talking-head-recut

First, keep this skill fresh — confirm with the user before running: npx hyperframes skills update talking-head-recut. 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.

Talking Head Recut

Talking Head Recut takes a local video that plays in full and layers a sequence of timed, designed graphic cards onto it — titles, lower-thirds, data callouts, quotes, side panels, picture-in-picture — synced to what's being said. The agent designs the cards (timing + content) and writes each card's HTML directly in the conversation, then assembles a single composition HTML and renders it to MP4 via hyperframes. There is no fixed archetype list and no prescribed card structure — the overlays emerge from what the transcript actually says.

The front door is /hyperframes. This skill packages an existing talking-head clip with designed graphic cards (titles, lower-thirds, data callouts, quotes, side panels, PiP) — not plain captions (the spoken words as text). The clip plays untouched. Any other intent — plain subtitles, a standalone graphic, a from-scratch video — or any uncertainty → read /hyperframes first: the intent layer owns every route decision.

Graphic-packaging sibling of embedded-captions. Captions add the spoken words as a readable subtitle; this adds designed graphics on top of the playing video. Plain subtitles → embedded-captions. Build a video from scratch → the creation workflows (product-launch-video / faceless-explainer / …).

Routed through /hyperframes, the intent layer confirms only the input (which clip) and announces the render-strategy questions as deferred asks — aspect, layout, style group, and card count stay at Step 7, where the probed footage and transcript ground the recommendations; the layer's run-shape questions don't apply. A BRIEF.md, when present, carries the confirmed input and any user notes — read it first.

Inspectable intermediate files in the work directory:

• metadata.json — duration / width / height / fps
• audio.mp3 — extracted audio
• transcript.json — a flat word array [{ text, start, end }, …] (Whisper; no segments, no words wrapper)
• storyboard.json — lightweight card outline (the agent's plan)
• public/cards/card-XX.html — one HTML fragment per card
• public/index.html — final assembled composition
• output.mp4 — rendered video

CLI Resolution

# hyperframes — transcription (local Whisper) + rendering the assembled HTML to MP4
npx hyperframes --help

This skill runs entirely on the hyperframes CLI plus system ffmpeg / ffprobe. Transcription is local Whisper via hyperframes transcribe — no third-party service, API key, or rate-limited proxy.

Workflow

1. Check Environment

npx hyperframes doctor          # ffmpeg, headless browser, render deps
# confirm bundled assets:
ls "<SKILL_DIR>/assets/fonts"

Required:

• ffmpeg / ffprobe (system)
• <SKILL_DIR>/assets/fonts/*.woff2, <SKILL_DIR>/assets/vendor/gsap.min.js (bundled inside this skill, staged to work dir in Step 9)

Transcription needs no key — hyperframes transcribe runs Whisper locally (Step 4).

Strongly recommended on macOS for hyperframes render:

2. Create a Work Directory

All artifacts live under videos/<project-name>/ — the same convention as the other video workflows (product-launch-video / faceless-explainer / pr-to-video). Keep the cwd at the workspace root; everything below writes under this one subdirectory.

3. Extract Audio and Metadata

Outputs: metadata.json (read width/height/duration; fps = the r_frame_rate fraction evaluated, e.g. 30000/1001 → 29.97) + audio.mp3.

4. Transcribe

Local Whisper — no API key, no proxy, no rate limit. Writes a word-level transcript.json into the work dir (word text + start / end timestamps). Read it for the word / sentence timings that drive card timing in Step 6; group words into sentences yourself at punctuation / pauses if you need segment-level chunks.

Clamp to media duration. Whisper can return the final word's end a hair past the actual clip length — clamp every card endSec and composition.durationSeconds to the metadata.json duration, or the render will show a black tail past the video.

5. Correct Transcript

transcript.json is a flat array of word objects — [{ "text": "...", "start": s, "end": s }, …] (no segments array, no words wrapper; the per-word key is text). Read it and fix obvious ASR errors:

• Homophones, product names, technical terms, punctuation
• Edit a word's text in place; preserve its start / end timestamps
• There is no pre-grouped segments array — group words into sentences yourself (split at terminal punctuation / pauses) when you need segment-level chunks for card timing

6. Draft a Lightweight Storyboard (in chat)

No CLI involved. Read transcript.json + metadata.json and design cards directly. storyboard.json is an agent-internal planning artifact — no CLI command consumes it; it exists so you can think clearly about timing and content before writing each card's HTML. Keep the shape consistent with the example below so the same outline can drive the composition you author in Step 9:

Required Card fields:

Five zone values:

When you assemble the composition in Step 9, resolve each card's zone into pixel bounds on the card-host wrapper following the table above. Video bounds are set once at composition level (videoTrack.bounds); to make video appear to "move between cards", author GSAP tweens against #video-wrap in the composition's <script> (see Step 9).

No prescribed card roles, no prescribed narrative arc. Cards emerge from what the video actually says — could be all quotes or all data, could open with a number or with a story. Let the transcript drive the rhythm.

How many takeaways? — auto-infer from duration + density. No fixed upper limit. Pick a base pace from the video duration, then adjust by information density. Only floor is fixed: minimum 5 cards so even short videos have rhythm.

Step 1 — base pace by duration (the natural sec/card for medium density):

Step 2 — density multiplier (multiplies the base pace):

Step 3 — compute:

Examples (notice — no upper clamp; long videos naturally produce more cards):

• 30s reel, single punchline (low density) → 7 × 1.5 = 10.5s/card → round(30/10.5)=3 → floor to 5 cards
• 60s reflective monologue (low density) → 10 × 1.5 = 15s/card → 4 → floor to 5 cards
• 121s talking-head with rich data (high density) → 10 × 0.7 = 7s/card → 17 cards
• 5 min interview, mixed density → 16 × 1.0 = 16s/card → 19 cards
• 10 min deep-dive, high density → 16 × 0.7 = 11s/card → 55 cards
• 30 min lecture, medium density → 28 × 1.0 = 28s/card → 64 cards
• 1 hr podcast, low density → 45 × 1.5 = 67.5s/card → 53 cards

When a card holds longer than ~15s, plan for a richer card (data block, multi-step reveal, several sub-points unfolding with staggered animations) — a static one-liner gets boring past 8s. For long pieces where many cards exceed 30s, consider chunking the timeline into sub-compositions (one .html per chapter, mounted with data-composition-src) so the GSAP timeline per file stays manageable — see the timeline_track_too_dense HyperFrames lint warning.

content can be a plain string ("Title: annualized 5.69%\nNotes: ...") or any JSON shape that captures the data. The agent decides the shape per card.

Optional outro. This skill ships no fixed brand outro. If the user wants a closing card, design a neutral one yourself (wordmark + one-line tagline, ~1.5-2s, fade in -> short hold -> fade out), append it to cards[], and extend composition.durationSeconds to its endSec. Otherwise end on the last content card.

7. Decide Render Strategy

Confirm Visual Direction with User (DO THIS FIRST)

Before you start designing cards or deciding bounds, ask the user to pick the output ratio, the layout, the style, and the card-density preset. Frames are auto-selected from the chosen layout × style combination (see "Auto-pick frame" table below). Before sending the question, precompute two things:

1. recommendedRatio from the source video's aspect ratio (metadata.json width / height):

   • sourceAspect = width / height
   • sourceAspect ≥ 1.5 (≥ ~3:2 wide) → recommend 16:9
   • sourceAspect ≤ 0.7 (≤ ~9:13 tall) → recommend 9:16
   • 0.7 < sourceAspect < 1.5 (near-square) → recommend 4:5

   Mark the recommended option's label with " (recommended · matches source video X:Y)" so the user sees why it's recommended.

2. autoCount from Step 6 (max(5, round(videoSec / (basePace × densityMultiplier)))) so the "auto" option's label can show the concrete number.

Environment compatibility — pick the best available question channel. Not every runtime exposes the same structured-question tool. Apply this order:

1. Native clarification tool — use the structured 4-question call below.
2. Other native clarification tool (e.g. ask_question, request_user_input, IDE-specific prompt) — use that tool with the same 4 question texts and option lists. Preserve the recommendation markers and the precomputed values.
3. No native tool (Codex CLI, plain text-only runtimes) — ask directly in normal conversation. Use the plain-text template at the end of this section. Keep it to one message, 4 numbered questions (the global cap is 2–5 questions per round; we stay inside it).

