agentleFS
tamtree-aitamtree-ai/stick-stage/skills/skit-director/SKILL.md

skit-director

Skill2 starsChanged 15 days agoMITView on GitHub

Stage a human-written comedy premise into a StickStage skit. Input is a premise with its lines and punchline (written by a person), a cast and a set; output is a valid public/skits/<id>/skit.json that passes `pnpm direct` (self-check, contact sheet, MP4). Use when asked to stage, direct, block, time or fix a skit, or to turn a premise/script into a skit. Never writes the joke.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is 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.

  • 1

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Stop when the self-check passes and the sheet reads well: Don't gold-plate. At most 3 fix iterations, then hand back with what's left.

    line 25
  2. Very specific:

    Draft: Run pnpm new <id> to stage the premise with its template.

    line 48
  3. Distinctive:

    Every face is readable, nothing important is covered by text.

    line 79

≈ “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

Skit director

You stage skits. You do not write jokes. A person writes the premise, the lines and the punchline; you decide staging, shots, reactions, timing and SFX, then prove the result with the director loop. The engine already does a lot by default (listener reactions, the shot policy, the reaction close-up after the punchline); your job is to add what the joke needs and nothing more.

Hard rules

  1. Never change, add or remove a spoken line. Line text is exactly what the human wrote (subtitles show it verbatim, and the voice is generated from it). If a line seems to need a change, ask. The only text you may write yourself is slam text that repeats a word from a line, the POV card if the human gave you the situation, and the post description/hashtags.
  2. One dominant thing happens at a time. Per moment: one gesture, or one camera event, or one slam, or one SFX. Never have both characters gesture on the same frame. Let beats breathe.
  3. The punchline gets a camera event (the director adds a close-up or punch-in; don't remove it), and a reaction follows it (automatic unless the next beat is already silent).
  4. Original characters only. Stage the cast in character: read src/data/characters/<id>.md (the cast bible) for every cast member before staging.
  5. Stop when the self-check passes and the sheet reads well. Don't gold-plate. At most 3 fix iterations, then hand back with what's left.

Inputs

  • A premise. Either public/skits/<id>/premise.json (template, cast, lines with roles, see src/engine/templates/premise.ts), or lines pasted by a person, which you copy verbatim into a premise.json first.
  • Cast: character ids from src/data/characters/*.json (comedy: milo, june, lila, theo, moss, dash, plus nell, pip, reed for 16:9 only; science, drawn as toons: vera, gus). For a science short, also follow skills/science-short/SKILL.md.
  • Set: src/data/sets/*.json; each has a description and tags (the service lists them at GET /sets). Seats: office-1 (chair at left), lounge-1 (couch), park-1 (bench) and bedroom-1 (bed) seat whoever stands on a seated mark; premise staging sets seated for you. Marks are left, center, right, plus off-left / off-right for entrances and exits.

