Video
HyperFrames renders HTML compositions to MP4. The upstream skills own the craft
(composition contract, animation, CLI, media). This skill owns only the
personal layer: setup, routing from daily work, media providers, where files
land, and delivery. Never restate upstream rules from memory - load the
upstream skill that owns them.
1. Setup (once per session)
node -v # >= 22
npx -y hyperframes@latest skills check || npx -y hyperframes@latest skills update
skills update installs the core set (hyperframes, hyperframes-*,
media-use) into ~/.claude/skills and ~/.agents/skills; creation workflows
install on demand at routing time. A failed install is a visible failure -
surface it, never continue from remembered upstream content. On a new machine
run npx -y hyperframes@latest doctor once (FFmpeg required; headless Chrome self-installs).
Never name a local skill hyperframes-* - the upstream installer claims that prefix.
2. Route
Pre-select from the table, then read the hyperframes skill and follow its state table
(resume, edit, and render-only rows still win). The route file
(references/routes/<workflow>.md) still has the final say.
| Ask |
Workflow |
Defaults |
pr - "video for this PR", after vd:ship |
pr-to-video |
16:9, developers. Angle from the PR title's conventional-commit type: feat → feature-reveal, fix → fix-explainer, refactor/perf → refactor-walkthrough, else changelog |
release - a tag or "this release" |
faceless-explainer, input = gh release view <tag> --json body -q .body |
square, 45-60s. One scene per theme, never one per bullet - visualize, don't list |
promo - product, docs site, or vd:show-off page URL |
product-launch-video |
16:9 master |
explain - concept, lesson, devlog idea |
faceless-explainer |
square |
motion - stat hit, logo sting, headline, under 10s |
motion-graphics |
square, unnarrated |
| captions, talking-head overlays, music video, Remotion port, Figma input |
let hyperframes route |
- |
| anything else, or "build it with me" |
general-video |
- |
Destination shorthand: x → x-feed (1080x1080), wide → 1920x1080,
vertical → 1080x1920.
Elsewhere: AI-generated footage (text-to-video) → vd:omnimedia; generated
clips can come back as footage. Slide decks → vd:opendesign. Recording a real browser flow →
vd:agent-browser; its recording can come back as footage for general-video.
A still promo image → vd:marketing-design.
3. Intent layer, compressed
Run upstream's intent layer (hyperframes → references/intent-interview.md)
with these overrides:
--auto or "just build it" is upstream's lock signal: flow: automation,
storyboard: no, every unanswered field becomes a defaulted decision listed
in the heads-up.
--storyboard locks storyboard: yes.
- Otherwise ask the route's open must-haves in ONE round - one question per
field, recommended option first - not one message per field. Values the
table above or the input already settles are recommendations with receipts.
- Pitch round only for a truly unformed request. A PR, tag, or URL plus an
angle is formed.
4. Where the project lives
Scaffold under the hook-injected Visuals: path: run the workflow's own
npx -y hyperframes@latest init <slug> ... command from that directory, adding
--non-interactive. No Visuals: path injected → ask once where to put it.
Never scaffold inside a repo's tracked tree. Run every later command
(media-use, check, preview, render) from the project root, since
--project . and .hyperframes/ resolve against the current directory.
5. Media without HeyGen
No heygen CLI here, so upstream's free catalog path (bgm, image search, TTS)
is off, and its local Kokoro voice is not installed. Do not suggest installing
either. Use:
| Need |
Source |
| image |
npx hyperframes media-use resolve --type image --provider codex --intent "<what>" --project . (ChatGPT sub, ~60s each) |
| logo |
upstream default (theSVG → GitHub avatar → favicon), free |
| sfx |
npx hyperframes media-use resolve --type sfx --intent "<what>" --project . - bundled offline set |
| voiceover |
vd:omnimedia MiniMax speech (speech-2.8-hd) |
| music bed |
vd:omnimedia Lyria (instrumental only) - or no music. Not MiniMax music: it sings |
Register every file generated outside media-use with its real type
(--adopt labels all audio as bgm):
npx hyperframes media-use resolve --type voice --from <file> --intent "<scene/line>" --project .
