rail (default) + embed (promotion). drop (filler, not shown) / rail (verbatim lower-third subtitle, in front, carries most text) / embed (a peak word composited behind the subject). Standard mode does both, embedding only the peak(s). See § Caption model.
The video is delivered UNTOUCHED (Standard/Cinematic; Theme mode's PLATE budget is the one sanctioned exception — register-gated reaction beats (charge-dim, punch, shake, grain) defined per theme DNA and applied AFTER the matte composite so subject+text+plate move as one frame) — captions are the only thing added; the matte just lets the subject occlude the embed track. Never grade/recolor/scanline the footage.
Two rulebooks: rail → references/rail.md (thin), embed craft → references/composition-craft.md (rich, embed-only). Skim by need.
SKILL.md
Plugin installs: Before setup or freshness commands, follow plugin execution rules when this skill is inside a HyperFrames plugin. Standalone installs keep the update instructions below.
First, keep this skill fresh — confirm with the user before running:npx hyperframes skills update embedded-captions. A fast no-op when everything is current; otherwise it refreshes this skill plus the core domain skills it depends on before you rely on them.
Embedded Captions
( — 35 identities; the engines behind it are backend detail). (default) builds a clean verbatim (lower-third subtitle carrying most text) + an climax composited the scene behind the subject at the peak. is pure embed — no rail, every caption composited behind the subject (hero typography, accumulation, occlusion as the effect). is a complete themed constitution — body paradigm × hero setpiece × front fx × plate reaction, composed from registries (): . Most explainer / voiceover is ; — embedding every word is the common mistake; Theme is for VFX-grade asks ("炸", "特效", "像 AE 做的").
Plugin installs use the bundled, manifest-pinned CLI for matting, transcription,
and rendering; no source checkout is required. The local preview and caption
measurement helpers also need Sharp, Puppeteer (with its Chromium browser), and
GSAP. Install these in the caption project, not inside the read-only plugin:
Keep the project's lockfile. If these dependencies already exist, use its locked
versions instead of overwriting them. Bash and FFmpeg/ffprobe must be on PATH.
Matting and transcription may download their own models on first use.
Rendering waits for the CLI to exit successfully before compositing. The old
HF_TIMEOUT_S shell watchdog is no longer used: a large partial file is not proof
that rendering finished. An explicit built-checkout argument or HYPERFRAMES_ROOT
selects the contributor CLI instead of the plugin pin. Cancel a stalled render normally through the CLI/terminal;
the caption helper does not force-kill or recover a render from a process snapshot.
Operational flow (TL;DR)
Routed through /hyperframes, the intent layer confirms only the input (which clip) and announces the identity pick as a deferred ask — the shortlist needs the probed clip, so it stays at step 1 below; the layer's run-shape questions don't apply (the footage is untouched, there is no storyboard to review). A BRIEF.md, when present, carries the confirmed input and any user notes — read it first.
The craft prose below is long; the pipeline itself is short — and everything deterministic is computed or compiled, never hand-written:
Decision gate (refuse bad clips) → pick ONE identity from CATALOG.md (35 identities; engine/compiler derived by lookup — never surface a mode/category question)
hyperframes init (skip it if the project dir already exists with the video inside — matte.cjs/transcribe.cjs adopt any video in the dir as source.mp4) → bash scripts/prepare.sh <project> (matte ∥ transcribe ∥ audio-envelope in parallel, then safe-zones v2 with scene palette/optics/lighting — one command, nothing forgotten)
author a small JSON of creative choices (read safe-zones.json first): Cinematic → plan.json → fill-timings.cjs → fit-fonts.cjs → make-composition.cjs; Theme → theme.json → make-theme.cjs (rail/panel/poem/takeover paradigms; anchor is the quiet rail default)
Visual QA: node scripts/preview-frames.cjs <project> → faithful composite previews in ~2s/frame (no render). Check § Visual QA before paying for a render.
rail (default) + embed (promotion).drop (filler, not shown) / rail (verbatim lower-third subtitle, in front, carries most text) / embed (a peak word composited behind the subject). Standard mode does both, embedding only the peak(s). See § Caption model.
