AI co-writer and the project canon

The co-writer on the project home is a room where you build the story together through conversation. You go from first ideas all the way to writing scenes in a chat, but only proposals you approve change the script and its structure. Things that are a record of the story — foreshadowing, what each character knows, established facts — are written to the project canon the moment they arrive, and each entry can be undone (see "Record-type proposals are recorded without asking" below).

In a short drama (script) project, each chat turn costs 1 credit — the badge next to the send button tells you in advance (a turn that continues automatically also counts as one turn). A turn that delivers a scene costs 50 credits per scene. If you asked for a scene and got a question instead, those 50 credits are refunded in full automatically (shown on screen). If you don't like a scene, you can ask for another proposal once per scene for free. There is no daily chat limit.

The flow — from ideas to writing

Conversational co-writer (experimental, entry point hidden) — a new path where you get drafts through free-form chat instead of cards, and apply them to an episode with Apply as is (D-306). Right now it only appears when an operator turns it on for a specific project; in those projects, the chat screen opens instead of the card flow below. A draft doesn't touch your script until you apply it.

>

Switching back — a conversational project can be switched back with Back to cards in the ⋯ menu at the top right of the chat screen (the writer in charge only); switching back keeps the chat, the script and the canon as they are. However, open questions and drafts you haven't applied don't appear on the card screen — they aren't deleted, and they come back when you turn the conversational path on again. We tell you how many there are before you switch back. You can't change the path while the co-writer is generating a reply.

>

What carries over when you turn it on — decision cards you haven't answered yet on the card path become "open questions" in the conversational path, and the scene summaries the card path remembered move to the conversational store (scenes whose text has changed are read again). Cards, the script and the approval history are not deleted. The move runs in the background right after the switch, and the switch stays even if the move fails. If you don't see your open questions, switch back to cards and turn it on again to move the same items once more (moving twice doesn't create duplicates).

1. Episode 1 readiness — above the chat, one line shows "Episode 1 ready: about N%" and what to decide next.

It looks at the logline, genre, protagonist, supporting characters, character personalities, locations, synopsis, the episode 1 incident and the episode 1 design. It's a rough estimate. For now, though, if a readiness item is empty, the co-writer asks the next question first instead of writing scenes or outlines (a step-by-step approach — a redesign is planned; see the plan document docs/cowriter-redesign-plan-2026-09-11.md). The indicator disappears once everything is filled in or a script exists.

1. Ideation chat — you refine the logline, characters and structure together. Facts settled during the chat come up as

"fact proposals", and approving them adds them to the project canon. Conversely, if you ask to drop a fact that's already there, a "Retire fact" card appears — applying it leaves that fact in the canon marked "retired", and it won't be used again in later episodes or revisions (it's ignored even if it still remains in the episode plan). Important decisions (about ten, such as the logline, genre, protagonist, supporting characters, locations, synopsis, episode 1 incident and design) are asked one per turn, in order — each time you answer, a proposal reflecting that answer arrives in the same turn. Details that aren't on that list (an episode's hook, twist, starting state and so on) aren't asked about; they're filled in as a draft and proposed as a card. If you don't like it, edit the card or ask for another proposal.

• The full episode plan comes after you set a target episode count (short dramas usually run

30–100 episodes). A plan never appears before you've set the count, and the full plan and the episode 1 outline don't pour in all at once — once you apply the plan, the outline is proposed next.

• Character names get settled too — if a character is only defined by role, like "insurance investigator" or

"younger brother", 2–4 name candidates that fit the project's tone are proposed as a choice card. Pick one or type your own, and a character card is created with that name. You can change the name later, so it's fine to settle on one and move on. Unnamed bit parts such as a secretary, reporter or doctor can be used freely by role. If the same bit part keeps appearing or grows in importance, you'll get a proposal to register them; once registered, you can create sheet and image assets for that character.

