hyperframes

HyperFrames — read this first

Start here for any HyperFrames task — especially with no project agent config (CLAUDE.md / AGENTS.md / .cursorrules) present. Capability map + video router below.

Capability map — which skill for which intent

You want to…	Go to
Make a video (from a URL, brief, topic, GitHub PR, existing footage, or a single element to animate)	the video router below (§ Video routing)
Author a slideshow / presentation / pitch deck — discrete slides, fragments, branching, hotspots	/slideshow
Author / edit an HTML composition — the data-* contract, clips, tracks, sub-compositions, variables	/hyperframes-core
Animate — atomic motion rules, scene blueprints, transitions, runtime adapters (GSAP / Lottie / Three.js / Anime.js / CSS / WAAPI / TypeGPU)	/hyperframes-animation
Creative direction — design.md, palettes, typography, narration, beat planning, audio-reactive	/hyperframes-creative
Media preprocessing — TTS voiceover, background music, transcription, background removal, captions	/hyperframes-media
CLI dev loop — init, lint, validate, inspect, preview, render, publish, doctor	/hyperframes-cli
Install registry blocks / components (hyperframes add)	/hyperframes-registry

The composition authoring contract (every timed element needs data-start / data-duration / data-track-index; timed elements need class="clip"; GSAP timelines are paused and registered on window.__timelines; deterministic logic only — no Date.now() / Math.random() / network) is not duplicated here — it lives in /hyperframes-core. Read that before writing composition HTML.

What HyperFrames cannot do — check this first

HyperFrames authors an HTML composition and renders it to MP4 from code. That model has hard outer edges. A request past one of them is not a routing choice — it is out of scope, so decline (or point at the right tool) instead of reaching for a workflow. These follow from the architecture, not from any single request:

• The render is deterministic and self-contained. Every value, asset, and piece of text is baked in when you author; the render does no network call and no live / at-render-time data pull (core rule: no Date.now() / Math.random() / network). "Refresh the numbers live at render time" is out — fetch the data once at author time and bake it in, or decline.
• Existing video is overlaid, never edited. HyperFrames composes frames on top of a source clip — /embedded-captions adds a caption layer, /graphic-overlays adds designed graphic cards (lower-thirds, data callouts, titles) — and the clip plays unchanged underneath, but neither post-processes the encoded video stream. Changing the footage itself (its timing, color, framing, order, or audio) is NLE-style editing and out of scope.
• Remotion import is one-way. /remotion-to-hyperframes translates the Remotion framework's source into HyperFrames. There is no reverse (HyperFrames → Remotion, or → any other framework — out of scope), and a non-Remotion React / web-animation source has no Remotion source to translate — re-create it via /general-video.
• It cannot produce inputs it does not have. No screen / session recording, no camera capture, no AI talking-head / lip-synced avatar generation. If the footage or asset does not exist yet, HyperFrames cannot conjure it — ask the user to supply it (or use the right capture tool) first.

Everything else — a video from a URL, brief, topic, PR, footage-to-annotate, or a single element to animate — is in scope; route it below.

---

Video routing

This section knows ONLY top-level workflows. It does not load workflow-internal phases, domain skills (hyperframes-* — see the capability map above), or technical references.

Decision table

INPUT type (intent) is the primary axis; OUTPUT length is only a ceiling, not a gate. For a matching input, the specialized workflows handle anything up to ~3 min — which workflow you enter is decided by intent (the input type, and for text the subject), not by length. Length matters only at the top end: a genuinely longer piece (a 3-5 min tutorial, a 5 min+ deep dive) is a different register and routes to /general-video. Within the ≤~3 min band, a third axis splits the two text-fed workflows — the subject: a product being marketed vs a topic being explained (see the disambiguation rule in step 3 below).

