Make voiced kinetic Shorts
Everything renders with the bundled pipeline in pipeline/ next to this file
(the "Base directory for this skill" shown when the skill loads).
0. Setup (once per project)
If the project has no video_gen/ folder with kinetic_build.py, create it:
mkdir -p video_gen && cp <skill dir>/pipeline/* video_gen/
cd video_gen
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt # drawing (Python 3.11+)
python3.12 -m venv .venv-tts && .venv-tts/bin/pip install -r requirements-tts.txt # Kokoro voice
ffmpeg must be on PATH. voice.py runs the TTS worker in .venv-tts next to it.
Run every command from video_gen/.
1. Read the rules
DESIGN.md- the kinetic visual style. Hard rules: boxes stack in one full-width column, never side by side; no accent bars on box edges; big text; each element appears on its spoken word; headings fit one line (about 30 characters); no burned-in subtitles.PARTS.md- thepart.mdformat, narration rules, and how to split a guide. Don't readengine_short.py,templates.pyorskin.pyunless debugging them.
2. New series: split first, get approval
- Analyse the source guide and propose a part list (one concept per part, about 90-125 s each). Stop for user approval. Never cut content to hit a length; add examples and the "why" instead.
- Write every
<series>/pNN_<slug>/part.md, plus<series>/README.mdwith the analysis and a parts table. - Add the series folder to
SERIESinbuild_part.py(kicker + footer). Optionally add a logo mark toMARKSinkinetic.py. - Lint narration:
.venv/bin/python build_part.py <part dir> --lintand fix every problem. Narration: no recap, no welcome, jump straight to the topic; no symbols or code; spell out acronyms for the voice; beats at most 45 words.
3. Scenes
Each part gets a scenes.py defining SCENES. Two styles (see examples/):
- Declarative stage blocks (preferred, fast):
SCENES = K.part(word, tag, chips, [[blocks...], ...])usingTrack, Pipe, Rows, Code, Lanes, Bars, Ring, Fan, Chips, Big, Note. A cue is a word prefix (searched after the block's previous cue),(word, beat)or a beat number. Example:examples/p03_*. - Hand-written
Sceneclasses for a metaphor the blocks can't show. Example:examples/p01_*. List cue words in spoken order, or cues come back MISSING.
4. Check, then render
.venv/bin/python kinetic_build.py <part dir> --stills /tmp/pNN.png # contact sheet + cue report
.venv/bin/python kinetic_build.py <part dir> --out <series>/output # final <slug>_voiced.mp4
.venv/bin/python cue_scan.py <series>/p*_* # text-only sync check
Fix every MISSING cue and overflow before the final render. For a whole series, render the first
two parts and get the user's OK on the look, then run the rest 3 at a time in the background
(xargs -P 3).
5. Report
A table: MP4 link, topic, length, longest silence (target under 1 s). Mark the parts rendered in the series README. Say plainly what was not checked (frames not viewed, audio not listened to).
Never
- Put boxes side by side, edge accent bars, small text, or burned-in subtitles.
- Propose per-video voice calibration; sync is word-timed and generic.
- Edit
engine_short.pydirectly;build_part.pypatches it in memory.