1. Choosing how to write (once, at the start) — when the story skeleton (the decisions above) is ready, a writing-mode

card arrives automatically: build the full episode plan first, set only the main arc, or start writing from episode 1 right away. On the turn where you answer the last planning decision (for example, locations), the writing-mode card arrives in the same reply as the card that applies that answer — once you've applied it, you can choose. Even when the skeleton gets filled by applying a decision without a card, the chat doesn't stall; the next-step turn continues automatically (counted as one chat turn). Even if you ask for a script first, this question comes first. If you choose the planned mode, you're then asked for a target episode count (for example, "50 episodes"); answer with a number and a Full episode plan card arrives in the same turn — it contains a logline and a 2-minute episode design for every episode, and applying it saves it as the project's per-episode plan (episodes you've already decided aren't overwritten). The plan is designed on a proven short-drama formula (hooks in the first 3 episodes, paywall conversion points, a regular twist cycle, a stepped emotional curve from suppression to release). As the episode-count question says, answer "switch to writing now" and it writes right away without a plan. The decision is saved to the project, so a new chat won't ask again.

• The server enforces the order — proposals that skip ahead, such as an outline before the writing mode is chosen or a

scene before the plan is applied, are filtered out by the server, which tells you why.

1. Scene outline — when you ask for a whole episode, like "Write episode 1", it doesn't start writing straight away; it

first proposes a scene outline (each scene's purpose, conflict and length). Approve it and it becomes the blueprint. Applying the full plan is also a trigger — once you apply it, the chat doesn't stop and moves on automatically to the next step (such as proposing an outline). Short-drama outlines follow the episode formula (a hook in the first 3 seconds → conflict → twist → final hook, 1–3 scenes, 2–3 central characters).

1. Writing in sequence — the outline's scenes are written one at a time at full length. If the episode was started from

the per-episode plan, the co-writer knows that episode's plan (its role, incident, ending, hook for the next episode, and the end state of the previous episode) — it doesn't use up the ending too early and sets up the next episode's hook at the end. If the plan and what you actually wrote diverge, what you actually wrote wins. When you apply a scene proposal, the next scene is written automatically (a light per-scene price). Uncheck "Request the next scene after applying" in the apply dialog to skip continuing this time — there's nothing to type, just one checkbox. If you want to stop, just don't apply — applying is the trigger. What you apply is saved to that episode's script, and you can check it right away on the script screen.

1. After finishing an episode, on to the next — once you apply the last scene of the outline, the co-writer recaps the story

so far and proposes the next episode. On the turn that starts a new episode, it doesn't start with scenes; it first asks 1–2 expansion questions based on the story so far (what grows in this episode, new characters or secrets coming in, and so on), then proposes that episode's outline.

2-minute episode design — the contract that makes an episode actually work

In a short drama, each episode's plan comes with a 2-minute episode design. If the logline is "what happens," the 2-minute episode design is the working standard for how that event starts, turns and ends within one short video.

• When the co-writer builds the full episode plan, it proposes each episode's logline together with the design above.
• Episode status (planned, writing, done) follows the moment you apply scenes through the co-writer —

applying the first scene marks that episode Drafting, and applying the outline's last scene marks it Written. Episodes not linked to the plan (created directly on the script screen) are appended in sequence.

• The co-writer designs only the current episode and the next 5 in detail. Distant episodes keep just a logline and

long-term turning points, so 50–100 episodes aren't locked in too rigidly from the start.

• When you start the script for a planned episode, the co-writer reads this design and writes the scene outline and dialogue.

If the script actually differs, the script you wrote and the scene memories you approved take priority, and the plan never reverts your script automatically.

The episode contract isn't a device for forcing the story into a formula. It's a design tool that checks whether a character's choices and their cost leave a real change within a short running time, so that episodes don't drag on as repeats of the same situation.

Editing gate — a structural check that sweeps 50–100 episodes automatically

Once you have an episode plan, the current-step guide at the top of the co-writer screen shows an editing readiness score and the problems to check first. The check reads the saved episode plan and is calculated automatically, so it makes no separate AI call and costs no credits. The co-writer gets the same result and proposes fixes, starting with the current episode and the ones nearest to it.