Rules that apply to every channel:

• Ask at most 2–5 questions per round. Our 4 here fits.
• Even if missing info doesn't block rendering, ask once to confirm the parameters that materially affect the final output (ratio, layout, style, cardCount).
• If the user has already pre-approved defaults ("just use defaults", "no need to ask", "auto-pick everything"), asked you not to ask, or the run carries an ongoing autonomous signal ("surprise me" / "decide for me" — ../hyperframes/references/brief-contract.md § 1) — skip the question entirely and use: recommendedRatio, layout="stack" (safest cross-ratio default), style chosen from transcript tone in the most neutral group (editorial/data), autoCount. Tell the user what you picked in one sentence and continue.

Channel A — native AskUserQuestion:

About "Other" — AskUserQuestion automatically adds an "Other" option to the card count question. The user can type a number directly (e.g. "8", "20") as the cardCount target. Parse the input as an integer: if parsing succeeds → use that value (minimum 5 as a floor); if parsing fails → fall back to "auto".

Channel B — plain-text fallback (Codex CLI, runtimes without a native question tool). Post this as one normal message, then wait for the reply. Bullet-style 1/2/3/4 keeps the reply parseable:

Parsing the plain-text reply:

• Accept loose formats: "1A 2C 3B 4A", "A C B A", "16:9 / pip / data / auto", full sentences, or default.
• If any answer is ambiguous → re-ask only the ambiguous ones (still inside the 2–5 cap).
• If the user says "default / auto / use all recommendations" → skip without re-asking.

After the user answers (any channel):

1. Resolve the output canvas from the ratio answer — these are the exact storyboard.composition.width / height values to write:

   user choice	composition.width × height	storyboard.layout field
   16:9	1920 × 1080	"landscape"
   9:16	1080 × 1920	"portrait"
   4:5	1080 × 1350	"portrait" (schema treats 4:5 as portrait — height > width)

   For 4:5 bounds inside references/layouts/*.html — those files only document landscape (1920×1080) and portrait (1080×1920). For 4:5 (1080×1350) derive bounds by proportional scaling from portrait: keep horizontal values, scale vertical values by 1350/1920 ≈ 0.703. Example: overlay portrait card = { x: 24, y: 1280, w: 1032, h: 564 } → 4:5 card = { x: 24, y: round(1280 × 0.703), w: 1032, h: round(564 × 0.703) } = { x: 24, y: 900, w: 1032, h: 397 }.

2. Map the style group to a specific style by looking at the transcript tone — pick the one that best fits, but stay inside the user's chosen group. If you're unsure between two specific styles inside the group, send a second AskUserQuestion with those 2–4 specific style options.

3. Resolve final cardCount from the density answer:

   user choice	final cardCount
   Auto (recommended)	the autoCount you already computed
   Fewer	max(5, round(autoCount × 0.6))
   More	round(autoCount × 1.5) (no upper clamp)
   Other = "" (integer)	max(5, parseInt(n))
   Other = anything else	fall back to autoCount

4. Auto-pick the video frame from this table (frames don't ask the user — they follow from layout × style):

   layout	warm-paper styles (academic / whiteboard / editorial / xhs)	clinical styles (audit / swiss / terminal / minimal)	experimental styles (geom / spotlight)
   split	polaroid	hairline	clean
   stack	polaroid	hairline	clean
   pip	clean (pip pill already has chrome)	clean	clean
   overlay	clean (full-bleed forbids deco frames)	clean	clean

5. Tell the user what you chose in one sentence — ratio (+ canvas size), layout, specific style, frame, and final cardCount — then proceed with the rest of Step 7 (per-card layouts, motion patterns).

6. Record the five values (ratio / layout / style / frame / cardCount) in working memory (no schema field needed); you'll reference them while writing each card's HTML in Step 8 and while reading the matching references/<dim>/<key>.html for tokens and structure.

If the user picks an answer via "Other" with a free-text style name not in the 10-style library, treat it as a hint to design a fresh card visual yourself, but still anchor on the chosen layout's bounds.

