Agent guide · You keep the decision gate

Ask, inspect, then decide ​

Agent works inside the open story project. It can read Lattice source, use story-aware skills, ask focused questions, edit files in Act mode, and show validation results. You choose the creative direction and decide which plans move forward.

Act mode for direct work ​

Act is the normal mode. Give it an outcome, the important constraint, and the evidence you expect:

Tighten the dialogue in opening without changing any Block IDs or route targets. Validate the story and summarize what changed.

Specific requests help Agent preserve the right boundary. You do not need to prescribe file operations; describe what should be true when it finishes.

Use Act when the request is bounded, reversible, and safe to implement immediately. Use Plan when the work needs review before any file changes.

Goal mode for persistent outcomes ​

Use Goal when you care more about the finished outcome than a single Agent turn. Goal mode keeps one objective active across Agent responses. The Agent must explicitly finish with completion evidence, report a repeated blocker, or enter an external wait; an ordinary final message does not silently complete the Goal. If information from you is required, TypeApe shows the usual question interaction instead of guessing.

Goal mode runs up to 10 automatic model responses by default. In System settings → AI → Goal, you can choose another positive number or Unlimited. Unlimited removes the automatic-response ceiling, but stopping the run, depleted Credits, a token budget, repeated lack of progress, runtime failures, blocking, and external waits still stop or pause it. The conversation shows the active state and configured ceiling. If Goal stops before completion, choose Restart Goal to clear that stopped run and start again from the saved objective. After completion, its evidence summary remains visible until you start another request.

When the progress card says that Goal is waiting, the Composer stays available with Goal selected. Send the new information or external result there to wake and continue the same run. If the Agent needs a specific decision from you, it uses the usual question interaction instead.

Use Goal for a verifiable multi-step result, such as completing and validating a story revision. Use Act for a quick direct change, and Plan when you want to approve the approach before files change.

Plan mode for consequential changes ​

The Plan toggle applies to the next submitted request. During that turn Agent can read, search, analyze the story, ask questions, and build a task checklist, but it cannot edit project files.

Plan mode is useful when you want to:

• add or restructure a CYOA route;
• introduce State, a Pool, or several Modules;
• split source across files;
• coordinate a change with important creative trade-offs;
• decide how you will know the change is complete before work begins.

What happens during a Plan turn ​

1. Agent inspects the project.
   It reads the current source and relevant structure before proposing work.
2. It resolves consequential ambiguity.
   Focused question groups cover choices that would materially change the result.
3. It creates a clear checklist.
   The full plan stays in the conversation, while the sidebar shows progress.
4. Plan review opens.
   The embedded review tab lets you annotate, approve, request changes, or leave the decision pending.
5. Agent follows the plan you approve.
   If the plan changes, you review the new version before work starts.

Closing Plan review does not approve or reject anything. The pending tab remains available. If you start a new Plan request while an unfinished checklist exists, TypeApe asks before replacing the active plan and keeps the earlier task in conversation history.

Annotate a plan ​

Use plan annotations to review the approach before source files change:

• Attach a focused comment to the relevant part of the plan when one step is wrong or unclear.
• Add a global comment when feedback changes the plan as a whole.
• Choose Request changes when the plan needs another pass. Agent returns with a revised checklist for review.
• Choose Approve when the steps, boundaries, and expected result are right. You can include final guidance with the approval.

Approval applies only to the plan on screen. If that plan is edited or replaced, TypeApe asks you to review it again.

Leave reading annotations in Story ​

Reading notes start from the story as experienced, not from a file path.

1. Select exact prose to anchor the marker beside that quote, or activate a passage/dialogue marker without a selection to annotate the complete item.
2. Write the required revision and save that note.
3. Repeat while reading. Each card keeps its quote and Block context.
4. When every note is saved, choose Send notes to Agent.

TypeApe starts a fresh Agent conversation containing the complete review. Quoted story text identifies where the note belongs; it is not treated as an instruction. Agent then reads the project files, applies the revisions, and validates the result.

Unsubmitted reading annotations are temporary. Restarting or leaving Story, switching projects, or closing TypeApe clears them. After you send the notes, the Agent conversation keeps a lasting record of the review.

Redirect work while Agent is running ​

The Composer remains available during a run. Choose the delivery that matches your intent:

Stopping a run also restores queued text before aborting. Use steering for a correction such as “preserve the existing narrator”; use follow-up for a distinct next task such as “after validation, suggest three titles.”

Questions, evidence, and recovery ​

When Agent needs a consequential decision, answer the in-product questionnaire; related questions appear one at a time and submit together. If execution or revision fails, the conversation and checklist preserve the visible state so you can retry. After a failed revision, the Composer returns to Planning mode; send a narrower instruction there to add it to the saved feedback and retry the same plan.

Before accepting story work, check three things:

• the files or story concepts that changed;
• parser and validator results, including any warnings you reviewed;
• a plain-language account of what the reader can now experience.

Return to the Lattice language guide when you want to inspect the source yourself.