Chat to proposal

Work out a motive, argue about where the plot should go, test whether a fact holds, hunt for holes—do that in the chat. Discussion changes nothing: it may read, compare, and hand you three directions, but until a confirmation card appears, your draft and your library stay untouched.

1. 1
   Open a session

   Enter the chat from the agent panel on the left. A project can hold several sessions—create, rename, and collapse the list; the current session belongs to this project only.
2. 2
   State the problem

   Give the problem and its range. There is no need to paste the project in; it looks records up itself. Typing @ pulls a specific character, scene, or passage of text into this round.
3. 3
   Ask for a proposal

   Once the discussion converges, say plainly: make this a project change. The confirmation card lists before / after per item, the target objects, and the impact range—up to 8 operations at a time.
4. 4
   Confirm, adjust, or drop

   “Apply changes” takes effect and can be undone in one step; “Adjust proposal” regenerates with your note; not confirming writes nothing. A card left too long shows “Proposal expired”—“Update and continue” rereads the current state and offers a new one.

• Chapter intent, scene summaries and lore descriptions you typed by hand are never overwritten by maintenance; when it writes back, the project’s current content wins, and old inferences are not carried forward.
• It does not invent facts. Short of information it goes and looks (graph, structure stages, craft notes) or asks the one question that actually blocks execution, instead of manufacturing an answer.
• Links and references stay your call: naming an object in discussion does not connect it in the project.

• “Special”: give this round a focused capability—brainstorm, continuation ideas, pacing analysis, consistency check, plot-hole check, generate character / location / prop. It applies to this round only and clears itself after sending; until then the chip reads “This round’s special: …”.
• “Style”: pick a structured writing style for this round alone, toggling and amending dimensions without touching the project’s template.
• Images: up to 4 JPEG / PNG / WebP attachments so the agent can look at a reference or a screenshot; each can be removed before sending.
• “@”: bring in the exact object this round is about—Character, Prop, Location, World, Full script, Chapter, Scene or Scene setting—instead of describing it in words and hoping it is recognised.

Before the answer streams, one status line says what it is doing (reading chapter · 12 / composing changes / generating image); “Stop” is available throughout and keeps what already arrived. Afterwards “view process” expands what the round read and wrote.

The project graph (objects and relations, with search, read, adjacency, trace and outline), the stage list and the local craft library (general writing knowledge, not your project’s facts), dedicated views such as lore and timeline, plus whatever you @ in. Full-text reads are chunked by offset, so a growing project does not narrow the view to its first pages.

The graph version the model sees is bound to the project version at the moment the task starts. Edits made after that take effect from the next round; the material never changes mid-run.

A proposal is one pending change, not a settled fact; check the targets and the impact range before applying. When the network or session misbehaves you get plain language (unavailable, session needs rebuilding, task interrupted) while the raw codes stay in the log and the admin detail view.

Which three kinds of object it may change: What AI may write. Cross-module effects: Cross-module links.