Repository navigation
Video Types
English · 中文
The agent routes every request to one of 9 video types and then reads that type's workflow. Type 09 is experimental. A type doc covers when to use the type, the engine, the steps, what looks good, what is banned, a prompt block that bin/vh new pastes into your brief (the type's defaults, weighed once the concept is set), the checks to run, and cases to learn from. Its bans come in two kinds: those about facts, licences, legibility and privacy are floors; those about looks (palette, texture, easing, chart style) are defaults a concept may override. The routing table is in CLAUDE.md.
| # | Type (bin/vh new name) |
For | Default engine | Look at | Doc |
|---|---|---|---|---|---|
| 01 | Math and science explainers (math) |
a principle, theorem, algorithm or physical process, 3Blue1Brown style; usually 1–10 min, 16:9 | Manim CE | showcase 03 (Fourier series), Code2Video case | 01 |
| 02 | Knowledge shorts, vertical or wide (short) |
30–90 s explainers for Douyin, Bilibili, Xiaohongshu, Shorts or TikTok, with no presenter on camera | HyperFrames | showcase 02 (LEO Doppler), black-hole explainer case | 02 |
| 03 | Product and launch films (promo) |
10–90 s launch films, app and SaaS promos, feature reveals, project intros | HyperFrames | showcase 00, showcase 04, HeyGen launch films case | 03 |
| 04 | Lyric videos and MVs (mv) |
a video for a song you made or licensed: a painted MV with lyrics off screen, or kinetic type | ClaudeAnimationBase (p5.brush) or HyperFrames | P(doom) case, Functional Emotions case | 04 |
| 05 | Data stories (data) |
one conclusion from data: a trend, comparison, ranking or distribution | HyperFrames + SVG | no showcase yet; the presets editorial-data, bubble-chart-story, isotype, and the whole-film skeleton data-along-time
|
05 |
| 06 | Paper and conference videos (paper) |
paper explainers, ~30 s teasers, 3–5 min explainers, talk recordings, supplementary videos | Manim (mechanisms) + HyperFrames (structure, results) | Paper2Video case, Code2Video case | 06 |
| 07 | Hand-drawn, watercolour, whiteboard, paper-cut (handdrawn) |
character shorts and hand-drawn explainers; Chinese characters written in stroke order | ClaudeAnimationBase (p5.brush, bundled) | showcase 01 (Clawd and the leaf), P(doom) case | 07 |
| 08 | Brutalist, meme and fast-cut edits (meme) |
15–40 s memes, opinion cuts, visual hooks for tech social media | HyperFrames |
Claude Pop case, the preset brutalist-meme
|
08 |
| 09 | Editing your own footage, talking heads (edit), experimental
|
cutting a talking head, interview or vlog you recorded yourself: remove filler and retakes, cut at the quiet gaps, add subtitles and explainer graphics, make a vertical version | HyperFrames | no showcase yet; community prompts (a talking-head graphics workflow) | 09, details in engines/editing.md |
About type 09. It was written from a survey of similar projects and has never cut a real film: its thresholds were calibrated on synthetic material only (a synthesised voice with noise), and no exported file has been opened in an editor. The tools the docs mention (a gap probe, seam metrics, an edit-list compiler) aren't written yet; the fastest route today is HyperFrames' transcript-cut.mjs. What the type does settle: cuts are placed on measured quiet gaps, not on speech-recognition word times; the edit list is stored in whole frames, and nothing renders until you approve its hash; and your footage is your face and voice, so transcription runs locally by default and cloud services come only after the agent asks you.
- One sentence is the normal case. The agent thinks of a concept first, fills the rest of the brief from the type's defaults and asks at most three questions that change the result.
- A video can span two types, say a paper explainer cut as a vertical short. The agent reads both docs, and the main type wins where they disagree.
- When nothing fits, the agent picks an engine from the table in playbook/00 and writes down why in
DECISIONS.md.
