pich

Drop a raw video. Get a finished Reel. Type /pich video.mov and your AI agent turns a raw talking-head clip into a finished vertical Reel/Short/TikTok — cuts, captions, cards, music — and checks it automatically before you post. A Claude Code skill (works with any agent that can run shell commands) plus plain Python/FFmpeg scripts.

Türkçe README

What you get

From one plan.json and one FFmpeg pass:

• Speech-safe cuts — pauses removed, never a clipped syllable (every cut edge is measured).
• Verified captions — two Whisper models plus a per-phrase isolated pass flag misheard words before they become captions.
• Hook card on frame 0, info cards, icons (200k+ open-source via Iconify), verified brand logos.
• Step slide with live PiP, or a brag-style Hyperframes motion segment (animated counters, charts) with the speaker still on screen.
• Stock B-roll from Pexels: full-screen cutaways, or video cards over the camera.
• Music ducked under the voice, SFX, broadcast loudness (-14 LUFS).
• Hook A/B previews: render the first 3 s of several hooks side by side, pick one.

Then qa_reel.py checks the result: a layer covering the eyes or mouth (macOS Vision on the clean source frame), captions the speaker doesn't actually say, music masking a word, A/V drift per segment, loudness, black last frame, estimated Instagram UI overlap. Broken renders fail loudly, so you never post one by accident.

Requirements

• macOS or Linux, Python 3.11+, ffmpeg/ffprobe, pip install pillow
• whisper.cpp (whisper-cli on PATH) + models: scripts/setup_models.sh (~2 GB)
• Optional: Node 20+ for Hyperframes motion segments (npx hyperframes), librsvg (rsvg-convert) on Linux for icons/logos
• Optional keys (free): PEXELS_API_KEY for stock, BRANDFETCH_API_KEY for logos Simple Icons lacks
• The face check uses macOS Vision; on Linux it is skipped with a warning

Install as a Claude Code skill

git clone https://github.com/<you>/pich ~/.claude/skills/pich
~/.claude/skills/pich/scripts/setup_models.sh

Then, in Claude Code:

/pich ~/Movies/take3.mov

The agent builds the first pass automatically (cuts, verified captions placed away from the face, music, loudness), checks the words the speech models disagreed on, writes a hook and adds cards/icons/B-roll for what you actually say, renders, runs the QA and hands you the finished file. Add direction in plain words: /pich take3.mov --lang tr, punchy, no music.

No agent? One command does the automatic pass:

python3 scripts/pich.py take3.mov --lang en --hook "TIP|Raw take|finished Reel" --cta "Follow|for more|Follow"

Try it without your own footage

python3 examples/demo/make_demo.py                     # synthetic 12 s source + B-roll + plan.json
python3 scripts/render_reel.py examples/demo/plan.json demo.mp4 --preview
python3 scripts/qa_reel.py demo.mp4 --plan examples/demo/plan.json --no-asr

examples/motion-template/ is a ready Hyperframes segment ("When does it pay off?": setup hours → minutes saved per day → break-even day → profit) to adapt for your own numbers.

How it works

Every overlay is anchored to source seconds inside a kept segment, so changing the cuts never slides a card away from the words it belongs to.

Rights and privacy

• No music and no brand logos are bundled. Music needs "rights_checked": true (or the generated pad); logos need a source URL and a verification date. Trademarks belong to their owners — show a logo only when the brand is named.
• Stock and icons record author and license in the render manifest.
• Keys are read from environment variables only; nothing is uploaded except the API calls you run.

Tests

python3 -m unittest discover -s tests -v

Synthetic fixtures (no footage needed), a golden filter graph for the demo plan, and QA that must FAIL on broken renders. Tests that need network, keys or models skip themselves.

License

MIT — see LICENSE. Third-party components: THIRD_PARTY_NOTICES.md.