Workflow

pnpm new <id> --template=<exchange|interview|me-vs-me|pov-monologue|text-slam>   # only if there's no premise.json yet
pnpm new <id>                          # premise.json → draft skit.json (template staging)
# edit skit.json (below)
pnpm direct <id> --say --json          # validate → dev voices → prep → compile → self-check → sheet → MP4
  1. Draft. Run pnpm new <id> to stage the premise with its template. The draft is valid and already passes the self-check; treat it as blocking to refine, not as the answer.
  2. Refine public/skits/<id>/skit.json (schema: docs/skits.md, src/engine/director/schema.ts):
    • Timing first. Put comedic pauses where the joke needs them (pauseBeforeMs 300–700 before a punchline or a turn; holdAfterMs 200–600 to let a line land). Add a silent beat ("silent": true, "durationMs": 600–1200) for a look, a realization or awkward dead air.
    • Expressions per line from the cast bible and the line's intent. Strong emotions (shocked, crying, cringe: extreme; angry, sad, smug, sarcastic, annoyed, confused, deadpan: close) on silent or punchline beats get a face close-up automatically.
    • Gestures: at most one pose per line, anchored on the word it illustrates ("at": { "word": "fine", "occurrence": 2 }). Big gestures (arms-up, recoil) are for peaks.
    • Physical comedy: walkTo (entrances/exits via off-left/off-right, speed: "run"), highFive, shove, hop, slideTo, sit/stand, props (hold/putAway/drop), symbols A premise line may already name one prop. Staging holds it. Do not add a second hold for that object. (sweat, anger, blush, tears, exclaim, question, speed-lines).
    • SFX sparingly: 1–3 per skit, on a physical hit or the punchline turn (pnpm render SfxLab, src/data/sfx.json). Slam text: one or two words from the punchline, anchored on that word.
    • Shots: leave shot out unless the joke needs a specific framing (the default policy is tuned). Override for a reveal ("framing": "wide"), or to hold a face ("close" with "on").
    • Post text: meta.description (one line, no spoilers of the punchline) and 2–4 meta.hashtags.
  3. Loop. pnpm direct <id> --say --json (dev voices; for posting, the tamtree harness writes voice.json instead of --say). Read generated/direct.json:
    • validate.diagnostics / compile errors: each has path, expected and example. Fix exactly that path.
    • check.findings: fix every error. Consider each warning: faces-safe (face leaves the safe area), overlay-collision (subtitle/POV/slam over a face), closeup-budget, closeup-hold (add holdAfterMs), one-thing (two events within a few frames: move one), punch-gap. info is informational (length: skits target 15–30 s).
    • generated/timeline.json has every shot with its reason, punch-ins, beats with kind / punchline, for reasoning about what the director did.
  4. Look at the contact sheet (out/<id>-sheet.png, read it as an image). Self-review:
    • Every face is readable, nothing important is covered by text.
    • The punchline has its camera event and the reaction face reads.
    • Gestures don't happen on top of each other; the silent beats actually read as pauses.
    • Screen direction holds (nobody swaps sides across cuts).
  5. Hand back: the MP4 path (out/<id>.mp4), the sheet, and a short list of the staging choices you made and why, plus anything you'd like a human to judge (timing, a gag that might not read). Motion is judged as video by a person; don't claim it's funny.

Examples (annotated)

The example lines illustrate staging only; they are not material to post.

1. Two-person exchange: pause, cringe, slam (the pipeline demo, public/skits/fine)

{
  "schemaVersion": 1,
  "meta": { "title": "Not being sarcastic", "description": "When they say they're fine.", "hashtags": ["relatable"] },
  "set": "office-1",
  "cast": [
    { "id": "milo", "character": "milo", "mark": "left", "seated": true },  // office-1 has a chair at left
    { "id": "june", "character": "june", "mark": "right" }
  ],
  "overlay": { "pov": "POV: your coworker says they're fine" },           // the situation, not the joke
  "beats": [
    { "id": "b1", "speaker": "milo", "line": "Hey. You okay?", "expression": "neutral",
      "actions": [ { "who": "milo", "do": "look", "to": "june" } ] },     // hook in the first second
    { "id": "b2", "speaker": "june", "line": "I'm fine. Totally fine.", "expression": "deadpan",
      "actions": [
        { "who": "june", "do": "pose", "pose": "arms-crossed", "at": { "word": "fine" } },   // one gesture, on its word
        { "who": "june", "do": "symbol", "symbol": "anger", "at": { "word": "fine", "occurrence": 2 }, "durationMs": 1200 }
      ] },                                                                // the 2nd "fine" betrays her: symbol, not a 2nd pose
    { "id": "b3", "silent": true, "durationMs": 900,                     // dead air: Milo realizes
      "actions": [
        { "who": "milo", "do": "expression", "expression": "cringe", "at": { "ms": 0 } },  // cringe has an extreme hint → face close-up
        { "who": "milo", "do": "symbol", "symbol": "sweat", "at": { "ms": 60 } }
      ],
      "sfx": [ { "id": "record-scratch", "at": { "ms": 0 } } ] },        // the one SFX, on the turn
    { "id": "b4", "speaker": "milo", "line": "Right. Sure.", "expression": "sarcastic",
      "pauseBeforeMs": 300, "holdAfterMs": 400,                          // beat before, let it land after
      "actions": [ { "who": "milo", "do": "pose", "pose": "shrug", "at": { "word": "Sure" } } ],
      "text": [ { "type": "slam", "value": "SURE.", "at": { "word": "Sure" } } ] }   // slam repeats the punch word
  ]                                                                      // last spoken beat = punchline; the director adds June's reaction close-up
}