Render Strategy Inputs

With ratio / layout / style / cardCount / frame locked from Step 7.0, the remaining per-card decisions are:

• Source-video fit inside the GSAP target: video element has object-fit: cover and is clipped to #video-wrap's tween bounds. If you want NO cropping (e.g. portrait source on landscape canvas shouldn't get its top/bottom chopped), aim the tween at a rect that matches the source's aspect ratio and let surrounding canvas show through (or fill with the card / a backdrop).
• card.zone per card: derive from your chosen composition layout (split → side-panel, stack → lower-third, pip → fullscreen, overlay → video-overlay), OR pick a different zone for one-off variants (fullscreen for hero / quote, whiteboard-area for dense data).
• accentIndex per card: each card pulls one of the 5 theme accent colors. Vary across cards for rhythm; reuse the same index when two cards belong to the same narrative beat.
• Motion vocabulary: pick 2–3 repeatable patterns from data-anim kinds (see the table later) and stick to them so the composition feels coherent.

Pick from these themeId palettes (use them as --accent-N / --bg / --text CSS variables in your composition <style> block):

Available fonts (woff2 in <SKILL_DIR>/assets/fonts/, staged to work dir in Step 9): Caveat (handwriting), LXGW WenKai TC (Chinese hand-script), Inter (modern sans), Virgil (geometric hand). Reference via @font-face or font-family directly.

For inspiration on visual patterns, <SKILL_DIR>/references/styles/ ships 10 self-contained reference cards (academic / editorial / minimal / spotlight / geom / whiteboard / audit / terminal / swiss / xhs) that you can copy as starting points — but do not feel constrained to match any of these. Each card is your own design.

Visual Design Library (<SKILL_DIR>/references/)

Beyond the composition-level themeId, the skill ships a richer reference library at <SKILL_DIR>/references/ covering three orthogonal visual dimensions you can freely mix:

Read <SKILL_DIR>/references/DESIGN_INDEX.md for the full matrix and a loose decision guide (interview / product launch / data analysis / social clip / technical tutorial / emotional story …). When you decide to use a specific style / layout / frame, Read the corresponding file:

• references/styles/<key>.html — self-contained card fragment with that style's CSS tokens (colors, fonts, padding, ornament) and a placeholder takeaway. Copy the .card[data-card-id="ref-<key>"] style block, rename the data-card-id to your card's id, swap the placeholder content for the real takeaway, and you're done.
• references/layouts/<key>.html — exact videoBounds + cardBounds for both landscape and portrait, with a copy-paste JSON snippet for storyboard.json's per-card layout field.
• references/frames/<key>.html — decorative HTML to add as a sibling of #video-wrap, plus placement instructions for the composition CSS.

Pick style × layout × frame per card — you can change all three between cards as long as the transitions read smoothly. A common rhythm: open editorial × overlay × clean, switch to audit × split × hairline for the data card, close on whiteboard × pip × polaroid.

The 10 styles are skill-side design tokens, not composition-level themes — they don't need to be declared in storyboard.composition; they live inside each card's HTML. The themeId field can still pick a composition-level palette (table above) that controls page-body background and video border chrome.

Layout Compositions (Card + Video)

Two coordinated decisions per card define how it shares the canvas with the source video:

• card.zone (declared in storyboard.json) — one of the 5 schema values; resolve it into pixel bounds (per the table in Step 6) when you write the card-host wrapper's inline style in Step 9.
• #video-wrap bounds at this card's time window (declared imperatively in the composition's GSAP timeline) — the agent tweens #video-wrap to a target rect for each layout transition.

Schema does NOT store per-card video bounds. videoTrack.bounds is one-time at composition level (defaults to full canvas). Video "moving" between cards is purely a GSAP animation authored in index.html. There is no card.layout field — earlier versions of this doc invented one; the real schema only has card.zone.

4 composition layouts (from references/layouts/) — each is a recipe pairing a zone with a #video-wrap tween target:

For 4:5 (1080×1350), scale portrait y/h values by 1350/1920 ≈ 0.703 (see Step 7.0 Channel A / Channel B recommendedRatio resolution table).

Other zone values for one-off variants (still uses card.zone; no fake "layout" field):

You can mix recipes per card — choose card.zone based on what suits the moment, then write the GSAP tween for #video-wrap between cards.