The video is delivered UNTOUCHED (Standard/Cinematic; Theme mode's PLATE budget is the one sanctioned exception — register-gated reaction beats (charge-dim, punch, shake, grain) defined per theme DNA and applied AFTER the matte composite so subject+text+plate move as one frame) — captions are the only thing added; the matte just lets the subject occlude the embed track. Never grade/recolor/scanline the footage.
clean lower-third subtitle, in front, readable. A punch word can get an inline emphasis highlight (accent colour / active-word pop) — it stays on the rail.
embed
a promoted peak — the headline beat
one big word composited behind the subject (matte occlusion), designed entrance + exit
The rail carries most of the text; embed is the scarce, earned peak. Scarcity is per beat/block, not per clip: ≤1 hero per block (thought), never two co-visible, ≥ a beat of air between hero windows (the compiler warns under 0.6s). A short clip → usually 1–2; a long explainer → ~one per section. Among multiple heroes, the largest authored one is the APEX (it alone gets the full lockup embed + width-fit raise); smaller ones are MINOR peaks that ride their column as oversized emphasis lines (fg, damped motion) — not every beat needs the matte showcase, which is exactly what keeps the apex an event. Embedding every word is still the common mistake.
Rail-surface identities build exactly this (rail = rail.html, embed = the climax in index.html). Column-flow identities drop the rail and make everything embed-style — recommend them only for mood-over-verbatim asks, never for explainer / voiceover where the words must read (CATALOG.md encodes this per identity).
Step 0 — pick ONE identity from the CATALOG
One front-end, three engines behind. The user picks an IDENTITY from CATALOG.md (35 entries: 10 classic + 25 themed); the engine, compiler and authoring file are derived by lookup from the catalog row. Never surface "Standard vs Cinematic vs Theme" as a question — those are backend names (a product has one UX even with several engines). The catalog encodes everything routing needs: reading surface, voice, recommend-for, scene needs, adjacency notes for the genuinely-close pairs (loud↔ordnance, neon↔neonsign, cream↔stardust).
The identity pick is a preference gate (../hyperframes/references/brief-contract.md § 1): in autonomous mode ("surprise me" / "decide for me"), pick from your shortlist yourself and state the one-line why instead of asking.
Procedure: probe the clip → shortlist 2–3 identities from the catalog → recommend ONE with a one-line why → the user picks (autonomous mode: you pick, stating the why) → author that identity's file. Identities are engine-locked (no cross combos; opening one is a validation event — see dna/README.md).
Always present your recommendation and let the user pick before you author. Don't silently default.
(The full identity table lives in CATALOG.md — single source of truth for routing. The engine docs below describe each backend's authoring contract.)
CATALOG.md is the whole answer space here: this workflow does not search the HyperFrames component registry. The composition workflows run npx hyperframes catalog before authoring a named look; this one must not. Its engines are locked compilers that consume cinematic.json / theme.json and emit the composition themselves, so a registry item — the caption-* blocks included — has nothing to mount into. A registry block styles text on a designed canvas; this skill burns captions into somebody's footage through a matte. When no identity fits the ask, say so and pick the nearest, rather than reaching outside the catalog.
Recommendation heuristic: use the "Shortlisting heuristics" in CATALOG.md — they are identity-level (e.g. "炸" shortlists ordnance/stomp/terminal/loud and picks by WHAT should explode), never category-level. Unsure → anchor.
Cinematic → write plan.json for a locked template, compiled by make-composition.cjs.
Multiple speakers / hard cuts (split & render each shot, or refuse)
No human subject (this skill is for talking-head)
Under 3 seconds, no speech, or face never clearly visible — transcribe.cjs warns when audio is near-silent (Whisper hallucinates words like "Thank you." over silence); heed it and refuse rather than caption fabricated words
Source already has burned-in captions / subtitles / heavy text graphics — adding a second caption system conflicts and the footage ships untouched (no covering/inpainting). Burned text often appears only mid-clip: sample a 1fps contact sheet (ffmpeg -i in.mp4 -vf "fps=1,scale=160:-1,tile=10x5" sheet.png), don't trust 3 spot frames.
Transcript is garbage — non-native/heavy-accent speech can transcribe into confident gibberish. Sanity-read transcript.json before authoring; if it doesn't parse as language, try WHISPER_MODEL=medium once, else refuse (a verbatim rail of fabricated words is worse than no captions).
Busy handheld with fast motion (matte flickers)
Pre-flight probes (cost nothing, prevent the worst failures)
Shot-cut probe. Sample frames at 20%, 50%, 80%. If a different subject/scene appears, trim the clip before the cut.
Letterbox / pillarbox probe. Black bars on the first frame? Compute safe content rect and constrain caption placement inside it.
Luminance probe. Sample the caption region's average luminance — under 60 → light text reads as-is, 60-180 → add the glyph scrim, 180+ → opaque text + scrim (never bare light text). Cinematic templates are cream+screen and LOCKED — use this probe to pick a fitting identity (bright scenes → ink, or the opaque-rail anchor theme), never to recolour one.
Identity recommendation by tone (you recommend; the user picks — see Step 0 + CATALOG.md). explainer / interview / must-read words → rail/panel-surface identities; poetic / social / "cinematic" → column-flow identities by register; "炸 / 特效 / VFX" / named worlds → themed identities. When unsure → anchor (words read, scene safe) — but present a shortlist and let the user choose.
Step 1's init checks the installed skills against the latest on GitHub and updates the global set if any are out of date.
Step 3 differs by mode:
Step 3 — Cinematic mode (pure embed)
Read safe-zones.json first. Narration planes go in zones.hugLeft/hugRight — clean strips ABUTTING the silhouette (text far from the body reads as floating, not embedded; far corners are the fallback, not the default). The hero defaults to heroAnchor/heroBands.best (centered ON the subject, ~30–55% occluded). recommendation:"fg" moves NARRATION in front for legibility; the hero stays embedded whenever heroBands.feasible — hero-fg is the last resort.
The DNA is the identity you picked in Step 0 (CATALOG.md) — do not re-open the choice here. Sanity-check it against the scene (bright hero band luma > 150 wants ink; full pick guidance lives in the catalog, covering all ten incl. neon / glitch / chrome / velocity). State your pick + why; the user decides. The DNA locks type/palette/blend/motion + hero three-act; safe-zones v2 (palette/optics/lighting) parameterizes it to THIS scene automatically.
Author <project>/cinematic.json — "dna": "<name>" + thought-BLOCKS, not raw groups: each block = lines of words (grouped 2–5 at clause boundaries) + the plane it stacks in + per-line css (size/weight/style only — no positions) + at most ONE line marked "hero": true (the promoted word; "text" for display form). Schema: scripts/make-cinematic.cjs header.
Compile: node scripts/make-cinematic.cjs <project> — lowers blocks → plan.json → index.html. Generated for you: transcript-sequenced timings, accumulate-within-block, page-flip-between-blocks, the hero LOCKUP (a hero block's pre-context, HERO and post-context stack as ONE bonded composition centered on the subject — reading order top→bottom = spoken order by construction; context floats in FRONT while the hero embeds BEHIND = the depth sandwich; a mass rule keeps the hero dominating its context), apex/minor hero split, reading order by construction, fg fallback per safe-zones. Then the gates run as usual. (Hand-authoring plan.json directly remains possible for designs blocks can't express — then run fill-timings.cjs + fit-fonts.cjs + make-composition.cjs yourself.)
Step 3 — Theme mode (themed constitution)
Read themes/README.md FIRST — paradigm/setpiece registries, linkages, hard rules, and the exact theme.json schema.
Pick a theme DNA by content register (each themes/<name>.json has voice + when). State your pick + why; the user decides.
Author <project>/theme.json — dna, lines (verbatim, transcript order; 1–5 words each — for takeover each line is one CARD), minors (emphasis words), hero:{match} (the climax word/phrase; leave it OUT of lines for embed setpieces, keep it IN for inline setpieces and panel+redact).
Render: bash scripts/render-theme.sh <project> — compiles (verbatim-completeness gate at compile time), renders both layers, composites, applies the plate reaction → final_fx.mp4. Use preview-frames.cjs between compile and render for Visual QA.
Visual QA — preview BEFORE you render
node scripts/preview-frames.cjs <project> [t…] composites faithful preview frames in ~2s each (caption layers screenshotted at seek-time + real video frame + matte occlusion + rail overlay = what the final composite will look like at that moment). Default samples = each group/climax window. A full render costs minutes — never use it to discover layout problems.
Check the previews (<project>/preview/sheet.png) against this list — these are the failures the geometric gates cannot catch:
Washout — light text over a bright region (window/sign/sky): unreadable → move the plane or change DNA/mode (bright scene → ink).
Text-on-text — captions over the scene's own text/graphics, or two caption groups colliding.
Reading order — on-screen vertical order must match spoken order; the hero must not sit below later words.
Hero presence — the climax should be BIG and visibly behind the subject (~30–55% occluded), not a floating label in a margin.
Balance — one coherent column/band, not scattered fragments; margins breathing; nothing clipped.
Then the 5 positive checks in references/reference-bar.md (poster test · timid test · one-glance hierarchy · scene handshake · dead-air audit) — the failure list keeps a render from being broken; the positive list is what makes it designed. Ship when both pass.