The editing gate checks the following.

• 2-minute feasibility — missing episode designs, and going outside 60–120 seconds or 1–3 scenes
• Story momentum — repetition of similar core incidents, end states, obstacles and twists
• Hook variety — whether the same opening or final hook is repeated with only the wording changed
• Emotional change — whether only the emotional intensity keeps rising while the relationship's direction stays the same
• Final episode convergence — whether the outcome, cost and relationship resolution of the long-term central question appear in the ending

Exact duplicates are searched across all episodes; near-repeats with slightly different wording are compared mainly against nearby recent episodes. Problems aren't listed as dozens of cards per episode; they're grouped by category. So even in a 100-episode plan, you can quickly find which type of repetition to fix first and which episode numbers it affects.

The editing readiness score from 0 to 100 isn't a score for hit potential or how entertaining it is. It shows how ready the structure is on things that can be checked mechanically — length, momentum, hooks, emotional change and the ending. A low score never edits your script automatically or reverts it to match the plan; the co-writer tells you the problem and a direction to fix it, then waits for your choice.

Episode 1 readiness above the chat looks at "do you have the assets to start writing the first episode," while editing readiness looks at "can the plan for many episodes run without repetition." They're different measures.

Writing AI follows the operator settings

There's no model selection on the co-writer screen. Operators manage the default model and availability for each task, and each request follows those settings. You can focus on the story's direction and on reviewing proposals.

Record-type proposals are recorded without asking

Among the AI's proposals, foreshadowing, character knowledge and facts aren't decisions but records of what was just written. These cards were almost never rejected, and when they were skipped without being handled they never made it into the canon, so the AI would forget that foreshadowing when writing the next scene. That's why these three types are recorded in the project canon the moment a reply arrives. The overview (logline, synopsis and so on) is filled automatically only when that item is empty; proposals that change a sentence you've already written stay as cards and wait for approval.

• What was recorded appears as one line under the reply, like Recorded · Foreshadowing 1 · Knowledge 2. Click Details to

see a summary of each item and an Undo button. Undoing removes it from the canon and leaves "Undone" in its place. A character's knowledge is merged into one entry, and a card that pays off existing foreshadowing doesn't create new foreshadowing; it marks the existing one as paid off.

• You can still edit things directly on the canon screen — automatic recording removes the approval step, not your

editing rights.

• A card that revives a retired fact, and an older card in the same reply for an item that has a newer proposal, aren't

recorded automatically; they stay as cards with a warning.

• Cards left in chats from before this feature existed remain cards (it isn't applied retroactively).

Order of handling proposals — planning updates first, one question at a time

One reply can bring two different kinds of things at once.

When one reply brings two or more planning updates, a bulk apply button appears above the cards. The button says what it applies by name, like Apply 2 characters and locations — episode outlines aren't included. An outline needs you to choose which episode it goes into, so it's applied separately from its card.

Because choices move the chat forward, the screen helps you keep the order:

• While unapplied planning proposals remain, the options are locked — you'll see "Apply or hold the N planning proposals

first." Holding also counts as handling, so hold proposals you don't want (applying isn't forced). Once handled, the options unlock.

• Only one question is open at a time. Even if several questions arrive, only the first is shown, and the next follows

when you answer (you're told in advance: "Choosing this answer leads to N more questions").

• The "N proposals not applied to planning" banner at the top counts only cards for the episode you're writing now.

Cards left from earlier episodes' turns are folded into "Includes N proposals from earlier episodes," and clicking it adds them to Apply all. Cards that would put a retired fact back are left out of the count and bulk apply, and carry a warning. The option lock ("Apply or hold the N planning proposals above first") uses the same rule — folded cards from earlier episodes don't block choices.

• The 2–4 options attached to a question are shown as they are — choosing is the point, so they aren't trimmed. If none

appeals, type a different direction yourself.