MiniMax calls cost money: confirm before an agent-initiated call; a
user-requested one just runs.
6. Finish
npx hyperframes check --snapshots passes, and look at the PNGs it saves
under .hyperframes/artifacts/snapshots/ before calling it done. Upstream examples can fail their own check, so never
trust a template blindly.
- Skip the HyperFrames desktop app / Framey lines the CLI prints. Mention the
app only if asked how to keep editing; the Studio preview is enough.
- Report the absolute MP4 path plus duration, resolution, and size from
ffprobe.
- Hand off:
- X:
vd:devlog ... no-cover drafts the post around the MP4 (the video is
the visual), vd:twitter posts it. Keep
it under 2:20 (X cap for non-Premium accounts).
- PR:
gh cannot upload video. Attach it to a release
(gh release upload <tag> <mp4>) or drag it into a PR comment via
vd:agent-browser.
- When the user likes a result, accept upstream's recipe-freeze offer with a
stable name (
pr-clip, release-reel, launch-promo) so the next run of
that format adopts it in one question.
1---2name: video3description: "Make, edit, or render videos with HyperFrames (HTML to deterministic MP4), tuned to this toolchain: PR and release videos after ship, product promos from a URL or show-off page, concept explainers, and short motion graphics for X devlog posts. Front door for HyperFrames - read this before the hyperframes skill: it bootstraps the upstream skills, pre-selects the workflow, sets media and output defaults without a HeyGen account, and hands the MP4 to devlog/twitter. Use when the user asks to make, create, or render a video, PR or changelog video, release reel, launch promo, demo clip, motion graphic, or explainer, or mentions hyperframes. Not for AI-generated footage or text-to-video clips (vd:omnimedia), slide decks (vd:opendesign), recording a browser session (vd:agent-browser), still images (vd:marketing-design), or a video the user wants built in Remotion (porting Remotion source to HyperFrames is in scope)."4license: MIT5argument-hint: "[pr|release|promo|explain|motion|edit|render] [<PR ref|tag|URL|path|topic>] [x|wide|vertical] [--auto] [--storyboard]"6metadata:7 source: https://skillmd.com/skills/vanducng/video8 author: vanducng9 version: "1.0.0"10---1112# Video1314HyperFrames renders HTML compositions to MP4. The upstream skills own the craft15(composition contract, animation, CLI, media). This skill owns only the16personal layer: setup, routing from daily work, media providers, where files17land, and delivery. Never restate upstream rules from memory - load the18upstream skill that owns them.1920## 1. Setup (once per session)2122```bash23node -v # >= 2224npx -y hyperframes@latest skills check || npx -y hyperframes@latest skills update25```2627`skills update` installs the core set (`hyperframes`, `hyperframes-*`,28`media-use`) into `~/.claude/skills` and `~/.agents/skills`; creation workflows29install on demand at routing time. A failed install is a visible failure -30surface it, never continue from remembered upstream content. On a new machine31run `npx -y hyperframes@latest doctor` once (FFmpeg required; headless Chrome self-installs).3233Never name a local skill `hyperframes-*` - the upstream installer claims that prefix.3435## 2. Route3637Pre-select from the table, then read the `hyperframes` skill and follow its state table38(resume, edit, and render-only rows still win). The route file39(`references/routes/<workflow>.md`) still has the final say.4041| Ask | Workflow | Defaults |42|---|---|---|43| `pr` - "video for this PR", after `vd:ship` | `pr-to-video` | 16:9, developers. Angle from the PR title's conventional-commit type: `feat` → feature-reveal, `fix` → fix-explainer, `refactor`/`perf` → refactor-walkthrough, else changelog |44| `release` - a tag or "this release" | `faceless-explainer`, input = `gh release view <tag> --json body -q .body` | square, 45-60s. One scene per theme, never one per bullet - visualize, don't list |45| `promo` - product, docs site, or `vd:show-off` page URL | `product-launch-video` | 16:9 master |46| `explain` - concept, lesson, devlog idea | `faceless-explainer` | square |47| `motion` - stat hit, logo sting, headline, under 10s | `motion-graphics` | square, unnarrated |48| captions, talking-head overlays, music video, Remotion port, Figma input | let `hyperframes` route | - |49| anything else, or "build it with me" | `general-video` | - |5051Destination shorthand: `x` → `x-feed` (1080x1080), `wide` → 1920x1080,52`vertical` → 1080x1920.5354Elsewhere: AI-generated footage (text-to-video) → `vd:omnimedia`; generated55clips can come back as footage. Slide decks → `vd:opendesign`. Recording a real browser flow →56`vd:agent-browser`; its recording can come back as footage for `general-video`.57A still promo image → `vd:marketing-design`.5859## 3. Intent layer, compressed6061Run upstream's intent layer (`hyperframes` → `references/intent-interview.md`)62with these overrides:6364- `--auto` or "just build it" is upstream's lock signal: `flow: automation`,65 `storyboard: no`, every unanswered field becomes a defaulted decision listed66 in the heads-up.67- `--storyboard` locks `storyboard: yes`.68- Otherwise ask the route's open must-haves in ONE round - one question per69 field, recommended option first - not one message per field. Values the70 table above or the input already settles are recommendations with receipts.71- Pitch round only for a truly unformed request. A PR, tag, or URL plus an72 angle is formed.7374## 4. Where the project lives7576Scaffold under the hook-injected `Visuals:` path: run the workflow's own77`npx -y hyperframes@latest init <slug> ...` command from that directory, adding78`--non-interactive`. No `Visuals:` path injected → ask once where to put it.79Never scaffold inside a repo's tracked tree. Run every later command80(`media-use`, `check`, `preview`, `render`) from the project root, since81`--project .` and `.hyperframes/` resolve against the current directory.8283## 5. Media without HeyGen8485No `heygen` CLI here, so upstream's free catalog path (bgm, image search, TTS)86is off, and its local Kokoro voice is not installed. Do not suggest installing87either. Use:8889| Need | Source |90|---|---|91| image | `npx hyperframes media-use resolve --type image --provider codex --intent "<what>" --project .` (ChatGPT sub, ~60s each) |92| logo | upstream default (theSVG → GitHub avatar → favicon), free |93| sfx | `npx hyperframes media-use resolve --type sfx --intent "<what>" --project .` - bundled offline set |94| voiceover | `vd:omnimedia` MiniMax speech (`speech-2.8-hd`) |95| music bed | `vd:omnimedia` Lyria (instrumental only) - or no music. Not MiniMax music: it sings |9697Register every file generated outside media-use with its real type98(`--adopt` labels all audio as bgm):99100```bash101npx hyperframes media-use resolve --type voice --from <file> --intent "<scene/line>" --project .102```103104MiniMax calls cost money: confirm before an agent-initiated call; a105user-requested one just runs.106107## 6. Finish108109- `npx hyperframes check --snapshots` passes, and look at the PNGs it saves110 under `.hyperframes/artifacts/snapshots/` before calling it done. Upstream examples can fail their own check, so never111 trust a template blindly.112- Skip the HyperFrames desktop app / Framey lines the CLI prints. Mention the113 app only if asked how to keep editing; the Studio preview is enough.114- Report the absolute MP4 path plus duration, resolution, and size from `ffprobe`.115- Hand off:116 - X: `vd:devlog ... no-cover` drafts the post around the MP4 (the video is117 the visual), `vd:twitter` posts it. Keep118 it under 2:20 (X cap for non-Premium accounts).119 - PR: `gh` cannot upload video. Attach it to a release120 (`gh release upload <tag> <mp4>`) or drag it into a PR comment via121 `vd:agent-browser`.122- When the user likes a result, accept upstream's recipe-freeze offer with a123 stable name (`pr-clip`, `release-reel`, `launch-promo`) so the next run of124 that format adopts it in one question.