Storyboard Render Contract

storyboard.json is an agent-internal planning artifact — no CLI command parses it. It exists to keep your timing and content decisions explicit before you write each card's HTML. Stick to the v3-style shape below so the same outline drives the composition you assemble in Step 9.

Required structure (see Step 6 for the full example):

• schemaVersion: 3
• composition: { fps, width, height, durationSeconds, layout, themeId, seed } — note durationSeconds/fps/themeId/layout live inside composition, NOT at top level
• videoTrack: { sourcePath, startSec, endSec, bounds? } — video bounds default to full canvas
• subtitles: { enabled, ... }
• cards[] — each card has the 6 required fields: id, intent, startSec, endSec, accentIndex, zone, contentHints

Rules:

• Card times stay inside composition.durationSeconds and should not overlap unless intentional (use data-track-index to control z-order when they do).
• Visual details live in card HTML fragments (Step 8), NOT in contentHints. contentHints is your own structured prompt for designing the card; the rendered look is the HTML.
• Keep the storyboard shape stable — even though nothing parses it, you read it back while authoring Step 8/9, and consistency keeps card IDs and timing in sync.
• Agent-side decisions like "I picked overlay × geom × clean" do NOT belong in storyboard.json — keep them in working memory and use them when authoring card HTML + GSAP tweens.

Transparent card backgrounds for cards that share canvas with video. When the GSAP tween leaves video visible behind/beside the card (overlay recipe, pip recipe, or any card.zone = 'lower-third' | 'video-overlay' moment), the card's .root MUST NOT paint a full opaque background — otherwise it occludes the video. Two patterns:

For side-panel-zone cards (split recipe), the card-host is already only half the canvas, so an opaque card bg is fine — it only covers its half.

8. Write Each Card's HTML

Create $WORK_DIR/public/cards/{card-id}.html for each card. Each file contains a single rooted HTML fragment that follows this contract:

Card HTML Contract

Hard rules (hyperframes lint will reject violations):

• Single root <div class="card" data-card-id="{cardId}">
• Inline <style> rules MUST be prefixed with the scope selector above
• No <script> tags
• No external URLs in src= / href= (no CDN, no remote fonts)
• No inline event handlers (onclick= etc.)
• All assets via relative paths into the same public/ directory
• Colors via var(--accent-N) etc. for portability across themes

Animations are declared, not coded. Use data-anim-* attributes only; never write <script> to animate. You compile every data-anim-* declaration into the single master GSAP timeline in Step 9.

Card Sizing — Mobile-First in Portrait

The 10 references/styles/*.html are sized for a 1920×1080 landscape preview. When storyboard.layout = "portrait" (1080×1920, the dominant case for social / mobile), scale every visual size up — phones hold the screen close, and the same pixel count reads smaller than on a landscape TV-style canvas.

Rule of thumb: portraitPx = round(landscapePx × 1.3), then floor to a nearby 4px multiple for visual rhythm. Hero headlines may go up to ×1.4; small meta text stays at ×1.2 to avoid crowding.

Padding shrinks slightly in portrait — the card is narrower so big landscape padding (40–64px) eats too much width. Use 24–36px horizontal padding in portrait.

If you're producing a single card that must work in both layouts, prefer a @container query on the card root over hard-coding sizes:

But for most cards, a single layout choice is fine — just pick the size table column that matches the storyboard's layout field.

Available data-anim Kinds

This list is closed, and deliberately so: a card is an HTML fragment whose motion this skill compiles into the shared overlay timeline in Step 9 (see the GSAP mapping table there). That is why this workflow does not search the HyperFrames component registry the way the composition workflows do — npx hyperframes catalog returns standalone compositions that carry their own timeline, and a card has no place to mount one. Reach a look the kinds below cannot express with plain CSS inside the card's scoped <style>.

data-anim-at is seconds relative to the card's startSec — when you compile each declaration into the GSAP timeline in Step 9, add the card's startSec to get the absolute time and quantize to 1/fps.

9. Assemble the Composition HTML

Stage the assets and write $WORK_DIR/public/index.html:

Composition Template

GSAP Statement Cheat Sheet

Compile each data-anim attribute into a GSAP statement. Times are absolute seconds = card.startSec + data-anim-at, quantized to 1/fps. Selector is .card[data-card-id="X"] #elementId.

