Generate a novel outline
Outline generation is available from Novels → Novel generation or an explicit creation request in standard chat and is the app’s only agentic novel-generation workflow. Lis Novel can turn a writing request into a reviewed scene-level outline. The current outline workflow creates a new novel with world planning, Lore, a story blueprint, a chosen structure, and a complete scene plot. It does not write Scene Beats or prose yet.
The generation screen and chat card show the same current stage, attention and request activity. When an outline finishes, both show the saved novel title and generated structure/Lore counts when available. These counts describe what the run published: they exclude starter content and later author additions, and may be omitted when publication details are unavailable. Closing or reloading either view leaves generation running.
Start an outline
- Open Novels and select Novel generation.
- Describe the novel you want to plan. Add constraints such as genre, tone, or approximate length in the optional field.
- Choose a planning model. You can let Lis Novel infer the structure and involvement mode, or choose them explicitly.
- Select Generate outline. The run opens immediately and remains available while it works.
The run starts before the novel appears. After Lis Novel accepts the story blueprint and structure choice, it creates the novel and shows Open novel. You can inspect it while the run continues. Editing is disabled until the run stops or reaches a terminal outcome.
For an OpenCode Go connection, set Custom Base URL to https://opencode.ai/zen/go/v1. Outline generation supports models served through Responses or Chat Completions, including GLM and Kimi; check the OpenCode Go endpoint table for each model's API. Lis Novel supplies the required session header and selects the model's supported protocol from bundled defaults. Use Model Traits to override the protocol if your gateway differs. You can resume a run previously paused by a protocol error after updating the app. Models served through Anthropic Messages remain unsupported. This support applies to novel outlines; Beat generation and rewrites still require Responses models; chat also selects Responses or Chat Completions from model defaults or a protocol override.
Start from chat
In a standard chat, including one without a novel, ask: “Create a new novel outline about a lunar courier investigating a missing delivery. Let me review each proposal.” Use a chat model that supports tools. The assistant submits a self-contained brief and uses that chat’s selected connection/model for planning; generation keeps its usual reasoning policy and involvement defaults unless you specify options.
Every launch creates a separate novel. It does not replace the chat’s current novel, change chat linkage, or automatically read the current manuscript, Lore or attachments. The original request is saved in the run.
Chat stays busy while its generation tool works, including automatic retries. The inline card shows current stage/activity and Open generation. When an approval, repair, pause or completion checkpoint is reached, the tool hands off and chat becomes sendable after its response settles. Review ordinary approvals in the originating idle chat or on the generation screen. Repair choices and precise lifecycle/retry controls are also available in the originating idle chat. Independent continuation updates the card. On completion, a Lis Novel · Generation update message names the saved novel, summarizes generated structure and Lore counts when available, and offers Continue with this novel, Open novel and Open generation. It confirms an outline with Lore; manuscript prose and Scene Beats have not been generated. This app-generated summary returns after reload, reopening the chat, or app restart without adding another AI reply.
Chat Stop pauses the still-attached run and preserves accepted work. Resume on the generation screen continues independently; Stop run permanently ends the run. Chat Stop after handoff only stops the chat response. Closing the chat window leaves generation working. Restart restores durable progress without restarting the interrupted chat tool. Regenerating or cloning that source message refers to the same run; a new user message can request another novel.
Continue working with the saved novel
When generation completes, choose Continue with this novel on its completion message to keep working in the same chat. Open novel and Open generation only navigate; they do not change the chat's target. Generation still produces an outline and Lore, without manuscript prose or Scene Beats.
If the chat already uses another novel, confirm Switch novel and continue after checking both titles. The switch applies to the whole conversation, including every branch and future tool request. Your earlier discussion and unsent text stay. Pinned attachments and temporary attachment selections are cleared; attach what you want again. Cancel leaves everything as it was. For an unlinked chat with attachments, choose Continue and clear attachments to confirm the reset.
The same chat stays selected in the generated novel. Other open windows update their target and keep their own typed text. Using this novel confirms the current association. If delivery is uncertain, use Retry exact continuation; it checks the same request. The association and saved draft text survive reopening and app restart.
Now ask “Does the plot make sense?”, “What could improve the finale?” or “Propose a stronger ending.” The assistant reads current saved outline/Lore using its normal tools; earlier generation proposals may no longer describe the saved novel. Outline and destructive Lore proposals still require their usual Apply/Dismiss decisions. No message is sent by the continuation action itself.
