Complete-workflow video compilation anchored to word semantics
Traditional AI video generation treats output as an uneditable black-box pixel file. Even a small script edit or change in speech rhythm requires all downstream text cards, B-roll, sound effects and shot transitions to be discarded and re-edited.
Hypit converts video production into a declarative compilation pipeline. Through SVML markup and the SVS style library, Agents anchor visual actions, overlays and animated text cards to semantic times from WhisperX word alignment (Semantic Anchors) instead of fixed seconds. When narration changes or speaking speed varies, all layers and sound effects automatically align to the corresponding words.
HyperFrames is the rendering core. Up to 64 headless Chromium processes concurrently draw DOM, Canvas and WebGL frames, which ffmpeg then encodes as pixel-level high-quality video.
Reference analysis (Probe)
Script semantics (Script)
Prompt library (Kits)
Build planning (Plan)
Concurrent rendering (HyperFrames)
Visual review (Studio)
Hypit gives AI agents a language and system to create video. Drop in a video, and your agent clones it as a complete workflow: footage, captions, B-roll and effects, all anchored to words instead of seconds.
Skill installation and Runtime setup
Hypit separates the Skill, which provides production knowledge, from the @hypit/hypit package, which provides the execution toolchain. Run the global installation command in a Coding Agent that supports Skills, such as Claude Code or Codex:
# Install globally in a terminal or agent environment
npx skills add hypit-ai/hypit -gMonorepo development: local source compilation and testing
To customize underlying components, extend Provider adapters or compile Hypit locally, copy the repository from GitHub, then install dependencies and check types in order:
# Enable corepack and install locked dependencies
corepack enable
pnpm install --frozen-lockfile
# Check TypeScript types and unit tests
pnpm check
pnpm testCLI commands and declarative modules
Hypit separates video engineering into three independent dimensions: project declarations (Markup/DSL), execution control (CLI toolchain) and visual review (Studio). Developers and Agents use the following core commands for the complete workflow from reverse-engineering a reference video and aligning speech to concurrent rendering.
Preparation · 01
hypit media
Media reverse probing
Extract frame sequences, text boundaries and layout grids from a reference video, deriving shot rhythm and visual hierarchy.
Acoustics · 02
hypit transcribe
Word-level speech alignment
Use WhisperX to resolve audio precisely to individual word timestamps, creating Semantic Anchors for video layers.
Speech · 03
hypit measure
Local speech-rate estimates
Estimate script duration in advance from language, pace (fast/normal) and rounding policy without calling external APIs.
Budget · 04
hypit plan & pricing
Cost budget assessment
Analyze the .svrun dependency graph, list required model call Endpoints and estimated generation costs, and confirm budget authorization.
Compile · 05
hypit build
Multithreaded compilation and build
Call generation models, normalize assets and start the 64-process Chromium rendering engine to output video.
Review · 06
hypit studio
Visual review canvas
Start a local Web editor on port 5179 with timeline navigation, live Inspector property adjustments and timestamped feedback.
Runtime · 07
hypit runtime
Environment configuration pipeline
Manage local and cloud Runtime Profiles, switching between your own API keys and the HypiHub hosted service.
Security · 08
hypit auth
Credential and key isolation
Centrally manage third-party model API credentials, supporting secure system storage and environment-variable isolation to prevent plaintext leaks.
Command selection by task stage
| Task | Core command | Output and purpose |
|---|---|---|
| Extract reference-video rhythm and shot structure | hypit media tile <video.mp4> --step 1 | Frame-grid image and temporal split points |
| Estimate script pronunciation duration and rhythm budget | hypit measure <story.svml> --pace fast | Estimated seconds and pronunciation labels without using API quota |
| Confirm generation-model dependencies and billing limits | hypit plan <run.svrun> --runtime <prof.json> | Assets to generate and itemized billing endpoints |
| Execute the complete pipeline and render video | hypit build <run.svrun> --follow | Generate A-roll/B-roll and encode mp4 through Chromium |
| Adjust timeline and provide asynchronous comments | hypit studio --run <run.svrun> | Interactive browser timeline and FEEDBACK.json annotations |
Semantic anchors and pipeline compilation principles
Hypit's compilation architecture uses linguistic semantics and a responsive canvas, without the traditional Timecode editing model. The core principles are defined in official specification documents and SKILL.md:
RULE · Word anchors instead of fixed seconds
In SVML, visual events, B-roll, sound effects and karaoke captions anchor to semantic labels in the script. When a speaker re-records or changes speed, WhisperX word alignment automatically recalculates absolute times for all child layers without manual frame-by-frame trimming.
Source · Official SKILL.md · Script and Time
RULE · Three files separate creative responsibilities
The project separates .svml (structure and script), .svs (reusable visual appearance and prompt Recipes) and .svrun (target execution and asset reuse strategies). These are decoupled, so the same viral shot structure can be quickly applied to different subjects and products.
Source · Official docs/guide/conventions.md
RULE · Asset reuse and continuation after failure (Candidate Reuse)
Generation-model calls cost money and time. Hypit's Run Candidates track existing Outputs. Even after rendering fails midway or only caption styles are adjusted, Build can select existing assets directly to avoid repeated paid API calls.