• If you don't like a proposal at all, you can ask for another — hover over a proposal card and the Propose again button

appears. Pick a reason (predictable, doesn't fit the canon, and so on) or write the direction you want, and that feedback goes to the AI, which sends a new proposal in a different direction (re-proposing a chat or planning card is also one chat turn, so in a short drama it's 1 credit. Only scene re-proposals are free once per scene).

• Propose again changes only that one card — when several cards (logline, intent, characters and so on) arrived in one

turn and you ask for another proposal on one of them, the rest stay and you can keep applying or holding them. The AI knows which cards haven't been handled yet, so it doesn't ask about the same item as a new question or send duplicate cards. Previously, re-proposing even one card made the AI regenerate the whole turn, pushing the other cards aside as stale.

• Questions you rejected or re-proposed don't come back — a folded question is also closed in the project's list of open

questions. Previously, questions only closed when you picked an answer, so rejected questions would resurface in the next chat.

• If you don't have enough credits or the AI is still replying, the re-proposal request doesn't start and you're told

why — a card never gets locked as "Re-proposal requested" without the request actually going out.

• Rejections are remembered too — when you reject, you can pick or write a reason, and the rejected direction is kept in

the project record so the same direction doesn't come back in the next chat. Previously only approvals were remembered and rejections left no trace, so after several rejections the AI would repeat the same proposal. Even if you reject without a reason, the fact that "not this direction" is kept.

• After you reject, you're asked what to do next — when you reject a proposal that drives the chat, such as a question,

scene or outline, three options appear right there: Propose a different direction · Skip this question and continue · I'll write it myself. Previously, rejecting only closed the card with no guidance, so the chat looked stuck. I'll write it myself doesn't call the AI (it costs no credits). If you reject several questions in a row, the prompts appear one at a time (with a queue count) — a later rejection doesn't overwrite an earlier prompt.

• Unhandled proposals from an earlier chat stay as a banner above the input box: `There are N planning proposals from an

earlier chat that weren't applied`. Click Go to it to jump to where that proposal is — even in a long chat, missed proposals don't quietly disappear.

• It prevents accidents from applying an old proposal too late: when a newer proposal arrives later for the same item (the

logline, a specific character and so on), the old proposal card gets the warning A newer proposal for the same item arrived later in the chat — applying this now will overwrite the latest value, it's left out of the banner count, and Apply all skips it (you can still apply it yourself, one at a time). (Proposals that accumulate, like facts, and foreshadowing proposals that update the same canon entry by planting or paying off don't get this warning.)

This way, "I answered a question and the character proposal got pushed up and left unapplied" doesn't happen.

When the chat seems stuck — two safety nets

If, during planning, the AI sends only apply cards and ends without a next question, the server adds a "Shall we decide ○○ next?" card in that spot. Choose Suggest options and you get 2–4 candidates for that item with pros and cons; choose I'll write it myself and nothing is sent, so just type in the input box. This card is a choice about how to proceed, not a story decision, so it isn't saved to the project record.

If there's still nothing to do anywhere on screen (no cards, no questions, and the reply is finished), the line "The conversation stopped here. Continue?" appears above the input box. Click Continue to the next step and a chat turn that continues to the next step goes out (it's one chat turn, so in a short drama it's 1 credit). It never goes out automatically — you can close it and type yourself.

Scene memory — how the AI remembers the flow of an episode

When you apply a scene to an episode, a memory of that scene is saved with it — a summary of what happened in the scene in 1–2 sentences and the state changes (who learned what, how relationships, belongings and locations changed). Under the scene card you'll see "Saved to scene memory," and if you don't want it, Delete removes just that memory (the scene text stays). Previously a separate card asked whether to save it, but that meant approving the summary of a scene you'd already applied a second time, so we removed it.

• Saved memories are what the AI relies on when writing the next scene — beyond the last one or two scenes, "what happened

in this episode (and the previous one)" is kept.

• If you edit the scene's text after saving, the memory is automatically excluded (a memory that no longer matches the

changed content isn't quietly used). Undo the edit and it comes back.