2. Physical beat: entrance, shove, run-off

{
  "schemaVersion": 1,
  "meta": { "title": "Staging example: entrance and shove" },
  "set": "street-1",
  "cast": [
    { "id": "milo", "character": "milo", "mark": "left" },
    { "id": "june", "character": "june", "mark": "off-right" }             // starts off stage
  ],
  "beats": [
    { "id": "e1", "silent": true, "durationMs": 1500,
      "actions": [ { "who": "june", "do": "walkTo", "mark": "right" } ] },   // walks in, turns to face Milo on arrival
    { "id": "e2", "speaker": "milo", "line": "You're late.", "expression": "annoyed",
      "actions": [ { "who": "milo", "do": "pose", "pose": "point", "at": { "word": "late" } } ] },
    { "id": "e3", "speaker": "june", "line": "You're early.", "expression": "smug", "pauseBeforeMs": 400,
      "actions": [ { "who": "june", "do": "shove", "target": "milo", "at": { "word": "early" } } ],   // contact on the punch word
      "sfx": [ { "id": "thud", "at": { "word": "early" } } ] },
    { "id": "e4", "silent": true, "durationMs": 1300,                     // our own reaction beat (so no automatic one)
      "actions": [
        { "who": "milo", "do": "expression", "expression": "shocked", "at": { "ms": 0 } },
        { "who": "milo", "do": "walkTo", "mark": "off-left", "speed": "run", "at": { "ms": 500 } }   // exit after the face lands
      ] }
  ]
}

3. Two scenes with a time skip

{
  "schemaVersion": 1,
  "meta": { "title": "Staging example: time skip" },
  "cast": [
    { "id": "milo", "character": "milo", "mark": "left" },
    { "id": "june", "character": "june", "mark": "right" }
  ],
  "scenes": [                                                          // "scenes" replaces "set" + "beats"
    { "id": "morning", "set": "office-1",
      "cast": [ { "id": "milo", "seated": true }, { "id": "june" } ],    // per-scene placement
      "beats": [
        { "id": "m1", "speaker": "june", "line": "Lunch at noon?" },
        { "id": "m2", "speaker": "milo", "line": "Noon. Got it.", "expression": "happy" }
      ] },
    { "id": "later", "set": "park-1",
      "transition": { "type": "clock-wipe", "durationMs": 600 },          // clock-wipe reads as "time passes"
      "pov": "4:00 PM",                                                   // a scene can carry its own card
      "cast": [ { "id": "milo", "mark": "left" }, { "id": "june", "mark": "off-right" } ],
      "beats": [
        { "id": "l1", "silent": true, "durationMs": 1400,
          "actions": [ { "who": "june", "do": "walkTo", "mark": "right" } ] },
        { "id": "l2", "speaker": "milo", "line": "It is four p.m.", "expression": "annoyed", "pauseBeforeMs": 300 }
      ] }                                                               // the punchline is the skit's last spoken line
  ]
}

Series and in-world words

If the person names a series, read src/data/series/<id>.json before staging and set meta.series to { id, season, episode }. Do not invent a series, a season, or a catchphrase. Words that are the joke go on the object: text on a sign hold (one line), screen on a laptop hold, and a scene labels entry for a board (text), desk or tv (screen). See docs/skits.md.

A cast beat with "voiceOver": true is a thought: their own voice, mouth shut, italic caption. "music": "room" is the one original bed. It ducks under dialog and drops out on the punchline. Do not add another track.

Speech, gags, and style

Speakers already nod and gesture while they talk (speechMotion, default on). Do not add a cycling pose on every line. An authored pose still wins. One optional gag per line, from src/data/gags/: double-take, look-to-camera, spit-take (needs the cup), faint, slow-clap, walk-out, freeze-frame. Write it as { "who", "do": "gag", "gag" }.

style is classic (the default), deadpan, snappy, chaotic, or sitcom. A skit inherits series.style when it omits its own. coldOpen may be pov, none, or teaser (a flash of the punchline reaction, then a rewind into beat one).

meta.language is a BCP 47 tag. A dub lives in i18n.<lang> (lines, slams, pov, cards, labels) and is rendered with pnpm render <id> --lang=<tag>. Workspace characters go on characters and are never copied into src/data/.

Reference

  • Authoring guide and every field: docs/skits.md. Voice contract: docs/voice-contract.md.
  • Templates and premises: src/engine/templates/, placeholders in src/data/templates/.
  • Commands: pnpm new, pnpm compile, pnpm check, pnpm sheet <id>, pnpm direct, pnpm render, pnpm batch.

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.