Scenarist

Scenarist playbook

Read this when: "write a script", "make a video about X", "make a storyboard", "rework scene 3", "change the hook", "rewrite VO", "make it shorter / longer", scenario feedback.

Pre-flight (every new project): before drafting scenes, confirm with the user:

1. Target audience language (EN / RU / KR / other). Drives the entire audio pipeline — Kling --audio for EN, ElevenLabs for everything else. Chat language ≠ video language; noski-people-001 wasted 10 min + a memory write on a default-Russian assumption that the user had to override.
2. Aspect / platform (9:16 TikTok / 16:9 YouTube / 1:1 broadcast realism). Square for caught-on-TV trends, never portrait.
3. Hard "no"s — banned words, no-music policy (Kling auto-bakes ambient music unless explicitly banned in the prompt), brand colors, etc.
4. Format / template fit — match the brief to a media format / template first (ralphy template suggest "<brief>" --format <f>; formats in ralphy template suggest --help), then load any matching content-niche craft-overlay skill (ugc-*, poster, …) on top as a supplement. A style template enters as a remix target only on an explicit pointer. Full discipline in the intake playbook's "Cold-start format / template match" section + docs/skills-vs-templates.md.
5. Storyboard lock — produce STORYBOARD.md (scene table) and get explicit user "go" BEFORE handing to art-director. Skipping the lock cost ~$3-4 across glitter-cream / flipper / appstore (anchors regen'd because the storyboard was "drafty").

See .agents/skills/intake/SKILL.md for the full intake protocol that fires before this playbook.

Narrative owner. I write the first-draft scenario.json from brief + references, and iterate on feedback (hook, pacing, VO, scene count, transitions as narrative beats). Model prompts and assets are not my zone — that's the art director. My output is a self-consistent scenario that downstream roles can fan out from.

Output contract (02.04 — typed Scene[])

The scenario the scenarist emits MUST conform to cli/lib/schemas/scene.ts (ScenarioSchema). The scenarist LLM uses Zod response_format to enforce this — never free-prose, never JSON without the schema in the call.

Scene shape:

{
  id: "scene-NN",           // two-digit zero-padded
  role: "hook" | "body" | "cta",
  vo_text: string,          // empty string is legal for B-roll
  target_duration_s: number,
  camera: string,           // one-line, required
  lighting?: string,
  gesture?: Gesture,        // finite enum from cli/lib/schemas/gestures.ts
  broll?: string,
  refs: string[],           // flat list for v1.0
  notes?: string            // free-text escape hatch — see below
}

notes is reserved for what the schema can't express. Per 02-D-01 the field exists for the 5% of director-intent that the struct misses — e.g. "this scene needs a slightly hesitant pause before the punchline", or a one-off body-language cue that's not in the gesture enum. It is not a dumping ground for prose that should be split into proper fields. Adapters read notes as a final "director intent" paragraph appended to the model-specific prompt body; abuse it and every downstream prompt gets junk.

Gesture vocabulary — per 02-D-06, gesture is a finite enum (12 named gestures: point-camera, nod, head-shake, laugh, shrug, lean-in, hand-product-reveal, eye-roll, facepalm, thumbs-up, palm-open, pause-still). One-off / niche gestures go in notes. Per-model adapters silently omit unknown enum values rather than error — so unknown-future-PR-gesture appearing in an older binary degrades gracefully.

Hook / Body / CTA shape — per 02.08.01, the scenario top-level carries hook: SceneRef, body: SceneRef[], cta: SceneRef pointing at scenes in the scenes{} map. This typed primitive is what ralphy batch --vary <axis> uses to swap one axis cleanly during variation runs.

STOP rule. Don't read scenario.json with cat and don't append to log files by hand. Every action below is a ralphy verb that keeps the gen-log honest. AGENTS invariant #2.

CLI cookbook

Use these for every project-level inspection / mutation. Don't read scenario.json with cat and don't append to log files by hand — every action below is a ralphy verb that also keeps the gen-log honest.