Length / Input	Product launch (URL / brief / script)	General website / URL	GitHub PR / code change	Topic / article / notes (no product, no URL)	Existing video (talking-head) †
≤ ~3 min	/product-launch-video	/website-to-video	/pr-to-video	/faceless-explainer	/embedded-captions · /graphic-overlays †
3-5min tutorial	/general-video	/general-video	/general-video	/general-video	/embedded-captions · /graphic-overlays †
5min+ deep dive	/general-video	/general-video	/general-video	/general-video	/embedded-captions · /graphic-overlays †
Static / loop	/general-video	/general-video	/general-video	/general-video	/general-video

Coverage today: the ≤ ~3 min band has dedicated workflows for product-launch / general-website / GitHub-PR / topic inputs (a URL splits by kind then intent — see step 3), and the existing video column is covered at any length — by /embedded-captions (captions / subtitles) or /graphic-overlays (designed graphic overlays), split by intent (see step 2). Every other cell is /general-video — the general HTML-composition authoring flow (input- and length-agnostic): everything longer than ~3 min (the 3-5 min / 5 min+ rows) and every static / loop format. The router never dead-ends on a creatable video; the only true "general / none" answer is a request outside HyperFrames itself (e.g. NLE-style editing of a finished video file — re-timing, recoloring, reframing, reordering, audio).

† Existing footage splits by intent, not length. Captions / subtitles → /embedded-captions; designed graphic overlays / packaging — lower-thirds, data callouts, titled cards, pull-quotes — → /graphic-overlays. Both overlay the clip, which plays unchanged underneath; neither edits the footage itself.

Genre short-circuit (precedes the table). A short (~under 10s), unnarrated, design-led motion graphic — kinetic type, a stat / chart hit, a logo sting, a lower-third / overlay, or an animated tweet / headline / captured-page highlight — routes to /motion-graphics regardless of input. It is an OUTPUT genre, not an input type, so it takes precedence over the input-type table above when the ask is clearly motion-first with no narration (see step 2). A longer or narrated treatment routes by input type, or /general-video.

Migrating an existing composition (special case)

The table above is for creating a video from an input. One workflow sits outside it: if the user explicitly asks to port / convert / migrate an existing Remotion (React) composition into HyperFrames → /remotion-to-hyperframes. This is source translation, not creation-from-input, so it has no INPUT × LENGTH cell. Route here ONLY on explicit migration language ("port my Remotion project", "convert this Remotion comp", "rewrite this as HyperFrames") — a passing mention of Remotion is not a trigger; default to the creation table or /hyperframes-core.