Source · Official docs/quickstart/run.md
RULE · HyperFrames 64-process concurrent Chromium rendering
All video layers use modern Web technologies (HTML, CSS, Canvas, WebGL). HyperFrames starts up to 64 headless Chromium processes to draw multiple frames in parallel, then ffmpeg combines the stream for pixel-level animated layouts.
Source · Official README.zh-CN.md · HyperFrames
RULE · Key isolation and budget approval prerequisites
The framework is open source and free, but external generation models consume quota. Before paid tasks, Hypit requires hypit plan to provide endpoint lists and price estimates, and isolates secret keys in secure storage to prevent plaintext leaks.
Source · Official SKILL.md · Standing Responsibilities
RULE · Dual Studio views and asynchronous comment feedback
hypit studio has two views. Studio displays the timeline and component properties; Comments (#comments) lets you click frames to leave timestamped feedback, automatically persisted in FEEDBACK.json.
Source · Official docs/quickstart/preview.md
Viral football ranking video recreation workflow
The official ranking-football example shows how to submit a reference video to an Agent and replace the presenter and comedic B-roll images. The workflow covers semantic script labels, speech-duration estimates, approval of model-call costs and text-card alignment adjustments in Studio.
$ You ›
/hypit Use this video as reference: https://storage.googleapis.com/hypit-public-assets/quickstart/football_tier.mp4. Replace presenter with a goth host, and keep the energetic ranking format.
claude: Claude(Hypit Director)›
Project structure and semantic script created. Replace the presenter with a goth football commentator while retaining the tier-ranking motion and word-by-word karaoke captions.
Measuring speech pace and estimating duration locally:
> hypit measure reference.svml --segment ronaldo --language en --pace fast --rounding ceil
hl: Segment 'ronaldo': estimated 9s (requested duration 10s · 1s padding for facial delivery)
> hypit measure reference.svml --segment messi --language en --pace fast --rounding ceil
hl: Segment 'messi': estimated 9s (requested duration 10s · 1s padding for facial delivery)
claude: Claude ›
Analyzing the execution plan and costs:
> hypit plan reference.svrun --runtime ./hypit.runtime.json
· presenter.image: GPT Image 2K (1 request · $0.04)
· broll-*.image: GPT Image 2K (8 requests · $0.32)
· presenter-voice: Fish Audio VoiceDesign (1 request · $0.02)
· ronaldo-take & messi-take: Seedance 2 Mini 720p (2 video takes · $0.70)
· HyperFrames Composition: 64 Headless Chromium processes (Local · $0.00)
hl: Total Estimated Budget: $1.08. Approve execution?
$ You › Budget approved. Start the build.
> hypit build reference.svrun --runtime ./hypit.runtime.json --follow
[1/4] Generating presenter image & 8 comedy B-rolls (GPT Image 2K)... done (8.4s)
[2/4] Designing presenter voice reference (Fish Audio)... done (2.1s)
[3/4] Generating A-roll video takes (Seedance 2 Mini 720p)... done (34.2s)
[4/4] Normalizing media & rendering 64-process Chromium HyperFrames... done (12.8s)
ok: Build completed: dist/final.mp4 (duration: 20.4s · 720x1280 · 30fps)
claude: Claude ›
Video generation completed. Hypit Studio is running for timeline review:
> hypit studio --run reference.svrun --port 5179
In a clone, discover what a cut, picture, reveal or sound responds to, then recreate that relationship for the target's words and intention.
Production benefits of semantic compilation
In traditional video production, editors manually adjust entry and exit timecodes for dozens of assets. Hypit converts this into a declarative dependency graph: replacing a character or translating into a new language automatically rearranges all visual and sound events through semantic labels.
Using a verified viral structure, creators can produce hundreds of variants with a single command for content production at scale.
Environment dependencies and commercial license boundaries
Custom components and production pipeline extensions
Hypit has a highly modular Monorepo architecture. Beyond calling existing templates directly, developers can extend project components, adapters and visual controllers.
Advanced extension options
1. Custom Author Packages and project components. Use @hypit/author-kit to package channel-specific visual assets such as custom text-card layouts, branded animated transitions or split screens. Components can be included directly as markup tags in SVML.
2. Studio Companion visual controls. Write Companion descriptors for components to expose timeline handles and property sliders on the Studio canvas. Edits are written back automatically to SVML source.
3. Private models and on-premises inference endpoints. Extend Provider adapters with @hypit/model-kit and @hypit/endpoint-kit to switch audiovisual generation to self-hosted ComfyUI, Ollama or dedicated internal services.
Three further resources
① docs/quickstart.md: Complete onboarding and core workflow guide for Agent users. ② skills/hypit/SKILL.md: Director and producer prompt specifications and system protocol definitions. ③ examples/ranking-football/reference.svml: Complete declarative source example for the official standard ranking video.
To be clear: cloning a video is the fastest way in, not the only one. You can start from our templates, or just describe the video you want and your agent writes the workflow from scratch.