX captions", and the Chinese variants "炸/特效/酷炫字幕". Not anchor 5 because common variations users would say are missing — e.g. "burned-in subtitles", "closed captions/CC", "add text to video" — and no file extensions are mentioned.

Distinctiveness Conflict Risk

A clear niche with distinct triggers: single-subject talking-head captioning with captions composited "behind the subject" via matting, including Chinese VFX trigger phrases unlikely to match any other skill. The qualifiers "without editing the footage" and "split multi-shot footage before applying it" cleanly separate it from generic subtitle or video-editing skills, so conflict risk is minimal.