Use a fresh chat if you want a clean conversation history. Changing branches or regenerating an older reply keeps the current novel target. A copied standard chat can continue with the saved novel, but cannot approve or resume the original generation. Busy chats or chats needing history recovery must settle first. Reopening generation keeps the chat linked while the usual generation editing restrictions apply.
Deleting a linked novel removes its scoped chats through the normal novel-deletion flow. Deleting a chat leaves its independently generated novel/run intact. A deleted or unfinished generation result cannot be selected for continuation.
Discuss a proposal in chat
When a chat-started run pauses for an ordinary approval, you can ask the assistant about the current proposal's motivation, causal logic, or wording. It reads the permitted saved proposal, identifies its version, and explains relevant limits: it has not read the full surrounding outline or manuscript, and some saved review evidence may be unavailable. This discussion is separate from a new plot review.
Ask for suggested changes. When the assistant prepares concrete revision wording, a Suggested changes · Version N card shows the full direction. Once chat is idle, select Review suggested changes to inspect that exact saved proposal and its review findings. Edit the direction, then select Send revision request. Cancel makes no generation change. You can still write your own direction through Review proposal → Request changes.
Preparing or opening a suggestion does not revise or approve the run. A suggestion remains bound to its original proposal: if the checkpoint or disclosed context changes, ask for a fresh suggestion. Review any parent-replanning consequence and the replacement version before approving. A copied chat retains readable suggestions but cannot review them for submission. Switching suggestions asks before replacing an edited draft; unsent drafts may be lost on reload. Saved suggestion cards return after reopening or restarting the app.
Full author previews remain inspectable in chat when the model cannot read restricted or oversized detail. Repair checkpoints have their own filtered context; completed novels use ordinary saved-novel tools after continuation. Stopping the chat discussion leaves independent generation running. Previously read content stays in saved chat history even after a later visibility change; fresh reads follow the current visibility.
Scene estimates and chapter pacing
The story blueprint includes an Estimated scene count and a rationale for the story’s planning needs. Scene planning may adjust this estimate as it combines related moments or separates developments that need more space. If you request an exact scene or chapter count, that instruction remains a requirement. An approximate word count does not set a scenes-per-chapter formula.
Above a scene-plot preview, you can see the total scenes and each chapter’s scene count. Act groups appear when the structure includes acts; scenes-only outlines show just the total. These figures describe that saved proposal or historical revision, so later manuscript edits do not change them.
Chapters can have equal or different scene counts when that suits their dramatic purpose. Reviews assess redundant scenes, rushed developments, transitions and chapter endings. Editorial pacing weaknesses are warnings; equal chapter sizes or a changed estimate alone do not trigger a story repair.
Continuity and story logic
Scene planning connects important objects, secrets, clues and abilities across the outline. The AI reviews whether characters have the knowledge and access their actions require, whether evidence supports their discoveries, and whether the ending follows established events and world rules. Planning and review work backward from the climax to check its prerequisites, including object availability and whether travel, rituals and deadlines fit together. Moving a crate does not automatically move something beside it, and one character discovering a secret does not mean everyone knows it.
An essential missing connection can block the outline when its impact is supported by evidence. Ordinary off-page handoffs, deliberate withholding and mistaken beliefs are valid storytelling choices. Minor timing ambiguities, repeated explanations and weak protagonist agency can receive warnings; the detective need not personally cause every object’s return.
A finding should identify the relevant passages and explain why the sequence does not work. Essential blockers can trigger automatic story repair.
Reviews can still miss problems or misjudge their importance. Evidence confirms where a finding comes from, rather than proving its interpretation. When inspecting the scene plan, pay particular attention to how important objects move, how characters learn secrets, and what makes the climax possible.
Review decisions
Autonomous proceeds without questions. Milestones pauses at the major decisions you select. Supervised asks you to approve each proposed story artifact or decision, including Lore and scene plans. The Activity feed shows the novel title and progress from Writing brief through Final review. Accepted items stay compact; select an item to inspect its contents. The expanded Your next decision card identifies the exact item and revision awaiting approval. A World revision and a Story arc revision belong to different items, so their revision numbers do not need to match.
Choose Approve [item] and continue to accept that proposal. Choose Request changes to open the revision field, describe what should change, and select Send revision request. Cancel closes the field without submitting. A revision creates a new draft of that item. Earlier revisions holds earlier and withdrawn proposals, with review findings attached to the draft they concern.
For Structure, you can select a mode and choose Use selected structure. Revising a Lore entry, story milestone, or hierarchy decision regenerates its parent proposal and restarts related approvals; the form explains this before you submit. You can also expand accepted work and choose Reopen [item]. Read the consequences and confirm Reopen and replan to withdraw dependent work and generate a replacement. If an affected published scene plot contains your edits, added structure, prose or Scene Beats, the app refuses to reopen it and preserves your work. This protection also applies when reopening an earlier stage or a child item would replace that scene plot.
When you open a run, it scrolls to the pending decision once. Later updates preserve your reading position. Use Jump to current decision to return to the approval card.
The run separates Story findings from Run errors. Each review checks the deliverable for its current stage. Early planning does not need finished chapters or scenes; the scene plot and final audit check the complete outline. A blocking story issue can trigger up to two targeted repairs in that review cycle. Pausing and resuming preserve the repair allowance; an author revision starts a new cycle. If it remains unresolved, the run ends incomplete and preserves accepted work.
You can expand Review evidence and repair goal beneath a finding to read the passages it cites and the intended correction. Reviews check story meaning; Lis Novel checks required fields, types, and scene links. Editorial weaknesses are warnings. Evidence makes a finding inspectable, but does not guarantee that the AI’s interpretation is correct.
If a review cites unavailable passages or does not meet the evidence requirements, Correcting the AI review means Lis Novel is requesting a new assessment of the same proposal. This requires no action from you and uses technical retries rather than story repairs. The proposal remains intact. If correction fails after three retries, The AI review could not be validated offers an action such as Retry Story blueprint review or Retry final review. A failed final review assessment preserves the accepted scene plot; it does not replace your scenes or ask you to select a story repair.
When those story repairs are exhausted, Choose how to continue identifies the affected item and revision. Select a saved suggestion to fill Repair direction, then edit that full direction or replace it with your own. Select Apply direction and retry [stage] to create and review a new revision. This does not approve it; earlier accepted stages stay intact, and your involvement mode still controls subsequent approvals.
Suggestions are prepared only when intervention is needed. You can type and submit your own direction while Preparing suggestions is shown. If none are available, use your own direction or select Retry suggestions. Choosing an option alone never starts generation.
If Final review is blocked, the card explains that continuing replaces the current scene plot and restarts its reviews and related approvals. Earlier planning stays accepted. If you have already edited that scene plot or added prose or Scene Beats, the app refuses replacement to protect your changes.
Repair choices survive an app restart. An interrupted suggestion request is not repeated automatically; the custom idea field and Retry suggestions remain available. Pausing or restarting does not renew automatic story repairs; explicitly submitting a new repair direction starts a fresh review cycle.
While a request is running, a brief status shows Waiting for model, Model activity, or Receiving output. It includes the attempt’s elapsed time and, after output begins, time since the last generated activity. Model activity can include reasoning, but Lis Novel does not show or save reasoning text. The timer is not an estimate of remaining time.
Temporary provider failures and interrupted or malformed responses recover automatically, with up to three delayed retries for each request. The run explains what it is fixing and shows the automatic retry count. Fixing scene links means the model connected a scene to itself or to a later scene; the app asks it to correct the plan. Correcting the generated proposal means a completed response did not pass JSON or proposal validation. Retrying after the model’s output limit means the model stopped before finishing and the app is asking for a complete proposal within the limit. Recovering the model connection means the provider had a temporary problem or its response stream was interrupted. These states require no action from you and preserve approved work. Restarting the app preserves the retry allowance; an explicit Resume run, or the stage-specific Retry action after an error, starts a fresh allowance. Credentials, exhausted quota, request-size problems, and failures that persist after those retries still pause for attention.