Routing procedure

1. Determine INPUT type + target length. Routing needs to know what the video is about — its subject and input. If the subject itself is unspecified (e.g. "make a video about our thing" with no URL, named product, topic, or asset to work from), or the input type is unknown, ask before entering any workflow — clarify first; do not invoke a workflow Skill and then ask, since committing to a workflow is itself the routing decision. Ask at most 2 clarifying questions:
   • "What's your input — a product (URL or brief), a general website / URL, a GitHub PR / code change, a topic or article to explain, or an existing talking-head video to caption?"
   • "Target length — about 3 minutes or under, or longer (a 3-5 min tutorial / 5 min+ deep dive)?"
   • Spec defaults — state, don't ask (these do NOT affect the route, so never block routing on them): aspect 16:9 (the engine also supports 9:16 vertical — switch only if the user names a vertical destination like TikTok / Reels / Shorts); narration / caption language = the user's. Confirm only if the user pushes back; the chosen workflow re-confirms its own specifics at its Step 0.
2. Pick by INPUT type (intent) first; length is only a ceiling, not a gate.
   • Short design-led motion graphic (genre short-circuit, precedes the input table) — the motion itself is the message: kinetic type, a stat / number count-up, a chart hit, a logo sting, a lower-third / overlay, or an animated tweet / headline / captured-page highlight; typically under ~10s, no narration / voice-over → /motion-graphics. This is an OUTPUT genre, not an input type — when the ask is clearly a quick, unnarrated, design-led motion piece, route here regardless of input. A longer or narrated treatment of the same material is NOT this (route by input type, or /general-video).
   • Existing talking-head video (the user has a clip and wants something added over it) → split by intent, at any length (input type wins over length): captions / subtitles (the spoken words as readable text) → /embedded-captions; designed graphic overlays — lower-thirds, data callouts, titled info-cards, pull-quotes, a graphics-packaged edit → /graphic-overlays. (Editing the footage itself — re-timing, recolor, reframe, reorder, audio — is NLE-style and out of scope; see § What HyperFrames cannot do.)
   • GitHub PR / code change (a github.com/<owner>/<repo>/pull/<N> link, an owner/repo#N ref, or "this PR") → /pr-to-video (up to ~3 min).
   • Otherwise (product URL / brief / topic text): intent picks the workflow via step 3, and it handles anything up to ~3 min — a short 15-30 s promo and a ~100 s explainer both route by intent, not by length. Route to /general-video (the length-agnostic fallback — see step 4) only when the target is clearly longer than ~3 min (a 3-5 min tutorial, a 5 min+ deep dive). Never force a genuinely long piece into a ≤~3 min workflow — intent decides within the band, /general-video covers the rest.
3. Disambiguate the ≤~3 min URL / text inputs (the intent split). Two splits:
   • URL kind + intent — a URL no longer auto-wins for PLV; its kind then intent decides: a GitHub PR link (.../pull/<N>, owner/repo#N, "this PR") → /pr-to-video; otherwise a website URL splits by intent — marketing / launching / promoting a specific product or SaaS → /product-launch-video; a general site → video (site tour, portfolio / blog / landing-page showcase, a social clip from the site's own visuals, or just "turn this site into a video") → /website-to-video. Both crawl with headless Chrome; PR URLs are read via gh. When it's genuinely unclear whether a site URL is a product launch or a general-site video, ask one question.
   • Product vs topic (text, no URL) — the decisive question is what the video is about, not the input format:
     • A specific product / company / SaaS / app / website being marketed, launched, or promoted → /product-launch-video.
     • A concept / topic / article / how-something-works being explained, with no product and no URL → /faceless-explainer.
     • Tie-breakers: "Promote / launch / sell / our product" wording → PLV. "Explain / teach / how X works / what is X" with no product → faceless. The shipped style for faceless is always pin-and-paper.
     • A named site without a pasted URL is still PLV. A script that mentions a product or its website ("our site is acme.io", "promote ") routes to PLV even with no clickable link — PLV can web-search the site and crawl it for brand assets (unless the user opts out, → no-capture preset mode). Not pasting the URL does not make it a faceless / no-capture job. The verbatim-vs-restructure choice for a supplied script is internal to PLV and never changes the route.
     • Conflicting cues → ask, don't guess. If the supplied source is a product's own marketing (its landing page, a promo blog about their platform) yet the user explicitly asks to strip the promotion — a neutral explainer of the underlying concept, not an ad — treat it as genuinely ambiguous (is the video about their product, or the general concept?) and ask one question, rather than resolving to faceless on the "neutral" cue alone. Contrast: a general topic where a product is merely an aside the user says to exclude ("explain how OAuth works — we sell an auth product but don't mention it") is unambiguously faceless — no need to ask.
   • Still unclear after reading the request → ask exactly one question: "Is this promoting a specific product, making a video from a general website, explaining a topic/concept, walking through a GitHub PR, or adding captions to an existing video?"
4. Fall back to /general-video. When no specialized workflow above matches, route to /general-video — the general HTML-composition authoring flow (the original hyperframes flow: design system → plan → layout-before-animation → build → validate), which is input- and length-agnostic. Do not fake-route into a specialized workflow (don't force a tutorial into PLV); /general-video is the correct general home, not a near-fit. The only genuine "no workflow / general" answer is a request outside HyperFrames itself — e.g. NLE-style cutting/editing of a finished video file (captioning a talking-head clip is /embedded-captions; overlaying designed graphics on it is /graphic-overlays).

