agentleFS
calebnewtonusccalebnewtonusc/Chewbacca/skills/creative-video/SKILL.md

creative-video

Skill8 starsChanged 3 months agoMITView on GitHub

Make a short vertical video for Instagram Reels, YouTube Shorts or TikTok, or a YouTube creative, from a brief written against the craft, with the pixels drawn in code, rendered headless in Blender, or generated through Higgsfield. Use when asked to make a reel, a short, a TikTok, an Instagram video, a YouTube video, a social clip, an ad creative, a 3D animation, a Blender render, a Higgsfield or Seedance video, or content for a page. Also use when a video already made is getting no views, feels boring, or got cut off by the app's buttons. Not for a product demo of an app, which is the demo skill.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every line reads at a fifth-grade level or under.

    line 52
  2. Distinctive:

    Every new fact or mistake goes into the Higgsfield log the same turn.

    line 72
  3. Distinctive:

    The caption said "and hid", the voice said "and it hid": Check 7.

    line 100

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Make a short vertical video that follows the craft

The failure this exists to fix

Asked to master Reels with Higgsfield or Blender, the obvious move is to learn the tools. The tools are the easy half. What makes a Short work is written down by people who make them for a living, and it is almost all decided before a frame renders: a first frame that works muted, hook and foreshadow in three seconds, a mechanism pulling to the end, a cut right after the payoff. Read crafts/short-form-video.md before anything else. Every rule below points back to it.

The order

  1. Write the brief before choosing a tool. Platform, length, the hook line at a fifth-grade reading level, the foreshadow line, the last line, the mechanism the viewer watches progress, and the twist. Draw the first frame in words: what is on screen, where the text sits. If the hook would not work as a long-form title and thumbnail, rewrite it now, while it is cheap.
  2. Pick the route from references/routes.md: code-drawn, Blender headless, Higgsfield, or Blender blocking into Seedance. Pick by what done means for the picture, not by which tool was named. For Higgsfield, read his live log first: ~/dev/gavin-context/research/higgsfield/LEARNING.md. For a shot whose camera or blocking must hold, follow references/blender-to-seedance.md. For a montage of footage cut to a song, measure three references with edit-dna and cut with edit-cut, following references/cutting-to-music.md.
  3. Stage one, keyframe sheet. Render five to seven stills across the timeline at half size, tile them, and look. The first frame is judged hardest.
  4. Stage two, motion sample. Render the hardest two to four seconds and read 12 to 24 adjacent frames as one strip: pops, jumps, easing.
  5. Full render, then reel-assemble edit.json for captions, voice, bed and loudness. Voice lines come from hud-speak --say "..." --out line.wav.
  6. Gate. reel-check video.mp4 --sheet sheet.png must exit 0. Then read the sheet and score the ten checks below. Under 8 of 10, fix and rerun; do not deliver it.
  7. Write the lesson down. A new failure goes into the worked examples here or into the craft file, the same turn.

The ten checks

  1. The first frame carries the whole hook with the sound off.
  2. Hook and foreshadow land within three seconds.
  3. Every line reads at a fifth-grade level or under.
  4. A mechanism shows progress toward the end.
  5. The ending twists the expectation it set.
  6. The cut comes within about a second of the payoff.
  7. Captions match the voice word for word and sit outside the shaded zones.
  8. The subject sits outside the shaded zones at the payoff.
  9. Speech recognition on the final mix recovers the script (whisper-cli -m ~/.bob/whisper/ggml-large-v3-turbo-q5_0.bin -f mix.wav).
  10. The disclosure decision is made and written next to the file.

Never

  • Never post, schedule or upload. Publishing goes out under his name; the deliverable is a file and a note of where it is.
  • Never generate a real person's face or voice without that person's consent, and never a public figure's at all.
  • Never put a number on screen the render does not make true. "1,000 dominoes" is the field's count, checked in timings.json.
  • Never run a Higgsfield job without higgsfield generate cost first and the number said out loud, and never buy a plan or credits on his behalf. Every new fact or mistake goes into the Higgsfield log the same turn.
  • Never repost footage he did not make, and never export with another app's watermark. Instagram stops recommending both.
  • Never open Blender's window. Headless only; he works on this Mac.

Worked example: the domino reveal, 2026-09-27

blender/domino_reveal.py, 1,000 dominoes whose falling wave reveals a hidden picture (--mask any.png, default a heart), and blender/domino_clicks.py, one synthesized click per impact time. 9.6 s, 1080x1920, file at ~/Movies/chewbacca-creatives/2026-09-27-domino-reveal.mp4. Scored 8 of 10: hook and foreshadow took 3.6 s, and the heart is the expected picture, with no twist.

What broke on the way, so it does not break again:

  • Keyframe sheets one to three opened on a black wall. Below about 61 degrees every row hides the gap behind it (tan e > height / pitch). For any repeated array, work out the angle where the gaps show before placing the camera.
  • The picture showed before the fall. A domino exposes the face on the side it falls away from. Colour the far face and look from the near side.
  • The white faces vanished into a white floor. The floor went a step darker than the faces.
  • The hook card read "My AI set up" alone. Split captions are for subtitles, not the hook: the first card carries the whole promise.
  • The end push-in drifted the heart into the like-button rail, and the last two seconds sat after the payoff. Cutting at 9.6 s fixed both.
  • The caption said "and hid", the voice said "and it hid". Check 7.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.