Fresh-eyes review (recommended for anything user-facing): you have confirmation bias about your own layout. If you can spawn a subagent, give it ONLY the preview sheet + this checklist and ask for PASS/FIX verdicts per frame ("review these caption previews against the 5-point checklist; answer PASS or the specific fix per frame"). Apply fixes in plan.json / theme.json, recompile, re-preview — each loop costs seconds. Render once, when the previews pass.
The DNA registry — ten visual languages (replaces the template catalog)
Both modes draw from dna/ — ten art-directed visual languages that parameterize per scene (accent sampled from the footage, contact shadow along the measured light direction, depth-match blur, RMS-coupled hero amplitude):
DNA
Register
Scene fit
Voice
cream
premium-warm
dark/mid warm scenes
Inter + warm cream + screen; glowing emergence hero (successor of cinematic-cream)
ink
premium
bright scenes (luma > 150)
near-black multiply — type printed ON the wall; the bright-scene answer
Anton + scene-sampled accent, single-unit slam + ripple; body ANNOUNCES in front (bodyLayer: fg)
neon
loud-neon
neon-noir / nightlife / tech-noir (dark scenes)
electric-cyan signage, ignition flicker, the hero powers ON like a sign
Pick by safe-zones.json (heroAnchor.bandLuma, palette.temperature) × content register — dna/README.md has the decision rule. Authoring: cinematic.json takes "dna": "<name>".
The engine generates the hero three-act from the DNA (no authoring needed): co-visible captions dim (setup) → per-letter entrance with amplitude ∝ spoken loudness (impact) → breathe + glow until exit (afterglow).
(Legacy: plan.template:"cinematic-cream" maps to dna:"cream" automatically. The retired 54-template library is archived outside this repo and is not distributed with the skill; _motion.md remains in-skill as the motion-verb reference catalog.)
Aesthetic decision — tone × shot × platform (input to the catalog shortlist, NOT a second router)
Classify the clip on 3 axes and feed the result into CATALOG.md's shortlisting — this section never picks a mode/engine by itself:
Composition craft (embed track) — read before embedding
The full embed-track playbook lives in references/composition-craft.md: transcript role-annotation, phrase grouping, planes & clean-zone anchoring, zone coherence, climax pop & readability, edge-breathing, the occlusion 3-step judgement, and accumulation/persistence. It governs how a promoted phrase sits INTO the scene — read it before authoring any embed (Cinematic plan.json or Standard index.html). The default rail track has its own, much simpler spec → references/rail.md.
Bugs already locked out (CoreML, letter-spacing reflow, etc.)
Read the aesthetic principles and direction catalog FIRST. Everything else is implementation detail.
Non-negotiables
Face must never be 100%-covered continuously — every 0.3s window, face bbox ≥30% uncovered.
WCAG contrast — final render lints; fix palette if it fails.
Deterministic — no Math.random(), no Date.now(), no repeat:-1.
Never grade/recolor the video. The footage ships untouched — captions are the only addition. No full-frame scanlines / duotone / darken / vignette over the a-roll. neon-noir/CRT texture belongs inside a caption element, not over the whole frame.
Rail-first for talking-head / explainer. Don't embed the whole transcript — most text is the rail; embed only peaks. Embedding everything is the default mistake.
Embed is scarce + spaced. ≤1 embed per sentence/beat, never two adjacent or co-visible, ≥ a beat apart, at most one apex. climax = per-beat peak, not "the single payoff of the entire clip."
Matte = the PERSON (hyperframes remove-background, u2net_human_seg, Apache-2.0). Human segmentation by intent, but not surgically: thin offset furniture (mic boom arms) is usually excluded — captions render over it, behind the person — while large salient objects NEAR the subject (a telescope, a desk rig) can still leak into the matte and occlude captions. Objects HELD by the subject (products, phones) may drop out intermittently, letting captions pass in front. NEVER assume: sample frames_fg/ at 2-3 timestamps before placing the hero, and prefer hero positions clear of any leaked furniture (heroAnchor can be skewed by leaks — cross-check against frames_bg).
safe-zones is PROP-BLIND — eyeball every band you use. Zones/heroBands score subject occlusion + luma only: a mic, telescope, or screen sitting inside a "clean" zone is invisible to them (and a prop leaking INTO the matte skews heroAnchor.centerXPct off the person). Before authoring, extract ONE frame of each band you intend to use; if a prop lives there, measure its bbox and move/shrink the plane. Two real cases shipped clean only because the agent did exactly this. (Auto prop-saliency is a known gap; zones' peakLuma only catches moving bright objects.)