Workflow descriptions (for disambiguation)

/product-launch-video

• Input: A product being marketed, supplied as one of: (a) a product URL → crawled with headless Chrome for assets, brand tokens, page structure; (b) a script / brief that names a product site (even without a pasted link) → PLV resolves the site by web search and crawls it for brand tokens + assets, unless the user opts out of searching; (c) a script / brief with no derivable site (or an explicit "don't scrape") → no-capture mode, you pick a style preset that supplies the palette + design system (text/typography scenes, no scraped assets). A supplied script can be used verbatim as the voice-over or restructured into punchier per-scene narration — PLV asks which.
• Output: product launch / SaaS explainer / promo video as a HyperFrames composition rendered to MP4 — up to ~3 min (sweet spot ~30-90s; longer still when a verbatim script runs long — verbatim length follows the script)
• Triggers: "make me a launch video for X", "promo for our website", "explain my SaaS in a minute", "feature reveal for X.com", "marketing video for our product", "I have a script — turn it into a 60s promo", "here's my launch script for , our site is ", "use my script word-for-word as the voiceover", "make a text-only launch video, no website / don't scrape anything"
• Do NOT use for: pure-text explainers about a topic / concept with no product (→ /faceless-explainer) — note a script that names a product or its site is PLV, not faceless, even when no URL is pasted; a general (non-launch) website → video — a site tour / showcase / social clip not centered on marketing a product (→ /website-to-video); a GitHub PR / code-change explainer (→ /pr-to-video); adding captions to an existing video (→ /embedded-captions); anything clearly over ~3 min (tutorials, deep dives → /general-video); customer interviews, motion graphics without a product context, static brand assets (a short product promo, even 15-30 s, is still PLV — length is not the gate, the product intent is)

/website-to-video

• Input: A general website / URL the user wants turned into a video — when the goal is a video of / from the site itself, not a product launch. Captured with headless Chrome for real screenshots + brand assets. (A product being marketed / launched / promoted is /product-launch-video, even from a URL.)
• Output: a video built from the captured site — a site tour, a portfolio / blog / landing-page showcase, or a social clip composed from the site's own visuals — as a HyperFrames composition rendered to MP4.
• Triggers: "turn this website into a video", "capture this site and make a video", "make a video from my site", "site tour / showcase from ", "a social clip from our homepage", "I just have a URL — make something"
• Do NOT use for: a product being marketed / launched / promoted — a launch / promo / feature-reveal / "sell our product" framing (→ /product-launch-video, even from a URL); a topic / concept explainer with no site (→ /faceless-explainer); a GitHub PR (→ /pr-to-video); adding captions to an existing video file (→ /embedded-captions); a short unnarrated motion graphic that just highlights / animates a captured page (→ /motion-graphics — a single quick page-highlight shot, not a narrated / multi-scene site video). When it's genuinely unclear whether a site URL is a product launch or a general-site video, ask one question.

/faceless-explainer

• Input: Arbitrary text — a topic line, an article, notes, or a brief — being explained, with no product being marketed and no site to capture. (If the text names a product or its site, that is /product-launch-video, which can resolve + crawl the site — even when no URL is pasted.) Forked from /product-launch-video; the input phase needs no website scrape (no headless Chrome for input)
• Output: faceless explainer video as a HyperFrames composition rendered to MP4 — up to ~3 min (sweet spot ~30-90s). Every visual is LLM-invented per scene (typography / abstract graphics / diagram / data-viz); ships the pin-and-paper style preset
• Triggers: "make a faceless explainer about X", "explain how DNS works as a video", "turn this article into an explainer video", "video explaining [concept], no product", "topic → short educational video", "explainer from my notes"
• Do NOT use for: anything centered on a specific product / company being marketed, or a script that names a product site even without a pasted URL (→ /product-launch-video, which web-searches + crawls it); a request that supplies a URL — a product site (→ /product-launch-video), a general website to turn into a video (→ /website-to-video), or a GitHub PR (→ /pr-to-video); adding captions to an existing video (→ /embedded-captions); anything clearly over ~3 min (tutorials, deep dives → /general-video); product ad / promo formats (→ /product-launch-video); a pre-recorded / user-supplied voiceover or other media to time visuals to — faceless invents every visual and generates its own narration (TTS), it does not sync to supplied audio (→ /general-video); videos that need real screenshots or scraped brand assets (a short explainer, even under 30 s, is still faceless — length is not the gate, the explain-a-topic intent is)