• Saved memories don't disappear as episodes go by — memories from more than 3 episodes back are condensed into

previous-episode summaries, and scenes from distant episodes that connect to the current scene (for example, foreshadowing planted in episode 1) are found automatically by related-scene search and shown alongside.

• The "What the AI saw" line above the input box shows how many of each asset were actually included in the last request

(facts, decisions, foreshadowing, knowledge, scene memories, previous-episode summaries, related scenes, the previous scene). If only part was included, like Facts 60/72, it also shows how much was cut.

Project canon — the project's memory

Chat history isn't memory. The project's facts, decisions, things not yet decided and foreshadowing are saved in the project canon, and the co-writer and episode checks rely on it to keep things consistent.

• You can view and edit it directly in the canon panel on the right of the co-writer screen. Each section is collapsed by

default, and the count next to its title shows how much has built up.

• Open questions are closed by the co-writer itself once they're answered, and you can resolve the remaining ones in bulk in the canon.
• When the co-writer pays off existing foreshadowing, the payoff card shows both what it pays off in the canon and where the payoff happens, and applying it marks the existing foreshadowing as paid off.
• The continuity check points out where the canon and the text don't match (foreshadowing that isn't paid off, and so on).
• When you change a fact, impact analysis shows which episodes and scenes are affected.

When can you start the script — any time

You don't have to answer every question before the script starts. Say "Write the first scene" at any time and it writes it on the spot — minor things that aren't decided yet are assumed to fit the story instead of being asked about.

There's no write button. Writing scenes follows from the flow of the chat — when things are ready, the co-writer opens that spot with a scene proposal, and applying it puts the scene into the script.

In rare cases, when a reply hits the length limit and the end gets cut off, it says so at the end of the reply — say "Keep writing" and it continues from where it was cut.

When an episode is done, move on to the next

When all designed scenes are written, the co-writer sends an "On to the next episode" card. Apply it and the next episode starts right there — no need to go to the planning screen and back.

• If there's a per-episode plan, it picks up that episode and first confirms with "The planned incident is …," and if

you agree, it sets up that episode's scene outline.

• If there's no plan, it creates the episode and starts by deciding "what should happen in this episode."
• If the episode plan has run out (for example, a 10-episode plan with all 10 episodes written), the next episode

doesn't start, and you're told "The episode plan ended at episode 10." Click Ask to extend the episode plan in that notice and an extension request goes to the co-writer; apply the episode plan card that follows and you can keep writing the next episodes. A lost connection and a server error are each reported as such, so just try again.

Scenes don't keep getting added to a finished episode — episode boundaries are units of the story. Even 50- or 100-episode projects can continue this way, just by chatting.

Where to see how far you've come

Episode progress (which scene you've written up to) is shown on the script screen. Above the chat, only episode 1 readiness is shown in one line (about N% · what to decide next).

• Confirming an outline leads straight into writing the first scene — the same principle as applying a scene leading to the

next one (applying is the trigger · the continuing turn is charged the scene price, shown in advance in the confirm dialog). To stop, just don't confirm or apply.

• The button doesn't disappear during ideation either. Questions continuing to come doesn't mean you can't write yet.

When a reply takes a long time

The "Thinking…" bubble shows how long you've waited, in seconds. A chat turn usually takes 5–15 seconds and writing a scene around 30 seconds; after 1 minute, the bubble tells you so. If the server can't finish a single AI call within 2 minutes, or a whole turn within 4 minutes, it cancels the request itself and replies "The server stopped the request" — no credits are charged in that case, so send the same message again after a moment. On the other hand, if the browser stopped after waiting 5 minutes ("Stopped waiting") or you left the screen by refreshing, the server may still be generating the reply — after a moment, check the chat history and your credit balance first, and send it again only if there's no reply.

Help inside the editor (refining a selection, episode checks) is covered in Writing a script — the co-writer is "a room for building together," while editor help is "a tool for your own writing"; they play different roles.