# Continue with the agent

> In chat, name a chapter, scene, or scene setting; continuation comes back as a confirmable proposal.

Source: https://loomworld.ai/docs/ai-continuation

## When to let it carry on

Two cases pay for themselves. The skeleton is up (chapter intent, scene summary, beats all written) and needs turning into prose; or you are stuck behind one node and want to see a few ways forward. Where the project is nearly empty, write the people, the conflict and this stretch’s goal first—control improves a lot.

## Name the target: chapter, scene or scene setting

Say how deep to go in the agent chat: “continue chapter 12”, “keep writing the ‘rainy night confrontation’ scene”, “three more beats after Scene setting · afternoon”. The agent then assembles context for that target—the scene’s main location and any of its descendants, the chapter intent, and the as-of-chapter state of the relevant characters and lore.

When the range is vague it queries the project or asks you one clarifying question instead of guessing. Before the answer streams, one status line shows what it is doing (thinking / reading chapter · 12 / composing); afterwards “view process” expands what this round read.

## How

- **Open the chat box**The floating panel on the left; drag it somewhere it does not cover your text. It does not close when you click outside, and collapsing the session list leaves the main workspace width alone.

- **State the goal and direction**Say which unit (chapter / scene / scene setting), what must move, what to avoid. Ready-made entries such as “continue this chapter” are a fine starting line; add your own requirement after it.

- **Read the proposal**The result returns as a pending proposal. Script changes are listed as patches, not a whole-scene replacement.

- **Confirm or take over**“Apply changes” appends from the current end of that unit. Do not want to wait? “Stop and take over” keeps what already generated, and you write the rest.

## Structure and format rules

- Continuing a scene or scene setting never opens a sibling unit mid-stretch. If a new scene is what the story needs, the proposal must declare “new scene” so you see the impact on the confirmation card.
- Appending is safe: continuation adds to the end of the target unit, never rewriting existing text or replacing a node’s identity.
- Headings match the script: “# Chapter N Title”, “## ordinal · location · time · short title”, “### location · time”, and character lines carry an actual name. A missing main location, an ambiguous place, or bad format refuses the whole operation.

## How to ask so the result is yours

- Name the unit and the move, not the mood: “continue chapter 12; this is the first lie Lin Mo tells” beats “make it tenser”.
- Say what must happen and what must not: which information has to land in this scene, which relationship must stay unspoken. Those two lines set the boundary, and afterwards you only judge whether it was crossed.
- Skeleton before prose: intent, scene summary and numbered beats stop the model inventing events. Rough beats are fine—expanding one beat into three paragraphs is the usual move.
- State point of view and tense: whose eyes this scene follows, present-tense now or retrospective. One sentence saves a whole round of rework.
- Take long stretches in batches: ask for three beats, then expand the one you want, instead of asking for two thousand words at once. Big single asks average out into padding.

## Limits and troubleshooting

When a run is interrupted or hits the output cap, the answer keeps what it generated and is plainly marked “incomplete”, and you can ask it to go on. That does not leave the project half-written: generated-but-unconfirmed text is not project fact, and nothing is stored until you confirm.

Openings and imported manuscripts share this delivery layer. After a project is created, the opening lands on the page in real stages (a thin idea still has a 300-character minimum), stopping early hands the pen to you at once, and a technical failure permits one “Complete opening” pass that only appends from the current end. TXT / DOCX script imports call no model at all.