Creative — “Writer’s Room Copilot” (Script Writing)

1) Executive snapshot

ICP: aspiring / professional screenwriters, showrunners, indie teams
Job-to-be-done: turn rough ideas + notes into a polished screenplay package (outline → drafts → revisions → pitch assets).
Wedge workflow: “Episode/pilot draft sprint” with structured writer’s room critique and rewrite plan.
Why an agent: multi-step iteration across roles + artifact generation + structured revision loops (not a single answer).
Autonomy: Copilot (agent proposes; user approves direction; subagents produce drafts).

North-star KPIs

time-to-first-draft (hours → <60 minutes for a usable draft)
revision cycle time (days → <1 day)
consistency score: character/plot continuity across scenes

2) Product experience & UX

UI paradigm: group chat “writer’s room” where each message is attributable to a role.
Primary UX loop: Plan → Draft → Critique → Rewrite → Package.

Core screens (minimal)

Group Chat: Producer, Story Editor, Script Supervisor, Director, Dialogue Punch-up
Artifact Panel: Outline, Beat Sheet, Draft v1/v2, Notes, Character Bible
Diff Viewer: highlight changes between drafts (scene-level diffs)
Pitch Kit Generator: logline, synopsis, episode bible, poster concepts

3) Agent design map

Skills (domain roles)

Producer (market fit, pacing, positioning)
Story Editor (structure, arcs, tension)
Script Supervisor (continuity, props/timeline)
Director (visual staging, shot intent)
Dialogue Writer (voice, comedic rhythm)
Sensitivity / Authenticity reviewer (optional)

Subagents (executors)

Researcher: check novelty, references, comps, similar works (web search)
Continuity Checker: validate character/plot consistency with rules
Screenplay Formatter: exports to industry format (Final Draft / Fountain)
Pitch Kit Builder: generates 1-pager + synopsis + bible
Poster Mockup Designer: creates style-aligned key art

Planner

decomposes requests into: research → outline → draft scenes → consistency pass → revision pass → export

4) Tool & data plane (MCP-centric)

MCP server: “Script Library” (project artifacts: bibles, outlines, drafts)
MCP server: “Media/Comps Search” (film/TV database search + citations)
MCP server: “Document Export” (Fountain / Final Draft templates)

5) Context engineering plan

Source-of-truth artifacts: character bible, timeline, outline, user notes
Working context: last 2–3 turns + relevant artifact snippets + scene being edited
Compaction: older discussions summarized into “decision log”
Isolation: subagents get narrow slices (e.g., continuity gets bible + latest draft only)

6) Evals & observability

Offline evals

rubric scoring: structure coherence, character consistency, formatting validity
regression set: “common writer’s room edits” (tighten dialogue, fix logic holes, pace act breaks)

Online metrics

user acceptance rate of revision plan
number of “undo” / reverts
continuity flags per page
cost per screenplay-page produced

7) Failure modes & mitigations

What breaks	Detect	Constrain	Prevent regression
Continuity errors	timeline validator; character bible checks	continuity gate before export	add failing drafts as replay tests
Tone drift	style classifier; dialogue voice checks	require “voice sheet” pinned	rubric eval + tone stress tests
Similarity risk	similarity/comps checks	enforce citations + “inspiration” logs	add adversarial similarity probes
Overwriting user intent	diff viewer + approval	rewrite only after plan approval	UX: plan-first; locked constraints

8) Governance posture & rollout

“Export screenplay” is a gated action (approval required)
versioned artifacts + audit trail of changes
canary to small cohort; track “regret rate” (reverts)

9) Business case + distribution loops

ROI model: hours saved per draft sprint × hourly rate + fewer rewrite rounds
Pricing: per project/month + export credits
Distribution loops

Shareable pitch kits and scripts (viral loop)
Collaboration invites (writer invites producer / editor)
Templates marketplace (community loop)

10) Where RFT helps

Consistency of character voice & scene structure across long arcs (reduces “episode drift”).
Better long-horizon planning: fewer plot holes introduced early that break later scenes (credit assignment problem).
Tool-use policy: when to research references vs write; when to ask the user for missing beats.

What you train on (signals):

Trajectories: outline → beat sheet → scene drafts → revisions → final screenplay artifact.
Graders (multi-grader):
- Format compliance (screenplay structure checks)
- Rubric voice/continuity scoring
- “Room quality” score (producer/editor/director rubric)
- Efficiency (revision count / token + tool budget)

Business metrics it can lift: