Rawi v2: how one chat message flows
Rawi now plans before it writes. A fast model reads the message and writes a short plan (the Brief). The main model then writes the reply from that plan and only the screenplay parts it needs.
AI models
Database, wallet, live updates
Writer's screen
Server and tools
Only in some cases
- SmarterPlans first, so it works on the right phase, section or scene.
- Faster and cheaperA small model plans; follow-ups like “yes” skip planning.
- ReliableIf planning fails, v1 answers. Stuck replies are cleaned up.
- Writer in controlAccept, Dismiss, Stop, Continue, Discard; see and delete memory.
- Clear costOne charge per message; cost per screenplay can be shown.
- EngagementSuggests open contests the screenplay can still enter.
1 / 4
The big picture
GET /api/v2/ai/turns/<id>/events/ for progress and the final reply. The message itself arrives at POST /api/v2/ai/chat/. Only writers with Rawi v2 turned on (beta) take this path; everyone else stays on v1.2 / 4
One message, step by step
3 / 4
Suggestions loop
| Phase | What Rawi can suggest |
|---|---|
| Start page / Info | Genre and sub-genre (including an “Other” sub-genre), logline, title |
| Characters | Several versions of character fields |
| Story outline | Content for an act or section, in every story structure |
| Scene outline | New scenes or edits, even inside a closed section |
| Script | Script pages for a scene, inserted as a [Rawi] draft |
4 / 4
Credits: hold, then settle
Writer presses StopCharged only for what was written.
Stage 2 AI errorNothing charged for the Narrator; the writer can retry.
Crash or deployA cleanup job closes the turn and releases the hold.