autolecture-skill
scao7/autolecture-skill
Claude Code skill: turn audio/podcast/script (optionally + PDF or GitHub repo) into a finished AutoLecture (https://autolecture.ai) video via the autolecture Python SDK. End-to-end upload + compile + download mp4.
Instalação
npx convez-ai scao7/autolecture-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
autolecture-skill
Claude Code skill that turns your material into a finished AutoLecture video. The entry point asks what kind of video you want, then routes to the matching workflow:
- plain text / topic → generated narration + hand-written visuals
- audio recording / podcast → transcribe → match visuals (rough re-synth or keep your voice)
- PDF paper → explain it (extract figures) or show it (render the real pages with react-pdf, zoom + locate + highlight — technique borrowed from pdf2video)
- real footage → overlay transparent motion graphics on top (
over=)
Claude generates the scenes, then — if you've connected the AutoLecture MCP connector — builds the project and compiles it in the cloud directly via MCP tools; otherwise it packages a zip you upload at autolecture.ai.
Install
# Add the skill to your agent — one line, works with Claude Code,
# Cursor, Codex, and 12+ other agents. Add -g for a global install.
npx skills add scao7/autolecture-skill
npx skills is Vercel Labs' open agent-skills tool;
it clones this repo into your agent's skills dir (~/.claude/skills/ for
Claude Code). Prefer git? Clone it yourself:
git clone https://github.com/scao7/autolecture-skill.git ~/.claude/skills/autolecture-skill
To let Claude compile in the cloud (so you get a finished video
without touching the web UI), connect the AutoLecture MCP connector
in your agent: Settings → Connectors → Add →
https://mcp.autolecture.ai/mcp → approve in the browser. Then Claude
creates the project, writes the files, and compiles via MCP tools end to
end. Without a connector the skill just packages a project zip you
upload at autolecture.ai — the path most
claude.ai web users take.
Per-input-type extras — install only what your input needs:
| If your input is | also install |
|---|---|
| audio (mp3 / wav / m4a) | pip install openai-whisper + system ffmpeg |
| PDF paper with figures | pip install pdfplumber Pillow + system pdftoppm |
| a GitHub repo | system git |
| plain text script | nothing extra |
System binaries: sudo apt install ffmpeg poppler-utils git (Ubuntu) /
brew install ffmpeg poppler git (macOS).
Migrating from an older copy of this skill? Wipe the old folder before cloning the new one:
rm -rf ~/.claude/skills/autolecture-demo ~/.claude/skills/autolecture-claude-skill.
Use
Open Claude Code in any working directory and ask it to do the thing, optionally attaching an input file:
"Use the autolecture skill to make a 30-second explainer on Bayes theorem."
"做个 autolecture demo"
--include recording.mp3"Show this paper in the video — flip pages, zoom, highlight."
--include paper.pdf"Overlay lower-thirds and callouts on my footage."
--include clip.mp4
Claude reads SKILL.md — a router — figures out which
input you have (asking if unclear), then opens the matching playbook in
workflows/ and follows it:
| Input | Workflow |
|---|---|
| plain text / topic | workflows/text-to-lecture.md |
| audio / podcast | workflows/audio-upload.md |
| PDF paper | workflows/pdf-paper.md |
| real footage | workflows/video.md |
Every workflow hand-writes a \manimFile / \htmlFile /
\remotionFile source per scene (no LLM-codegen render macros — the
skill bans those; see SKILL.md HARD BANS), then converges on the shared
delivery step workflows/_delivery.md: with the
MCP connector, Claude creates the project, writes the files, uploads
assets, compiles in the cloud, and fixes failing scenes by fetching
frames — all via MCP tools. Without it, Claude packages a zip you upload.
With the MCP connector on, a session ends like:
✓ compiled — 6 ✦ spent
open in Studio: https://autolecture.ai/studio?id=…
Follow the Studio URL to play it or keep tweaking (regenerate scenes, swap voices, add BGM, re-render). In zip mode you get a zip to drop on autolecture.ai instead.
Where this runs
| Surface | Works? | How |
|---|---|---|
Claude Code CLI (claude in a terminal) | ✓ | reads ~/.claude/skills/ |
| Claude Code desktop app (Mac / Windows) | ✓ | same |
| Claude Code IDE extensions (VS Code / JetBrains) | ✓ | same |
| claude.ai web → Customize → Skills → + | ✓ | upload this repo as a Personal Skill (zip or git URL) |
The web upload keeps a copy on Anthropic's side; the CLI / desktop /
IDE paths all read your local ~/.claude/skills/. Upgrade with
npx skills update (or re-clone if you installed via git).
Troubleshooting
- No cloud compile happening? — Claude only compiles when the AutoLecture MCP connector is connected (Settings → Connectors). Without it the skill produces a zip for you to upload at autolecture.ai.
- Compile fails — Claude inspects the failing scene (fetches a frame via MCP), fixes its source, and re-renders just that block. If it can't, it hands you the Studio URL — open it, fix the red block, recompile.
(429) rate_limitedor(402) insufficient_credits— the message includes window + limit + your current balance. Top up at https://autolecture.ai/account or wait for the daily reset.- Missing system binary (ffmpeg / pdftoppm / git) — the helper scripts hard-exit with the install command. Follow the message.
Links
- Skill spec — SKILL.md (authoritative; read this if you're extending the skill)
- Remote MCP — https://mcp.autolecture.ai/mcp (connect it in your agent for cloud compile)
- AutoLecture web app — https://autolecture.ai
- VideoTeX DSL reference — https://autolecture.ai/docs/dsl
License
MIT — see LICENSE.