Quantize: T = Math.round(absSec * fps) / fps. At 30fps the smallest step is 1/30 ≈ 0.0333s; rounding to 4 decimals (.toFixed(4)) is fine inside the JS literal.

Video Framing Reference (per layout value)

The selector for the video container is #video-wrap. Animate its bounds between cards using tl.to('#video-wrap', { ...bounds }, T). Initial bounds should be set inline on the element to match card-01's layout. Pick a transition duration of 0.5–0.7s with ease: 'power2.inOut'.

Decorative frames (clean / hairline / polaroid) sit as a sibling of #video-wrap and follow it through layout transitions. See references/frames/ for each frame's placement HTML, suggested CSS, and which layouts it pairs with. Quick rule: overlay layout suppresses decorative frames (the full-bleed video clashes with chrome); PiP layouts already have their own pill treatment (border-radius + white ring + shadow), so add a decorative frame only on top of split / stack.

GSAP target lookup table for #video-wrap per composition layout (landscape 1920×1080 — for portrait & 4:5 see references/layouts/*.html which list all three ratios):

To toggle the pip-pill chrome (border-radius + white ring + drop shadow) when entering or leaving a pip moment:

Card-host bounds match the zone. Resolve the card's zone into pixel bounds using the table at the top of Step 6, then write those into the card-host's inline style="left:Xpx;top:Ypx;width:Wpx; height:Hpx;...". For video-overlay zone (overlay recipe), the card-host fills the full canvas — your CSS inside .card .root decides where the actual visible card sits.

HyperFrames Layout / Animation QA Rules

• Build each card's static hero frame first: the moment where the card is fully visible and readable.
• Confirm video, cards, subtitles/captions, and diagrams do not unintentionally overlap.
• Confirm hidden video areas are clipped by the frame and not visible outside intended bounds.
• Register one paused master timeline as window.__timelines["talking-head-recut"].
• Build timelines synchronously at page load; no async, setTimeout, Promises, or media play() calls.
• Do not use Math.random() or Date.now() in render paths.
• Do not use repeat: -1; calculate finite repeats from the video duration.
• Prefer GSAP transforms and opacity (x, y, scale, rotation, opacity) over layout properties (top, left, width, height) for motion.
• Animate wrappers such as #video-wrap, not the video element dimensions directly.
• Avoid animating the same property on the same element from multiple timelines at the same time.
• Use data-track-index, not data-layer; use data-duration, not data-end.
• Every timed element (card-host, sub-composition, etc.) should include class="clip" alongside its own classes — e.g. class="card-host clip". Visibility itself is driven by data-start / data-duration: the runtime gates every [data-start] element to its window whether or not this class is present. .clip is the marker Studio and the GSAP clip-ownership rules read to recognise a clip, so leaving it off makes the element harder to edit and to lint (lint: timed_element_missing_clip_class, a warning).
• For body / global font-family, list concrete font names ('Inter', 'Caveat', …) — not a CSS variable like var(--font-family). The HyperFrames font resolver doesn't expand CSS vars during static analysis (lint: font_family_without_font_face). Cards may still use var(--font-family) internally since their @font-face declarations are loaded.

10. Render to MP4

hyperframes render <dir> reads <dir>/index.html and produces the MP4. The canonical composition keeps the visual <video> muted and mounts the same source as the root #source-audio track, so the rendered MP4 preserves the talking-head audio without a manual remux. This uses a separate audio track rather than data-has-audio="true" so its volume and ducking remain independently controllable on the timeline. The flag PRODUCER_BROWSER_GPU_MODE=hardware (or --browser-gpu) is strongly recommended on macOS — software-only Chrome rendering times out on most laptops.

For a sanity check before the full render, capture a single frame at a specific timestamp:

11. Report Results

Tell the user:

• Work directory path
• storyboard.json (the card outline you designed)
• public/cards/*.html (one HTML per card)
• public/index.html (the assembled composition)
• output.mp4 (the final video)
• ASR provider used
• Card count + how you chose them (in 1 sentence)
• Any missing keys or quality caveats

Optional live preview (on request only). The clip plays unchanged inside public/index.html with the overlays on top, so it previews faithfully. Don't open it during the run. When the user asks, start a long-lived server after render and report the URL:

Do not delete the work directory unless the user asks.