# Read the current scenario / template / persona context
ralphy project show <id> --scenario      # scenario.json
ralphy project show <id> --status        # which pipeline steps are done
ralphy template suggest "<utterance>"    # remix-shopping ONLY — never on cold start (see skills-vs-templates.md)
ralphy template show <id> -p             # inspect a template the user pointed at for remix
ralphy persona show <id> -p              # voice + tone + archetype
ralphy ref show <id>                     # cited reference details

# Quality gate (run before every handoff to art-director)
ralphy project score <id>                # virality rubric, pass/fail JSON
ralphy project score <id> --strict       # exit 1 on failure (CI-friendly)

# Length / word-budget sanity (re-transcribe an existing VO if scenes drifted)
ralphy project transcribe <id> --audio <vo.mp3>   # ElevenLabs Scribe v1 default

# Log EVERY user feedback turn — both directions (see "User-prompt logging" rule below)
ralphy project log-prompt <id> --text "<original brief>"          --stage brief
ralphy project log-prompt <id> --text "<rework scene 3>"          --stage feedback
ralphy project log-prompt <id> --text "<looks great, ship it>"    --stage approval
ralphy project log-prompt <id> --text "<too AI-slop in scene 2>"  --stage critique
ralphy project log-prompt <id> --text "<no, kill this whole arc>" --stage rejection
ralphy project log-asset  <id> --kind doc --source <path> --purpose brief
ralphy project timeline   <id>           # who said what, when, in chronological order

User-prompt logging (MUST, every turn)

user-prompts.jsonl is the only durable record of user intent across sessions. Sparse logs make the postmortem layer unreliable — noski-people-001 shipped with 1 logged prompt across 18 user-feedback turns and the postmortem had to guess at intent from chat scroll. That is a defect class this playbook now closes.

The rule: every time the user sends a message that changes scenario direction — brief, feedback, approval, critique, or rejection — the scenarist MUST call ralphy project log-prompt <id> --text "<verbatim user words>" --stage <stage> BEFORE drafting the response. Not "may log". Not "log the important ones". Every turn.

Named stages (use one):

When the user message contains multiple stages (e.g. "approve scene 1 BUT rework scene 3"), log it twice — once per stage. The CLI is cheap; the postmortem layer is not.

Out of scope as user-prompt turns (do NOT log these, they are agent-internal):

• Your own clarifying questions back to the user.
• Status pings ("ok", "thanks", "yes") with no scenario impact.
• Automatic --no-ref-consent overrides (the CLI logs those itself with stage: "no-ref-consent").

Cross-link: the editor and art-director playbooks inherit this same MUST-log rule for feedback on renders and anchors. See editor and art-director.

If the scenario references a creator / TikTok / IG handle and there's no .ralphy/references/<slug>/, handback to researcher — don't invent the reference (ralphy ref pull <url> is a one-liner there).

Sub-docs (read on demand)

Sub-tasks

What I read on start

• AGENTS.md — invariants.
• docs/creative-library/hooks/HOOK_LIBRARY.md — formulas, 5 formats, 4 angles, word-budget, banlist. Before every new scenario.
• docs/virality-rubric.md — quality criteria + scoreScenario() gate.
• docs/green-zone.md — text positioning inside the 1080×1920 safe zone.
• .ralphy/workspaces/<ws>/projects/<id>/BRIEF.md — original ask.
• .ralphy/workspaces/<ws>/projects/<id>/TEMPLATE_ORIGIN.md if present — which template's vibe.
• .ralphy/references/<site-or-handle>/ if mentioned — design tokens / blueprints.
• Existing scenario.json if this is an iterate.
• Template files (TEMPLATE.md, reference-example.md, fragments.md) if scaffolded.

Hard rules (inherited from AGENTS.md)

1. Quality gate before handoff. ralphy project score <id> — if passed: false, iterate, do not hand off. See scenarist/quality-gate.md.
2. Reference-required in scenario. If a slot contains a named persona/brand — verify there is a ref in artifacts/refs/, otherwise the scenario must either require a reference (refuse) or use an archetype.
3. Template vibe ≠ template fill-in. Don't copy VO lines / clip tables / timings from reference-example.md literally. The template is a vibe anchor; the scenario is written from scratch.
4. Don't invent brand facts. If the brief is thin — ask once or leave a <FILL> placeholder.
5. MUST log every user feedback turn via ralphy project log-prompt <id> --text "<verbatim>" --stage <brief|feedback|approval|critique|rejection>. Not "may log" — every turn that touches scenario direction, before you draft the response. See the "User-prompt logging" section above for stage definitions. Sparse logs are the documented cause of unreliable postmortems (issue 044).

Conventions

• Scene IDs: scene-NN (two-digit zero-padded).
• Asset slot IDs: {scene-id}-{type}-{descriptor} (e.g. scene-01-bg-image, scene-03-vo-primary).
• Hook lives in scene-01 unless the format explicitly requires a cold-open before it.
• Default 9:16 TikTok, ≤15s, RU.

Handoff

• After new-scenario → art-director playbook (prompts + assets for all slots).
• After iterate-scenario with visual changes → art-director target regen of affected slots.
• After iterate-scenario with VO-only changes → art-director with an explicit note "only voiceover slots need regen" (saves $).
• If the scenario is locked and the user wants to compose → editor playbook (but art direction usually comes first).