agentleFS
tamtree-aitamtree-ai/stick-stage/skills/science-short/SKILL.md

science-short

Skill2 starsChanged 15 days agoMITView on GitHub

Make or fix a science/maths short for the science channel: a myth-flip dialog with Vera (host) and Gus (skeptic) as toons, figures that change on the spoken word, claims with sources, and the owner's approval gate. Covers the premise, figures, hook card, sets, staging, render and the hand-back. Use when asked for a science video, an explainer short, a myth-busting short, or to restage/fix fine-print, coin-due, old-light or slow-electrons. Never approves a skit or marks a claim as checked.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Don't rebuild a hand-staged skit: If skit.json was edited by hand (fine-print is), never run pnpm new <id> --force.

    line 32
  2. Very specific:

    spoken: when the caption and the voice differ ("5%" vs "five percent", equations).

    line 62
  3. Distinctive:

    Add a missing word there instead of misspelling the line.

    line 64
  4. Very specific:

    Every line gets one change on screen, anchored on the word it illustrates ("at": { "word": "matter" }).

    line 69
  5. Distinctive:

    Keep the screen changing every 3–5 s, one pause at most (≤ 0.6 s) and no dead air.

    line 84

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Science short

A science short is a ~30 s myth-flip: Gus states something most people believe, Vera shows what really happens with a figure or prop, says why, says why the myth felt true, and lands a takeaway the viewer can repeat. Read these first:

  • docs/science-channel.md: format, roles, steps, accuracy, the look
  • docs/figures.md: every figure kind, its params and cues
  • src/data/characters/vera.md, gus.md: the cast bibles
  • skills/skit-director/SKILL.md: the general staging rules (one thing at a time, pauses, the loop). They all apply here.

Reference skit: public/skits/fine-print/ (hand-staged to a storyboard; the storyboard and its retention reasons are in ../planning/2026-10-02-toon-characters/plan.md). Channel plan and gate: ../planning/2026-09-29-science-channel/.

Hard rules

  1. Never approve, never mark a claim checked. checkedBy, approval and --by= belong to the owner. You may run pnpm approve <id> --check to report status.
  2. Every number or fact on screen or in a line has a claim with a real source. A spoken number with no claim is a claim-missing warning: fix it, don't ignore it. If you wrote a source from memory, say so in the hand-back so the owner checks it first.
  3. Say where the picture breaks in simplifications (it goes in the pinned comment).
  4. No AI images. Images are NASA/ESA (or similarly licensed) inserts with credits (src/data/images.json). Everything else is drawn.
  5. Don't rebuild a hand-staged skit. If skit.json was edited by hand (fine-print is), never run pnpm new <id> --force. Edit skit.json directly.
  6. Any edit after approval clears it. The owner approves again.

Steps

pnpm new <id> --template=myth-flip     # premise.json from src/data/templates/myth-flip.json
# write lines (with roles), claims, simplifications and figures in premise.json
pnpm new <id>                          # stage it into skit.json
# refine skit.json (below)
pnpm voice:say <id>                    # dev voices; the harness writes the real voice.json
pnpm compile <id>                      # diagnostics, shots and why, length
pnpm render <id>                       # out/<id>.mp4 (draft renders are not gated)
pnpm approve <id> --check              # what the owner still has to check

The owner then checks the sources, runs pnpm approve <id> --by=<name>, and posts with pnpm render <id> --post.

Writing the premise

  • Roles in order: myth, pushback, prediction, demo, reaction, why, why-it-felt-true, takeaway, optional loop. A myth-flip needs myth and takeaway.
  • Hook in 0–2 s. The premise is clear by line 2, and the first laugh lands by ~10 s.
  • The surprise is on screen at frame 0. The hook card (or overlay.pov, ≤ 34 characters) says it in one glance ("A year with 400 days"); a vague teaser loses the muted swipe.
  • Title: 25-50 characters, the fact stated, never a Why/How/Is…? question. One number a viewer can picture, or none: no ms/micro-/millionths, no unrounded big numbers. Every number on screen and in the title has its claim.
  • spoken when the caption and the voice differ ("5%" vs "five percent", equations). Names and jargon get pronounce hints on the voice request from src/data/pronunciations.json. Add a missing word there instead of misspelling the line.
  • Length: aim for 30 s with the real voice. Dev voices run long; cut filler words before cutting a beat.

Staging: one visual event per beat

Every line gets one change on screen, anchored on the word it illustrates ("at": { "word": "matter" }). Tools, all used in fine-print:

NeedUse
Muted viewers know the premise at frame 0overlay.hook: a short card that leaves when line 1 ends
A number grows on its worda figure set cue with durationMs (tweens)
Shares of a whole (they add to 100)progress, not histogram; readout, ask ("?"), dim, here chip
An object the line namesgalaxy, orbit, wave, particles, image, … (docs/figures.md)
An impactburst figure + one SFX (thud, pop, zap)
An inset that survives a punch-infigure layer: "screen"
A prop that changes stateswap props with a second hold (book-open → book); start state via cast holding
The first laughthe reaction line: close shot on the skeptic, double-take gag, exclaim symbol
The punchlineslam text repeating the punch word; the director adds the punch-in and the reaction

Keep the screen changing every 3–5 s, one pause at most (≤ 0.6 s) and no dead air.

Sets: space-2 (Nebula night, dark sky; the palette's rim keeps the cast readable), space-1, lab-1, void-1, and the themed science-dawn / -tide / -cosmos / -lab. blueprint-1 is retired. Props: categories maths and science (docs/skits.md, prop table).

Checking it

  1. pnpm compile <id>: no errors. Every warning other than approval-unapproved is fixed or explained.
  2. Render, then look at frames at the hook (frame 0–15), the first laugh, and the punchline. To get stills from the MP4: ffmpeg -i out/<id>.mp4 -vf "select='eq(n\,0)+eq(n\,330)',scale=540:-1,tile=2x1" -frames:v 1 out/check.png.
  3. Text never covers a face; captions, the hook card and figures don't overlap.
  4. pnpm test && pnpm typecheck && pnpm lint if you changed anything outside public/skits/.

Hand back

The MP4 path, the length, the list of claims with the ones whose sources you wrote from memory marked, the simplifications, and what the owner should judge: does the hook read muted, does the first laugh land by ~11 s, and is it accurate? Say plainly that it is not approved.

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.