/embedded-captions

• Input: An existing talking-head / single-subject video (MP4) the user wants captioned — actual footage, not a URL or a text brief. Transcribed locally (Whisper, no API key) and matted (RVM) so the subject can occlude captions; no website scrape, no headless Chrome.
• Output: the same footage, untouched, with a caption layer added — Standard (default): a verbatim lower-third rail carrying the transcript plus an embedded climax composited behind the subject at the peak; or Cinematic: pure embed, every caption composited into the scene behind the subject. Any length — short reel to long explainer.
• Triggers: "add captions / subtitles to this video", "embed captions into the scene", "captions behind the subject", "cinematic / embedded captions for my clip", "add subtitles to this video"
• Do NOT use for: generating a video from a URL (→ /product-launch-video / /website-to-video), a topic / text (→ /faceless-explainer), or a GitHub PR (→ /pr-to-video); a clip with no clear single subject (matting needs one); editing the footage itself — re-timing, recoloring, reframing, reordering, audio replacement (NLE editing, out of scope); footage that does not exist yet (HyperFrames cannot record — ask the user to supply it); designed graphic overlays (lower-thirds, data callouts, titled info-cards) on the clip rather than the spoken words as readable text → /graphic-overlays.

/graphic-overlays

• Input: An existing talking-head / interview / podcast video (MP4) the user wants packaged with designed on-screen graphics — actual footage, not a URL or brief. Transcribed locally (Whisper). The clip plays in full underneath; nothing is cut, re-timed, or recolored.
• Output: the same footage with a sequence of timed graphic-overlay cards composited on top / beside it — kinetic titles, lower-thirds, data callouts, pull-quotes, side panels, picture-in-picture — synced to the transcript, via a design system of 10 styles × 4 layouts × 3 frames. Any length. (This is the replacement for the removed /footage-recut info-card overlay flow.)
• Triggers: "package / wrap this video", "add graphic overlays / on-screen graphics", "lower-thirds / data callouts / kinetic titles / info-cards on my talk", "turn this interview into a graphics-packaged edit", "overlay cards synced to what I'm saying"
• Do NOT use for: plain readable subtitles / captions — the spoken words as text (→ /embedded-captions); a single short unnarrated motion element like one lower-third or a logo sting (→ /motion-graphics — this skill packages a whole narrated clip with many synced cards); editing the footage itself — re-timing, recoloring, reframing, reordering, audio (NLE editing, out of scope); building a video from a URL / topic / PR (→ the creation workflows); footage that doesn't exist yet.

/pr-to-video

• Input: A GitHub pull request — a code change, given as a PR URL (github.com/<owner>/<repo>/pull/<N>), an owner/repo#N ref, or "this PR" in a checked-out repo. A URL, but a PR link read via the gh CLI — NOT a marketing site to scrape.
• Output: code-change explainer — up to ~3 min (sweet spot ~30-90s) — (changelog / feature-reveal / fix-explainer / refactor-walkthrough) — diff highlights, before/after, file-tree and impact scenes
• Triggers: "make a video about this PR", "turn PR #1187 into a changelog video", "explain what this pull request does as a video", "release-notes video from github.com/org/repo/pull/123", "turn this PR into a video"
• Do NOT use for: a product / marketing website URL (→ /product-launch-video) or a general website to turn into a video (→ /website-to-video); a topic / article / text with no PR (→ /faceless-explainer); adding captions to an existing video (→ /embedded-captions); a whole-repo tour or multi-PR release (no workflow yet → /general-video)

