Script Compose Agent Skill

Run only on explicit user intent, never on file drop. Dropped text/PDF already exists as a note (data.body) and mirror (./assets/notes/<note_id>.md).

Defaults: a 30s beat is one moment; match the input language; with characters, prefer meaningful dialogue and let narration support rather than carry the scene, but let wordless action carry a beat instead of packing every one with dialogue.

Stop at script capture, shot notes, and anchor extraction. Route multi-stage work back to story-to-video-workflow.

Capture target duration when observable:

• Explicit user duration wins ("30 seconds", "2-minute short").
• Timestamp blocks come next; sum them.
• Otherwise estimate roughly and mark it as an estimate.

Store target_duration_sec and duration_basis when known. If implied runtime is >~3 minutes, call out scope before shot/video planning.

1. Triage → Capture

Classify the input, then capture as in §2. Never skip straight to §3.

• Screenplay (INT./EXT. + ALL-CAPS cues + dialogue) → use verbatim. For dropped text/PDF, read workflow.json data.body or ./assets/notes/<note_id>.md. Pick a 2–5 word title; identify duration basis; do not rewrite to fit.
• Story / concept (prose, pitch, logline) → sketch ONE paragraph back (setting, characters, conflict, target duration) and ask if it's the shape. Iterate. On "yes/go", rewrite using the rules below, then capture.
• Neither → don't run; defer to image-compose / video-compose.

Torn between screenplay and story? Prefer screenplay — safer than rewriting.

Rewrite rules (story → screenplay):

• Format: INT./EXT. LOCATION - TIME slug, present-tense action, ALL-CAPS cue + dialogue. No scene numbering. No camera directions (that's video-compose).
• Preserve user-quoted dialogue verbatim.
• With characters, include dialogue that reveals motive/conflict/relationship; avoid narration-only exposition unless VO-driven.
• Pace speech at ~2.2-2.5 words/sec plus reaction/action room.
• Duration: match if stated; default 30–45s. Don't overshoot.
• Short input, longer target? Keep verbatim and ask "reads as ~Ns; extend?" — don't silently pad.

2. Capture — canvas note + title

ONE note. No split. Canvas writes go through the mutator, never direct workflow.json.

• read ./workflow.json (read-only inspection — see if title is already set).
• Append the script note via the mutator with subtype: "script":

   node "$PAI_REPO_ROOT/server/cli/canvas_mutate.js" \
     --op addNode \
     --payload-json '{"node":{"type":"note","data":{"subtype":"script","label":"Script: <title>","body":"<full screenplay verbatim>","metadata":{"author":"agent","timestamp":"<ISO>","target_duration_sec":45,"duration_basis":"estimated from script length"}}}}'

Omit target_duration_sec / duration_basis only when there is no defensible signal.

Stdout returns assigned.node_id — keep it for §3 (shots derive from this id).

• Set the workflow title if empty:

   node "$PAI_REPO_ROOT/server/cli/canvas_mutate.js" --op setTitle --payload-json '{"title":"<title>"}'

• Confirm with Captured., then offer the next step as a choice rendered per the project PROJECT_AGENT.md § "Recommendation and choice shape". Recommended option: "Split it into <=15s shots and extract characters/locations/voices." Plus an escape to do something else.

STOP. Do NOT proceed to §3 without an explicit user command.

3. Analyze — on explicit user command

Triggers (judge intent): "split into shots / clips", "break this up", "pull the characters / locations", "who's in this", "analyze this script", "design the characters from this script".

Not triggers: "what's in this", "summarize", "tell me about it" — those are read-and-reply.

When triggered:

• Slug — kebab-case of the working title. Collision → suffix -2, -3.
• Shot splits (≤15s each): use metadata.target_duration_sec or estimate. Split on natural beats (slug/dialogue/location/time/appearance changes). For >15s material, keep resulting shots as close to 15s as natural (default ≈ ceil(total_seconds / 15) shots); split shorter only for hard cuts, dialogue turns, continuity shifts, or strong beats — don't over-fragment just because the script's own time markers say so. Pace speech at ~2.2-2.5 words/sec plus reaction/action room; silent action ~3–5s. If dialogue cannot fit naturally, split it; reduce only when the user asked for compression. Never rewrite — shot bodies are verbatim slices. Each shot note has subtype: "shot". Build one addBatch with N shot notes + N derived edges:

   node "$PAI_REPO_ROOT/server/cli/canvas_mutate.js" \
     --op addBatch \
     --payload-json '{
       "nodes": [
         {"type":"note","data":{"subtype":"shot","label":"Shot 1 (0–15s)","body":"<slice>","metadata":{"author":"agent","timestamp":"<ISO>"}}},
         {"type":"note","data":{"subtype":"shot","label":"Shot 2 (15–30s)","body":"<slice>","metadata":{"author":"agent","timestamp":"<ISO>"}}}
       ],
       "edges": [
         {"from":"<script_note_id>","to":"$0","kind":"derived"},
         {"from":"<script_note_id>","to":"$1","kind":"derived"}
       ]
     }'

$N placeholders are 0-indexed positions in nodes; the mutator resolves them to the assigned ids after running. Reply's assigned.node_ids is the array of shot ids in the same order.

• Anchor extraction — from the shot bodies, extract only downstream needs:
• Characters: recurring/visually important people/entities. Include one-line base visuals only when given.
• Variants: same character with materially different on-screen look by scene/shot: age jump, costume change, injury, disguise, transformation, wet/dirty/bloodied state if it must persist across shots. Do not create variants for transient expressions or tiny props.
• Locations: distinct settings plus same-setting variants when framing/scale, time, weather, lighting, dressing, story state, or close/detail coverage matters.
• Voices: every speaking character and narration/V.O.; preserve speaker labels and dialogue language.
• Missing anchors: first character, variant, location, or voice that blocks rendering Shot 1.
• Parse offer — ONE compact planning line plus a soft next step:

> Plan check: ~<seconds>s, <shots> shots, <N> character(s), <V> variant(s), <M> location(s), <S> voice need(s). Missing: <first blocker>.

If N>0, V>0, M>0, or S>0, offer next step with project choice shape. Recommended: "Design the character/location anchors, then voices." Anchors include base/variant character sheets, detailed location/location variants, and speaker/VO voice anchors.

On approval, route to image-compose first (base character sheets, needed character variants, and location stills) with --source-node-id <script_note_id> so the new nodes wire back to the script. After image anchors land, route speaking/narration needs to voice-compose. Don't generate inside script-compose. Skip the offer if every count is 0.

If the user's command was narrower ("just the shots", "only characters"), do only that sub-step and skip the offer.

4. Revisions

Surgical (title still fits): update script-note body + affected shot bodies in place. Use the mutator's updateNode op (one call per node, or batched via updateBatch).

Structural (title no longer fits): new script note (addNode); old→new edge addEdge with kind:"derived"; new shot family via addBatch against the new script note. Leave old shots; delete only if asked (deleteNode cascades edges for you).