embedded-captions

Embedded Captions

One catalog, picked up front (CATALOG.md — 17 identities; the three engines behind it are backend detail). Standard (default) builds a clean verbatim rail (lower-third subtitle carrying most text) + an embed climax composited into the scene behind the subject at the peak. Cinematic is pure embed — no rail, every caption composited behind the subject (hero typography, accumulation, occlusion as the effect). Theme is a complete themed constitution — body paradigm × hero setpiece × front fx × plate reaction, composed from registries (themes/README.md): ordnance terminal neonsign stardust stomp. Most explainer / voiceover is Standard; embed is the scarce, earned peak — embedding every word is the common mistake; Theme is for VFX-grade asks ("炸", "特效", "像 AE 做的").

---

Operational flow (TL;DR)

The craft prose below is long; the pipeline itself is short — and everything deterministic is computed or compiled, never hand-written:

1. Decision gate (refuse bad clips) → pick ONE identity from CATALOG.md (17 identities; engine/compiler derived by lookup — never surface a mode/category question)
2. 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)
3. 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)
4. Visual QA: node scripts/preview-frames.cjs <project> → faithful composite previews in ~2s/frame (no render). Check § Visual QA before paying for a render.
5. render-and-composite.sh → gates (timing / occlusion+hero / overflow / hand-off) → final.mp4

Load-bearing rules people miss:

• 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.

---

Caption model — rail + embed

Every spoken phrase is one of three things:

	What	How it's shown
drop	filler — um/uh, stutters, self-corrections	not shown
rail	the default — ordinary spoken content (verbatim)	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 (17 entries: 12 classic + 5 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).

Procedure: probe the clip → shortlist 2–3 identities from the catalog → recommend ONE with a one-line why → the user picks → 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.)

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.
• Theme → read themes/README.md, author theme.json, run scripts/render-theme.sh (compiles + renders + plate reaction → final_fx.mp4).

---

Decision gate — RUN FIRST

Probe the video and classify the scene before either mode.

ffprobe <video.mp4>                    # specs
ffmpeg -ss <t> -i <video.mp4> -vframes 1 sample.png   # at 20/50/80%

Read the samples. Refuse if:

• 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)

1. Shot-cut probe. Sample frames at 20%, 50%, 80%. If a different subject/scene appears, trim the clip before the cut.
2. Letterbox / pillarbox probe. Black bars on the first frame? Compute safe content rect and constrain caption placement inside it.
3. 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.
4. 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.

---

Pipeline — 5 steps

1. hyperframes init <project> --non-interactive --video <video.mp4> --skip-skills
2. bash scripts/prepare.sh <project>       # matte ∥ transcribe (parallel) → safe-zones. One command.
                                           #   → frames_fg/ transcript.json safe-zones.json
3. [AGENT STEP — the only creative step] author a small JSON; see below by mode
   Cinematic: author plan.json → node scripts/fill-timings.cjs → fit-fonts.cjs → make-composition.cjs
   Theme:     author theme.json → bash scripts/render-theme.sh <project>   (compiles + renders + plate fx)
4. node scripts/preview-frames.cjs <project>   # ~2s/frame composite previews → § Visual QA (BEFORE the render)
5. bash scripts/render-and-composite.sh <project>  # gates → final.mp4 + history/ snapshot
   (Theme mode: SKIP steps 3b/5 — render-theme.sh already runs compile + render-and-composite
    + _postfx.sh; the deliverable is final_fx.mp4, final.mp4 is pre-plate-reaction)

Step 3 differs by mode:

Step 3 — Cinematic mode (pure embed)

1. 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.
2. 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.
3. 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.
4. 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.

1. Pick a theme DNA by content register (each themes/<name>.json has voice + when). State your pick + why; the user decides.
2. 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).
3. 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:

1. Washout — light text over a bright region (window/sign/sky): unreadable → move the plane or change DNA/mode (bright scene → ink).
2. Text-on-text — captions over the scene's own text/graphics, or two caption groups colliding.
3. Reading order — on-screen vertical order must match spoken order; the hero must not sit below later words.
4. Hero presence — the climax should be BIG and visibly behind the subject (~30–55% occluded), not a floating label in a margin.
5. 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 c

...