📺 Example video: Meituan's LongCat-2.5: This 1.6T AI Agent Beats GPT-5.5 — Full Breakdown — produced end-to-end with the model-release pipeline from this repo.
Linkmotion is an opinionated, instruction-driven workflow for producing short-form and long-form AI videos from a brief, URL, or topic. Instead of hard-coding a single pipeline, it gives an AI coding assistant a project structure, pipeline manifests, review checks, and tool registry so it can research, script, render, and checkpoint work end-to-end.
This repo is designed for:
- AI shorts and explainers
- model release explainers
- YouTube-style AI storytelling videos
- pipeline-driven production with human approval loops
The repo does not try to replace the agent. It gives the agent:
- a set of production tools
- a pipeline manifest describing stages and outputs
- skill files explaining how to perform each stage
- schemas for validating outputs
- checkpointing for resumable work
In other words: your AI assistant becomes an orchestrator for a video production pipeline.
pipeline_defs/— pipeline blueprints for different video typestools/— concrete production tools (search, capture, TTS, render, etc.)lib/— project setup, config, pipeline loading, checkpointing, and style logicskills/— authoring guidance and stage-by-stage instructionsschemas/— JSON schema contracts for proposals, scripts, storyboards, report filesstyles/— visual identity and motion style definitionsprojects/— generated video workspaces, one folder per projecttests/— repository validation tests
You need:
- Python 3.10+
- Git
- ffmpeg and ffprobe installed and on PATH
- A working AI coding assistant or terminal environment with access to the repo
For local render and browser workflows, the repo expects Python packages such as Playwright, edge-tts, pyyaml, and jsonschema. These are installed with the setup commands below.
Clone and set up the environment:
git clone <repo-url>
cd linkmotion
make setupThis creates a local virtual environment and installs the Python dependencies from requirements.txt.
If you want the dev environment too:
make install-devRun the repo preflight check:
python -m ai_news_agent preflightThis checks external dependencies and the discovered tool registry. It also flags missing render or browser support.
List available pipelines:
python -m ai_news_agent pipelinesInitialize a new project with a slug, pipeline, topic, and optional source URL:
python -m ai_news_agent init 2026-11-gpt5-short ai-news-short "GPT-5 launch coverage" "https://example.com/article"This creates a new project folder under projects/<slug>/ and writes progress/state files. Project names should be kebab-case.
The repo ships with several video production pipelines:
| Pipeline | Manifest | Best for |
|---|---|---|
ai-news-short |
pipeline_defs/ai-news-short.yaml |
45-75 second AI news short / vertical reel |
ai-news-long |
pipeline_defs/ai-news-long.yaml |
7-8 minute deep-dive or roundup |
model-release |
pipeline_defs/model-release.yaml |
model / LLM release explainers |
You can inspect the pipeline config directly in pipeline_defs/*.yaml to see stages and requirements.
The workflow is intentionally stage-based and review-heavy.
- Read the pipeline manifest
- Run preflight
- For each stage, read the matching skill file before making changes
- Generate or select a style for the video
- Review the output and save a review artifact
- Update checkpoint/progress state
- Request approval when the stage requires it
The routing and stage guidance are defined in AGENT_GUIDE.md.
Example flow:
python -m ai_news_agent preflight
python -m ai_news_agent pipelines
python -m ai_news_agent init my-ai-news-short ai-news-short "Claude Code release analysis"Then work inside projects/my-ai-news-short/ and follow the stage instructions in skills/pipelines/....
.
├── AGENT_GUIDE.md # routing and production workflow rules
├── AGENTS.md # high-level agent instructions
├── README.md # this file
├── Makefile # setup / preflight / test shortcuts
├── config.yaml # default config for model/budget/output setup
├── requirements.txt # runtime dependencies
├── requirements-dev.txt # development dependencies
├── pipeline_defs/ # pipeline manifests
├── skills/ # agent playbooks and stage instructions
├── tools/ # tool implementations
├── lib/ # core libraries
├── schemas/ # JSON contracts
├── styles/ # YAML style definitions
├── projects/ # generated project workspaces
├── tests/ # test suite
├── ai_news_agent/ # CLI entry point
└── out/ # rendered output directory
The repo auto-discovers tools via tools/tool_registry.py. This makes tool availability explicit and lets the agent adapt to what is installed.
Key capability families include:
web_search— research + verificationyoutube_source— fetch and normalize YouTube source materialbrowser_capture— Playwright-based page capturebroll— image/video / social asset search and capturetts— edge-tts narration and word timingsfx— sound effects and audio polishvideo_compose— final render / compositionmetadata— title, description, tags, thumbnail metadata
Each video project is created under projects/<slug>/ and usually includes:
proposal.json— project brief / conceptresearch.mdorresearch_brief.json— research-backed notesscript.json— narrative scriptstoryboard.json— beat-by-beat visual planprogress.json— stage tracking and checkpoint statedecision_log.jsonl— append-only log of decisionsreview-*.md— review notes generated for each stageassets/,audio/,captures/,renders/,src/— project output folders
The project folder is where the actual production work happens.
The repo uses a style engine and YAML playbooks to create a unique visual identity per video. These lives under styles/ and are associated with each pipeline.
The standard behavior is:
- generate a style identity for the project
- tailor the motion, typography, and look to the topic
- keep the style consistent across the script, storyboard, and composition stages
This repo expects iterative review, not blind execution. Outputs are meant to be checked by the agent and by you.
Important patterns:
- every stage has review criteria
- progress is checkpointed in JSON
- decisions are logged append-only
- human approval is expected for creative gates
- quality reviews happen after each stage
See AGENT_GUIDE.md and the files under skills/meta/ for the review/checkpoint protocol.
# install deps
make setup
# install dev deps
make install-dev
# verify tool availability
python -m ai_news_agent preflight
# list manifests
python -m ai_news_agent pipelines
# scaffold a project
python -m ai_news_agent init <slug> <pipeline> "<topic>" [source_url]
# run tests
make test
# compile check
make lintShort AI video:
python -m ai_news_agent init gpt5-launch ai-news-short "GPT-5 launch coverage"Long AI analysis:
python -m ai_news_agent init claude-code-deep-dive ai-news-long "Claude Code internals explained"Model release explainer:
python -m ai_news_agent init gemini-omni-flash model-release "Gemini Omni Flash release breakdown"Then open the generated project and proceed stage by stage.
- This repo is optimized for agent-driven production rather than manual scripting.
- It is best used with an AI coding assistant that can read pipeline manifests and skills and execute tools.
- The design philosophy is grounded claims, structured review, and resumable work.
- The repo prefers explicit checkpoints over implicit state.
MIT — see LICENSE for the full text.
This project expects a pipeline-first convention. If you add new tools, skills, or pipeline stages, follow the patterns already used in tools/, skills/, and pipeline_defs/ rather than creating ad hoc workflows.