A type names a default engine, and the concept can ask for another. The repo has five: HyperFrames (HTML + GSAP); canvas-film (canvas 2D, for long narrated films whose shots and captions hang on the narration's words, with an optional Three.js layer for 3D); ClaudeAnimationBase (p5.js + p5.brush), bundled; Manim CE, with ManimGX as an optional faster backend; and Blender 5.2, optional, for path-traced shots or for tidying a 3D asset. What each draws and where it is used is in the README's Engines table. Remotion appears in a few type docs as an alternative; it is set up by hand, and no showcase film uses it.
| You want | Where the guidance is |
|---|---|
| Realistic people, real physics, a live-action feel | playbook/05: a generative video model with a code layer on top, organised by task (taking a clip in, matching it to code-rendered footage, keying, a character across shots, artefact checks, cost). Dated prices and specs are on genvideo-market, licences and disclosure on genvideo-ledger. It costs money and results vary, so the budget is confirmed with you in the brief first. Experimental: no video model has been called yet. Footage you recorded yourself isn't generated: that's type 09 |
| Effects, transitions, particles, shaders, audio-reactive picture | playbook/08 |
| How should this shot move? Pacing for a whole film |
recipes/: 71 shot recipes (camera, seams, rhythm, type, UI, data, effects, openings and endings), 24 shot-grammar cards (shot sizes, angles, the axis, camera moves) and 10 whole-film skeletons: three by energy arc (a 15 s launch, a 30 s explainer, a 60 s product film), five that start from a concept (time-dilation replay, machine-eye portrait, making-of-self, mascot style tour, workflow as story) and two by subject (a data story along a timeline, a live-countdown split screen). bin/vh recipes list filters them by what the shot must say and how loud it is. See Style Library
|
| A story with rises and falls, or a film of 3 minutes or more | playbook/09: structures, fillable beat sheets, the curiosity gap and energy curve, act breaks |
| Posting to a short-video platform: the hook, title and cover |
playbook/10; bin/vh cover-preview shows a cover at real feed sizes |
| Music with chapters, a theme you can hum, real rises and falls | playbook/11; see Sound and Voice |
| 3D or shader films (Three.js) | A narrated film with 3D in it: --engine canvas-film --three draws GLB models under the 2D with their animation clips set to the narration's words; bin/vh model3d check reads a model before it goes in, and bin/vh hdri takes a CC0 environment map to light it. There is no 3D type doc yet: use the promo type's motion rules plus playbook/08 (one-take camera paths, effect presets, sub-frame motion blur); showcase 04 is the worked example. For path-traced light (glass, skin, volumetric light), physical simulation or real depth of field there's a Blender guide, partly verified: Blender 5.2.2 rendered the tabletop-miniature swatch and the intro film's opening, a galaxy driven frame by frame from numpy; some of its commands haven't been run yet. A 3D type doc is on the Roadmap
|
| "How was this video made?" |
playbook/07: break a reference film down; the result is a new case study in cases/
|
bin/vh types # the 9 types and their docs
bin/vh new short leo-doppler # projects/<date>-leo-doppler with a portrait HyperFrames scaffold
bin/vh new math fourier --style dark-math # with a style preset as a reference
bin/vh new promo launch --effort studio # with an effort level
bin/vh new short explainer --aspect 16:9 # short, promo, data and meme: choose the frame (16:9, 9:16, 1:1, 4:5, 3:4)
bin/vh new promo launch --watch feed --res 4k # where it will be watched (sets the text-size floor) and the output resolution
bin/vh new paper talk --engine canvas-film --three # the narration-driven canvas engine with a Three.js layer; any type, 16:9
bin/vh new short episode-2 --from <earlier-project> # the next episode of a series: spec, style and outline skeleton carry over
bin/vh new edit my-interview # type 09; then bin/vh hf-init projects/<…> portraitYou rarely need to run these yourself: the agent does it once it has routed your request. bin/vh new -h lists every option. Choosing a look comes next: Style Library.