agentleFS
nocodehacknocodehack/make-short/SKILL.md

make-short

Skill71 starsChanged 13 days agoMITView on GitHub

Make an animated YouTube Short end to end (9:16, bean characters in a painted Blender world) from one idea. Claude measures a reference clip, writes a paradox story, records or generates the voice, builds the shot list and acting, checks stills by numbers AND by eye, renders natively, mixes sound, then an independent critic scores it and one director fixes each flagged beat. Supports a target length (for example a 15 s quick run). Never uploads without the user's explicit go. Use for "/make-short <idea>", "make an animated short about ...", "turn this into a Short".

  • Reads credentials
  • Installs packages

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. hassancs91/claude-faceless-shorts-creator★ 201same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 1 is distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No false claims: never write "1 prompt", "one prompt", "one click", "fully automatic".

    line 12
  2. Very specific:

    Read references/lessons.md (binding) before step 4 and before every fix round.

    line 16
  3. Very specific:

    Target length: ask once; default 30-60 s.

    line 23
  4. Very specific:

    Write the targets to critic/targets.md (one screen).

    line 30
  5. Very specific:

    Write story.md (layout: example-60s/story.md.

    line 33
  6. Very specific:

    Write CONTRACT.md (model: references/contract-example.md): length in frames, sets (camp, insert, night, tent ship in lib/sets.py), …

    line 49

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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-short (v3: Blender engine)

SKILL = this skill's folder ("Base directory for this skill"). P = the user's project folder, a copy of SKILL\v3 (step 0). Paths below are relative to P unless they start with SKILL. Install problems: SKILL\INSTALL.md. What to test after install: SKILL\CHECKLIST.md. The older 2D Chrome route (no Blender) is references/legacy-2d-route.md.

Rules, always

  • Nothing is uploaded or posted unless the user types "go" (step 9). Never print or ask for API keys; they live in the user's environment only.
  • No false claims: never write "1 prompt", "one prompt", "one click", "fully automatic". True: the user approved the story, and Claude wrote the shots, acting and code.
  • Generic IP only: original characters and world.
  • LOOK at every still and strip yourself (read the PNG). Passing numbers is not passing (lesson 1).
  • Never self-score. Builders rate about 2 points too high; only the independent critic scores (step 6).
  • Read references/lessons.md (binding) before step 4 and before every fix round. Append new lessons after every round.
  • Quote ETAs from the clock (Get-Date), at 1.5x your first guess.
  • At most 5 short lines to the user between steps.
  • go.sh loads the GPU. If the user is screen-recording or editing, ask before rendering.

0. Set up the project (once per Short)

  1. Copy-Item -Recurse "SKILL\v3" "<work folder>\<slug>" (outside OneDrive, for example C:\shorts). That is P. Work inside P.
  2. Target length: ask once; default 30-60 s. Quick run: 15 s = 450 frames at 30 fps, about 5-8 shots, 3-4 narration lines, about 40 words. Everything scales: go.sh reads the length from the last f_out in shots.json.
  3. Fresh install: run CHECKLIST.md section A first.

1. Reference measure (about 10 min)

Skip if the user gives no reference clip.

  1. Extract 1 fps frames and a contact sheet of the reference with ffmpeg; read the sheet.
  2. bash measure.sh <frame files> prints SATAVG, YAVG, YLOW, YHIGH. Note saturation, luma, shots per 10 s, face height, caption style.
  3. Write the targets to critic/targets.md (one screen). gate.py has default bands per mood (BAND); change them only with a reason from the reference.

2. Story: the paradox formula (GATE 1, the user approves)

Write story.md (layout: example-60s/story.md; do NOT reuse its story):

  • Sources first: only facts you checked; list the caveats you dropped.
  • Hook = a paradox in line 1: two true things that should not both be true ("canned food existed 48 years before the can opener"). The hero and the want are in frame 1.
  • Middle = escalating failed attempts, each one visible gag per line (one noun or verb), with a hold on the reaction.
  • Payoff resolves the paradox with the real fact; the last frame loops to the first.
  • ## Narration: numbered lines, at most 3 words per second (15 s is about 40 words).
  • Show the logline, the paradox, the numbered lines and the sources. Ask "Approve or change?". Continue only on approve.

