Stick Figure

/stick-figure

Copy code
┌──────────────────────┐
│   ☀      ☁           │  bright sky, sea, props drawn in code
│      ◯   ◯           │  stick people with big white heads, expressive faces
│     /|\ /|\          │
│     / \ / \  🛢       │
│     eye patch        │  1-2 word captions, lowercase, ~70% down
└──────────────────────┘

A worked example comes with it: engine/template is a finished 37 s reel, "What if pirate eye patches had nothing to do with missing eyes?"

First use

This skill's folder is the "Base directory" Claude Code shows when it loads it.

bashCopy code
SKILL_DIR="<base directory of this skill>"
ENGINE="$SKILL_DIR/engine"
KIT=$(python3 -c 'import os,sys;print(os.path.realpath(os.path.join(sys.argv[1],"..","..","kit")))' "$SKILL_DIR")
[ -x "$KIT/.venv/bin/python" ] && [ -d "$KIT/models" ] || bash "$KIT/setup.sh" --voice   # ~1 GB of voice models, once
PY="$KIT/.venv/bin/python"; export NODE_PATH="$KIT/node_modules"

Needs ffmpeg, Node and Python 3.9+. To see the example: cp -r "$ENGINE/template" ~/reels/pirate && bash "$ENGINE/make.sh" ~/reels/pirate.

How it fits together

A reel is one folder:

• script.json: the voiceover, one sentence per entry.
• captions.json: optional text swaps for captions ({"twenty five": "25"}).
• web/shots.js: the shot list. Each shot is SHOTS.push({start, end, tr, draw(ctx, lt, t)}) in seconds of the voiceover; lt is time inside the shot, tr is the transition out ('whip' or 'cut').
• web/lib.js (easing, keyframes kf, shapes, text), web/chars.js (the stick rig drawStick, hats, faces, arm IK, the eye patch), web/env.js (sky, clouds, sea, ships, deck, barrels, cannon, lantern, below deck), web/main.js (whip blur, vignette, grain, captions). Add new props to env.js or a new env2.js (list it in web/index.html).
• audio.py: music and SFX, all synthesised in code and placed in voiceover seconds.

make.sh makes the voice, aligns every word (audio/words_new.json), mixes the sound, renders the shots in parallel and writes <folder>/<folder name>.mp4.

Timing: audio/words_ref.json is the voice the shots were timed against. When the voice changes but the words don't, make.sh warps the shots onto the new take, so a new recording needs no edits. If the script changes, delete audio/words_ref.json (the next take becomes the reference) and re-time shots.js.

Steps

1. Script. About 110 to 120 words for 38 to 40 seconds, one idea per sentence: a "what if" hook, the common belief, a turn on "But", examples that build, one shot that shows the mechanism, and a payoff that answers the hook. Show it to the user before animating.
2. Facts. Check every claim with sources. Hedge theories ("scientists think").
3. Folder. cp -r "$ENGINE/template" <reel>, then empty its audio/ and replace script.json (and captions.json).
4. Voice. In the folder: "$PY" "$ENGINE/tools/tts_kokoro.py" (free) or the user's own recording, or ElevenLabs with ELEVENLABS_API_KEY ("$PY" "$ENGINE/tools/tts_elevenlabs.py" audio/vo_el.mp3). Put the take at audio/vo.wav, then "$PY" "$ENGINE/tools/align.py" audio/vo.wav and read the word times in audio/words_new.json.
5. Shots. Rewrite web/shots.js: about 10 to 12 setups, most 2 to 4 s, the explainer shot longer. A new idea gets a new setup. Start the hook in action on frame 1. Keep faces and key props out of the caption band (y about 1250 to 1420) and inside Instagram's safe area (x 100 to 980, y 220 to 1500).
6. Stills. From the folder: node "$ENGINE/tools/render.js" stills 1.5,4.2,9.0 writes stills/. Look at the busiest moments and fix them before a full render.
7. Sound. Edit the cues in audio.py against the word times.
8. Build. bash "$ENGINE/make.sh" <reel> (keeps audio/vo.wav; pass kokoro, elevenlabs or a file to replace it). WORKERS=8 on a machine with more cores. AUDIO_ONLY=1 remixes the sound without re-rendering.
9. Hand over <reel>/<name>.mp4.

Style

• Captions: Poppins Bold, white with a dark outline, 1 to 2 words, lowercase, no punctuation (capText() in main.js does it; digits and % stay). No em dashes anywhere on screen.
• Bright saturated palette, ink outlines (#221c22), soft contact shadows, slow push-ins with handheld drift, whip transitions with motion blur. An impact is a burst, a shake and a flash.
• SFX sit under the voice (gain about 0.38). The music is a quiet plucked underscore, ducked under the voice, never a loud loop at the start.