Captions stay on-frame. Cinematic mode hard-gates frame-overflow; Standard mode runs check-overflow.cjs as a WARNING (intentional bleed is the only exception — read the warning).
Each caption ≥ 0.5s on screen — shorter = unreadable.
Word timings must match transcript.json within 80ms — a caption firing 500ms off-beat destroys the scene illusion. Cinematic runs check-timing.cjs --strict before rendering (via render-and-composite.sh); THEME mode enforces the same timings at compile time instead (make-theme's sequential transcript matcher + verbatim completeness gate — drift is a compile error). Never pack multiple transcript words into one entry (e.g. "FUTURE OF" or an IT + line-break + ALL stack with one start/end) — the second word inherits the first's timestamp and fires early. Split them into separate word entries with their own timings, even if you want them on the same visual line (use CSS white-space / natural wrap instead of <br>). Creative substitutions where caption text ≠ transcript (e.g. "15%" replacing "fifteen percent") are supported — register them in CREATIVE_SUBS inside check-timing.cjs.
Group windows must envelop their words — group.in ≤ min(word.start) and group.out ≥ max(word.end) for every group. If group.in is later than a word's start, the word is silently delayed until the container mounts (we've shipped 800ms lag bugs from this). The validator enforces this.
No two caption groups may overlap in both time AND screen region — overlapping-in-time captions create text-on-text pileups. Options: (a) spatial separation — place each group in a non-overlapping vertical band so they can coexist (memory-wall cascade style); (b) handoff — set the earlier group's out ≤ the next group's in so only one is on screen; (c) deliberate layered typography — add "allow_overlap": true on one of the groups to silence the validator. The validator estimates each group's vertical bbox from its CSS and flags collisions. Pick (a) by default — it's what makes cinematic-cream feel like a poem accumulating, not a subtitle track replacing itself.
Screen-blend fails on bright backgrounds (>180 luminance).Cinematic templates are cream + screen and that DNA is locked (the plan can't recolour them) → on a bright backdrop they wash out, so pick ink (letterpress built FOR bright surfaces) or the anchor theme (opaque rail surface) rather than overriding a look.
Don't animate letter-spacing or filter:blur on word entrance — inline-block reflow causes line-jumps.
CoreML banned for matting — the onnxruntime CoreML EP's mixed-precision partitioning corrupted face alpha (observed with the previous RVM engine; don't re-try it). Matting is CPU-only (~2 fps @1080p ≈ 2-3 min per 10s clip; budget for it on long clips).
Dependencies
HyperFrames CLI: plugin installs use the bundled manifest-pinned launcher. Source contributors can use a built checkout (packages/cli/dist/cli.js) via HYPERFRAMES_ROOT, the skill’s source tree, or ~/Downloads/hyperframes.
Node-first; two Python touchpoints via uvx (no manual installs): transcription runs WhisperX through uvx (word-level timings; falls back per SKILL §transcription), and Theme's drawon setpiece shells python3 scripts/gen-stroke-path.py at compile time. Everything else runs on the toolchain hyperframes already ships: matting via the hyperframes CLI's remove-background (u2net_human_seg; weights auto-download once, ~168 MB, to ~/.cache/hyperframes/), image/alpha math via sharp, layout/occlusion/overflow via puppeteer, plus ffmpeg. Install Sharp, Puppeteer, and GSAP in the caption project as described in Runtime prerequisites above. The helpers check that project first and retain checkout dependency lookup for source contributors.
Transcription = WhisperX via uvx (word-level timings + alignment; no manual install — transcribe.cjs drives uvx whisperx). Falls back to an existing word-level transcript.json if present.
Source video — matte.cjs / transcribe.cjs auto-resolve source.mp4 (or glob the clip / read hyperframes.json), so hyperframes init --video X.mp4 needs no manual rename.
fps — matte.cjs extracts at the source's native rate and records matte.fps; render-and-composite.sh uses that so the matte stays frame-aligned.
Matting weights are NOT bundled: matte.cjs shells the hyperframes CLI's remove-background, which downloads u2net_human_seg (~168 MB, Apache-2.0) once to ~/.cache/hyperframes/background-removal/models/. First prepare on a fresh machine needs network for that one download.
If a hard dependency is missing, STOP and ask the user — don't silently skip steps.
glitch
loud-neon
digital / hacker / AI
RGB-split echoes snap together on landing; machine-percussive timing
chrome
loud-luxe
Y2K / fashion-tech / music
liquid-metal gradient hero + one sheen sweep during the hold
velocity
loud-sport
sport / auto / fitness
every word arrives along its motion vector (streak+skew), hero passes with speed trails