raycfu/reel-builder

/reel-builder

Wraps 4 to 8 edited clips in a chat-style reel. It is the format where a phone screen types /glass-widgets, Claude says "Editing with glass-widgets...", the clip builds out of orange dots, zooms to full screen for 2 seconds, and the reel ends on a "Comment EFFECT" grid. Everything is drawn in code, so it renders the same on any machine.

SK is this skill's folder. It reuses the Python environment of the effects skill (~/.claude/skills/effects/.venv). If that does not exist, run bash ~/.claude/skills/effects/scripts/setup.sh first. PY = ~/.claude/skills/effects/.venv/bin/python (on Windows .venv\Scripts\python.exe).

1. Collect the clips

For each effect you need two files of the same moment:

• raw: the clip before the effect (shown as the attached thumbnail in the chat)
• edited: the clip with the effect, from the effects skill (yourclip_effects.mp4)

Make each edited clip at least 4 seconds long and render it a little longer than needed, so the effect never fades out on screen. Vertical 1080x1920 is ideal; anything else is scaled to fit.

2. Pick the moment for each effect

Every effect shows its clip for T_FULL + full_seconds (2.2 s in the chat card, then full_seconds full screen, default 2). offset is where in the edited clip that window starts. Choose it so the full-screen part lands on the payoff: the last checklist tick, the graph hitting its number, the person pointing at the sign. Check with stills (step 4).

3. Write the config

Copy SK/examples/config.json next to the clips and edit it. Ask the user for:

• their first name for the greeting ("greeting": "Evening, <name>")
• the comment keyword for the CTA (cta_word, e.g. EFFECT) and the pill line under it (cta_pill)
• which effect opens the reel (cold_open.effect, an index into effects) and its three words (cold_open.lines)
• up to 8 clips for the end grid (grid, each with a label, a clip and the time of a good frame)
• optional more_label like +43 more effects

Keys you can set: output, full_seconds, end_seconds, cold_open (effect, from, seconds, lines; leave it out for no cold open), effects[] (name, desc, raw, edited, offset), greeting, end_command, end_desc, cta_word, cta_pill, more_label, reply_line1, reply_line2, grid_title, grid_subtitle, grid[] (label, clip, t). Paths are relative to the config file.

4. Check, then render

1. "PY" "SK/scripts/reel.py" config.json --info prints the length.
2. "PY" "SK/scripts/reel.py" config.json --test 3.0 5.5 9.0 ... saves check_XX.XX.jpg stills next to the config. Check the start and the end of every full-screen window (cycle k starts at cold_open.seconds + k * 4.2, full screen runs from +2.2 to +4.2). Look at them yourself. Adjust offset until every window ends on its payoff.
3. "PY" "SK/scripts/reel.py" config.json renders three files: the reel with typing sounds, a silent copy and the typing sounds as a WAV, so the user can mix their own music. Use --crf 5 for the highest quality, --jobs N to use more CPU cores.

Rules

• No brand logos. The working indicator is an original spinning ring, on purpose. Do not draw or paste another company's logo into the reel, even if the user supplies the file.
• No burned-in captions over the full-screen clips. Let the effect carry the moment.
• Keep the typing sounds as the only audio unless the user asks for music.