3. Voice (before the shots: the shots lock to it)

Choose with the user:

  • Free route (default): record yourself reading the narration, one calm take, save as voice/voice.mp3. Transcribe locally: whisper voice/voice.mp3 --model small --language en --word_timestamps True --output_format json --output_dir voice, then node voice/whisper-to-timing.mjs voice/voice.json. Read voice/voice-timing.json against the script. (Whisper is a separate install: pip install openai-whisper.)
  • ElevenLabs (optional, paid per character): the user sets $env:ELEVENLABS_API_KEY in their own terminal. Tell them the character count first. node voice/take.mjs <voiceId> voice writes voice/voice.mp3 and voice/voice-timing.json.
  • Sound effects: voice/sfx/*.mp3 is a starter set (crate, bend, boot, rock, pan, clang, pop, ding, night, music). New ones: node voice/sfx.mjs (same key rule), or skip.
  • Frame of a word = round(seconds*30).

4. Contract + build

  1. Write CONTRACT.md (model: references/contract-example.md): length in frames, sets (camp, insert, night, tent ship in lib/sets.py), characters (marcel, gaston, soldier), props, the numbers to hit. A story needing a NEW set, prop or character means editing lib/sets.py or lib/chars.py: tell the user and budget the time.
  2. Beat list first: one narration line = one beat = one readable frame (line, noun/verb, impact word, frame).
  3. Write shots.json and acting.json from the beat list (schema: example-60s/, shots/NOTES.md; shots/gen.js is the generator the 60 s film used). Then node shots/validate.js. Always: faces 30-40% of frame in character shots; react/squash/lean on every line; a 12-24 frame reaction hold after each gag; whips about 6 frames; no puff or blur on a gag frame.
  4. Team (30 s and up): one subagent per frame range. Each reads references/beats-brief.md, writes only beats/<name>.json, and reports the exact impact frame. For a 15 s run, build it yourself.

5. Gate: numbers AND look

  1. bash gate.sh renders one still per shot; gate.py prints PASS/FAIL per shot (saturation/luma bands, blown whites, crops). Fix FAILs; tune.sh tries grade overrides.
  2. bash contact.sh 1 24 makes gate/contact_1.png. Open it and look. Per shot: gag object visible and big, faces readable, nothing cropped, no giant foreground clutter, characters not overlapping, the mood colour right.
  3. At most 3 gate rounds. Show the user the sheet; render only after they say "render".

6. Render, mix, independent critic

  1. bash go.sh r1 (native 1080x1920, 3 processes; about 21 min for 60 s on the author's GPU, so roughly 5-6 min for 15 s; give a clock-time ETA). Output out/rounds/r1-silent.mp4.
  2. Mix: write voice/cues.json from the impact frames (format in its _note), then node voice/finish.mjs. Output out/short-v3-final.mp4; copy it to out/rounds/r1-final.mp4.
  3. Critic = a fresh subagent that did not build the film. Give it the final mp4, the reference clip, critic/RUBRIC-blank.md and your suspicions to verify. It makes frames and cmp sheets, fills the rubric, writes critic/critique-r1.md with 3 fixes by FRAME RANGE. Verify each claim at full resolution before acting (lesson 28).
  4. Pass = criteria 1-4 each >= 7 and mean >= 8. The first Short took 12 renders; a 15 s one should need far fewer, but time the real run and report the real number.

7. One director per flagged beat

Per flagged frame range, one subagent: reads references/beats-brief.md, references/lessons.md, the critique and the stills; writes only beats/<range>.json; re-renders only its range (bash render.sh A B, output in render/); makes a strip; LOOKS at it twice; reports the impact frame. Then redo steps 5-6 for that range. Stop at the pass bar or when the user says stop. Never "ship the best we have, call it a first try".

8. Lessons

Append to references/lessons.md (numbered, one line: what the critic found, what fixed it). Copy new lessons back into SKILL's references/lessons.md.

9. Upload (manual by default)

Hand over out/short-v3-final.mp4. Titles start with a verb; the description has the fact, its source and #Shorts. The user uploads by hand. Auto-post only if the user asks: copy the final to out/<slug>-9x16.mp4, then node "SKILL\scripts\post.mjs" "P" --platforms youtube is a dry run; --go only after the user types "go" in this conversation. It needs their own $env:BLOTATO_API_KEY (generating it starts a paid plan). Private by default; --public only if told. Never handle the key.

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.