AI collaboration overview

AI collaboration

AI collaboration overview

Three doors: selection, continuation, chat-to-proposal. Nothing writes without your OK.

Three collaboration entries

Loomworld’s AI has exactly one entrance: the chat box at the left of the workspace, where the agent answers. Three moments run through it—right-click a selection to adjust it, ask for a continuation of a chapter or scene, or talk something through first and let the conclusion become a cross-module proposal. It reads the records the current task needs; anything that would write into the project must come back as a proposal first.

Selection edit

Select a passage, right-click, and choose rewrite, expand, condense, or polish; add a more specific requirement if you have one. This is local wording work. Changes that cross chapters or modules should name their range directly in the chat box.

The result arrives as a before / after diff. It replaces the original only after you confirm. Details: Selection rewrite.

Continuation: down to a chapter, scene or scene setting

Name the target in the chat (which chapter, which scene, which scene setting) and what this stretch has to move. The agent writes on with that unit’s context—the scene’s main location and its descendants, the chapter intent, the as-of-chapter state of related records—and returns a pending proposal; once confirmed the text appends from the end of that unit.

The companion panel on the right carries no generation button at all—it holds structure. See Continuing with the agent.

Talk it through, then make it a proposal

Work out a motive, argue about direction, check whether a fact holds—discussion writes nothing. When it converges, say plainly “turn that into a project change”, and you get one confirmation card listing before / after, the target objects, and the impact range.

While it works there is a single status line (thinking / reading character · Ah-Zhen / searching the project / composing changes); afterwards “view process” expands the steps this round took. Sessions stack up, rename, and collapse; the conversation before project creation stays with the project as the “Project start” session. Details: Chat to proposal.

Confirmation: AI does not write your draft

Whatever the entry, a write into the project passes through one card: before / after is visible, confirming applies it, refusing or closing applies nothing. A card that outlives its validity shows “Proposal expired” and can be regenerated with “Update and continue”; if the target changed elsewhere before you confirmed, the whole card is refused instead of quietly overwriting.

One confirmation carries at most 8 atomic operations, committed as a single transaction; after it lands, the card sits on the project undo stack, so one undo withdraws the whole set—including deletions.

AI output can be wrong, repetitive, or off. Review it before you take it in. Do not rely on AI output for facts, law, medicine, or other specialist content. Final cut and factual judgment stay with the author.

What AI may change, and what it may not

In scope: create, edit and delete across the six library objects (characters, locations, props, lore, foreshadows, timeline); create and delete character relationships; patch script text; and chapter, scene and scene setting structure fields, including creating and deleting them.

Out of scope: per-chapter keyframes, a relationship’s chapter states, leaderboard scores, hand drawing and placements on maps, the background maintenance pipeline, and canvas positions. The list is a whitelist—what is not on it cannot be written. Full list and validation rules: What AI may write.

Background maintenance and project settings

Stop writing for a while and chapter summaries, the story overview, entity dossiers, continuity reports and beat summaries update themselves (maintenance calls are tiered by complexity and read the project’s current state). The switch and “update all profiles / check this chapter / check every chapter” live in project settings → AI settings; chapter intent and scene summaries you typed by hand are never overwritten.

Writing style is a separate thing: the project default is a nine-dimension structured template (the six built-in templates become the project’s own snapshot once loaded), and it changes only how prose sounds, not project facts or plot outcomes. Interface language and project language are separate too—an English interface writes a Chinese script just fine.

What you pay for

Three ledgers, clearly split: the software itself stays usable long term, with the script editor and reference modules complete and records on this machine; subscriptions include unlimited AI archive maintenance and cloud sync, plus a monthly grant of AI collaboration credits; beyond the grant, credits are bought by volume at the rate shown in the product.

The Global edition uses Creem and the CN desktop website uses Alipay. Confirm the product and price after sign-in. Subscriptions last 30 days and renew manually, without automatic charges. Billing and credits and Pricing.

Limits and troubleshooting

  • AI features require a signed-in cloud account; signed out, those entries do not quietly fall back to a local model.
  • A call first drains local changes to the cloud and checks the version; a mismatch blocks the task openly instead of running on stale data.
  • If content is too long for the model, it tells you instead of truncating your text and sending it anyway.
  • Up to 4 images can accompany a task; if the runtime cannot see images, it refuses rather than sending text and pretending it looked.
  • The agent is not an auto-writer: it never schedules the next scene for you, and it changes nothing while you are away.