/remotion-to-hyperframes

• Input: An existing Remotion (React) video composition's source — the user explicitly asks to port / convert / migrate / rewrite it as HyperFrames. Direction is one-way (Remotion → HyperFrames) and specific to the Remotion framework; this is NOT a creation-from-input workflow.
• Output: A HyperFrames HTML composition translated from the Remotion source, graded against the Remotion render with an SSIM eval harness + tiered test corpus
• Triggers: "port my Remotion project to HyperFrames", "convert this Remotion comp", "migrate from Remotion", "rewrite this as HyperFrames HTML"
• Do NOT use for: authoring a NEW composition (even while A/B-testing a Remotion video), a passing mention of Remotion, or "the same video as my Remotion one" without an explicit migrate request (→ creation workflows / /hyperframes-core); the reverse direction — exporting HyperFrames back out to Remotion or any other framework (out of scope, see § What HyperFrames cannot do); a non-Remotion React / web-animation source (no Remotion source to translate → re-create it via /general-video)

/motion-graphics

• Input: A request for a short, design-led MOTION GRAPHIC where the motion itself is the message — typically under ~10s (up to ~30s), no narration / voice-over. Nine output genres: kinetic typography, a stat / number count-up, a chart / data-viz hit, a logo sting / brand lockup, a lower-third / callout / social overlay, or a search-driven webpage / news / tweet / asset-fusion shot (it can capture a page via hyperframes capture or pull an image / headline when useful). An OUTPUT-genre short-circuit — it spans inputs, so it precedes the input-type table when the ask is clearly motion-first and unnarrated.
• Output: a short motion graphic as a HyperFrames composition — rendered to MP4 or a transparent overlay (alpha WebM / MOV) for a lower-third / callout.
• Triggers: "an 8s logo sting", "animate this stat / number", "a kinetic-type intro", "a quick stat / chart hit", "turn this headline or tweet into a motion graphic", "a motion poster", "a transparent lower-third / callout overlay"
• Do NOT use for: a longer, multi-scene, or narrated piece, a brand reel, or any custom composition past ~10-15s (→ /general-video); a narrated video OF a website / site tour (→ /website-to-video — motion-graphics' webpage genre is a single quick page-highlight shot, not a narrated site video); a narrated topic explainer (→ /faceless-explainer); a product launch / promo (→ /product-launch-video); a GitHub PR (→ /pr-to-video); adding captions to existing footage (→ /embedded-captions)

/general-video

• Input: Anything not handled above — a creative brief, a single element to animate, an edit to a composition you're building. Input- and length-agnostic.
• Output: A HyperFrames HTML composition (any length / format) authored with the original hyperframes flow: design system → prompt expansion → plan → layout-before-animation → build (delegating to the hyperframes-* domain skills) → validate.
• Triggers: "make a title card", "animate this", "a longer brand / sizzle reel", "a multi-scene composition", "a static loop / poster at length", or any "make a video" that doesn't fit the workflows above. (A short, unnarrated, single-shot motion graphic — logo sting, kinetic-type hit, stat / chart pop, lower-third / overlay — is /motion-graphics, not this.)
• Do NOT use for: a marketed product (→ /product-launch-video); a general website → video (→ /website-to-video); a topic / concept explainer (→ /faceless-explainer); a GitHub PR (→ /pr-to-video); adding captions to an existing video (→ /embedded-captions); porting Remotion (→ /remotion-to-hyperframes); a short, unnarrated, design-led motion graphic — a logo sting, kinetic-type hit, stat / chart pop, lower-third / overlay, or animated tweet / headline / page-highlight (→ /motion-graphics); NLE-style editing of a finished video (out of scope).

Out of scope for video routing

• Domain skills (/hyperframes-core, /hyperframes-animation, /hyperframes-cli, /hyperframes-creative, /hyperframes-media, /hyperframes-registry) — these are NOT routed here, but they ARE in the capability map at the top of this skill; a workflow's build phase loads them as technical references.
• Workflow-internal phases — phases live inside each workflow's folder and are dispatched by that workflow's orchestrator, not by this router.