A Waiting for your approval of [item] message identifies an approval pause. If automatic correction cannot finish, the run explains the problem and offers a stage-specific action such as Retry scene planning. This continues from that stage and preserves earlier approved work. Connection problems offer Open AI connection settings: check the API key and access permissions, or the selected model’s availability and supported context limit, as directed. Exhausted usage allowance requires checking your provider’s allowance or billing before retrying. Saving failures are identified separately. Timeouts, interrupted connections, output limits, and invalid proposals have separate explanations. An output-limit message means the model stopped before completing its response; check the model’s supported output limit before retrying. Collapsed Technical details contains individual model requests with their timing, finish reasons, available token usage, and safe provider identifiers, along with historical events. Token usage can be unavailable when the provider interrupts a response. Earlier story findings stay with their own revision rather than appearing as problems with the latest proposal. Automatic recovery applies in all three involvement modes and respects Pause run and Stop run. Stopping a run preserves its accepted outline and makes an existing novel editable.
An outline is complete only after the scene plot passes its endpoint audit. The accepted scene summaries appear in the novel's structure; the manuscript has no generated prose. Scene Beats and full-prose generation are planned for later phases.
Model defaults and JSON output
Known model IDs receive context and response-limit defaults even on custom connections. In LLM Providers → Model Traits, you can override limits, choose the outline/chat protocol, and enable or disable JSON-object output. The built-in JSON list initially enables glm-5.3-flash; other models stay unknown until configured. Ordinary prose and chat remain text.
An explicit protocol or JSON-parameter rejection before generated content can trigger a bounded fallback. Technical details records it. JSON responses must still pass proposal and scene-link validation; a valid JSON object can contain story-planning errors that require correction.
Review checkpoints from chat
If you started the outline in a standard chat, its live card can show Review proposal at ordinary approval checkpoints. Inspect the complete saved proposal, exact revision, and review findings before choosing Approve, Request changes, or an offered structure. The chat must be idle before a new decision. Copies of the chat can inspect proposals; decisions belong to the original chat or this generation screen.
Request changes accepts up to 20,000 characters of nonblank directions. Revising a child proposal can regenerate its parent and supersede related approvals. Your decision starts independent generation; the chat remains available and its card updates at the next checkpoint. After handoff, Chat Stop stops only the chat response. Use chat controls or explicit instructions for pause/resume, retry, repair assistance and confirmed permanent stop. Reopening accepted work remains on this screen.
If delivery is uncertain, retry the exact decision. A recorded decision confirms the saved answer, not successful publication or completion. If publication needs recovery, open generation and use its Retry/Resume action. Restart restores saved decisions and pending previews without automatically approving a proposal. See Chat tools for the complete chat flow.
Manage generation in the originating chat
Ask what went wrong to discuss saved findings and their permitted evidence. Story problems and connection/output failures are separate. The assistant names the current revision, explains missing coverage and distinguishes saved findings from its own advice. The card shows full author findings and saved repair suggestions even when the assistant cannot read every source.
Selecting a repair suggestion fills Repair direction without starting work. Edit that wording or write your own, then submit it once. Replacing edited text with another suggestion asks Replace draft. Missing suggestions leave the direction field usable; Retry suggestions is explicit and does not restart the story stage.
You can discuss another direction, then say “Proceed with that.” The assistant submits the exact prepared direction without requiring another click. If several directions are still plausible, it asks which one. An explicit request for a concrete change and execution can prepare and submit in the same turn. Discussion alone changes nothing. The replacement still receives a fresh review and your saved involvement mode controls subsequent approvals; suggestions and command receipts do not establish that a finding is fixed.
Use Pause run, Resume run, the stage-specific Retry action, or equivalent typed instructions. Pending approvals and repairs cannot be bypassed by resume. If “retry” could refer to several operations, the assistant asks which one. Credentials and model limits remain in Settings. UI controls work without a tool-capable or available model once the chat is idle and its history is healthy.
“Stop generation” first asks for confirmation after explaining that the named run ends permanently, cannot resume, and retains accepted content. Confirm in a later chat reply or use the dedicated confirmation control. Decline cancels. Unrelated replies, branch/linkage changes, another management action in any window or on the generation screen, and restarting the app invalidate pending confirmation. UI Stop run permanently supplies the same consequence dialog without a model. If the initial request loses its reply, Retry exact command restores that dialog when confirmation is still required. Chat Stop still stops the current reply and pauses generation only while its initial launch remains attached.
Commands finish quickly while the card observes independent generation. If delivery is uncertain, inspect the run or use Retry exact command; do not submit a second request with a new identity. Saved commands and preparations survive restart; unsent directions are never submitted automatically. Copied chats can discuss permitted context and retain readable drafts but cannot manage the original run. Runs created only on the generation screen cannot be attached to an arbitrary chat.