Writing & Editing Commands

Commands for generating prose, summaries, and making revisions.

/write

Generate prose from summaries. For scenes, writes PROSE.md from SUMMARY.md. For chapters and acts, writes all descendant scenes.

Usage

/write act 1 chapter 2 scene 3   # Write a single scene
/write 1 2 3                     # Shorthand
/write act 1 chapter 2           # Write all scenes in chapter 2
/write 1 2                       # Shorthand
/write act 1                     # Write all scenes in act 1
/write 1                         # Shorthand

How it works

For a single scene:

  1. Reads the scene’s SUMMARY.md
  2. Loads sibling scene summaries for continuity
  3. Loads parent-level .md files (chapter outlines, notes)
  4. Loads db/core/ entries (style guide, metadata)
  5. Generates PROSE.md in an isolated coder session

For a chapter:

Each scene is written with context from all sibling summaries plus any .md files at the chapter level.

For an act:

Writes all scenes across all chapters in the act.

Requirements

  • The target scene(s) must have a non-empty SUMMARY.md
  • An empty PROSE.md is created automatically if missing

/summarize

Generate summaries from prose. For scenes, reads PROSE.md and writes SUMMARY.md. For chapters and acts, summarizes all descendant scenes.

Usage

/summarize act 1 chapter 2 scene 3
/summarize 1 2 3
/summarize act 1 chapter 2       # Summarize all scenes in chapter 2
/summarize 1 2
/summarize act 1                  # Summarize all scenes in act 1
/summarize 1

How it works

For a single scene:

  1. Reads the scene’s PROSE.md
  2. Generates a concise synopsis for SUMMARY.md

For a chapter or act:

Finds all descendant leaf scenes that have prose and generates summaries for each one in a single batched LLM call.

Summary content

The generated summary captures:

  • Key events and plot points
  • Character dynamics and emotional beats
  • Important revelations, decisions, or turning points

Summaries start directly with content—no title heading (the title lives in the directory name).

/edit

Make targeted edits to manuscript files. Works like chatting but makes the editing intent explicit.

Usage

/edit Change Tom's dialogue to be more confrontational
/edit Tighten the pacing in the second half of this scene

With no arguments, switches to the model’s preferred edit mode.

How it works

Spins up an isolated coder with the model’s preferred edit format, runs the prompt, and returns to your current mode. Files currently in the chat are available for editing.

/query

Query the manuscript without editing any files. Delegates to a read-only query mode.

/query What are the main themes established in act 1?
/query Does Elena's character arc feel consistent across these scenes?
/query Identify any plot holes in the chapter summaries

/compose

Enter compose mode, which uses two models: one to plan the edits and another to execute them. This is Aider’s architect mode with novel-mode prompts.

/compose
/compose Restructure the dialogue in this scene to build more tension

With no arguments, enters persistent compose mode. With arguments, runs a single compose operation.

/agent

Plan and execute complex multi-step tasks autonomously. See the Agent Mode guide for full details.

/agent                           # Enter persistent agent mode
/agent Rewrite act 1 in first person POV

/select

Set a text selection for focused editing. Used primarily by IDE integrations.

Usage

/select filename:start_line:start_col-end_line:end_col

Line and column numbers are 1-based.

/select novel/Act 1 - Title/Chapter 1 - Title/Scene 1 - Title/PROSE.md:5:1-12:45

Switches to selection editing mode focused on the selected text.

/instruct

Load an instruction file and inject its content into the current chat.

Usage

/instruct tone-shift             # Load instructions/tone-shift.md (or .txt)

The instruction content becomes a user message in the chat, influencing subsequent AI behavior. Useful for recurring directives like tone shifts, pacing changes, or style overrides.

Tab-completion shows available instruction names.

/lint

Check prose files for style issues using Vale. See the Prose Linting guide for setup and configuration.

Usage

/lint                            # Lint files in the chat (or dirty files)
/lint 1 2 3                      # Lint a specific scene
/lint act 1 chapter 2            # Lint all prose in chapter 2
/lint path/to/file.md            # Lint a specific file

After displaying issues, offers to fix them with AI assistance.

/lint-level

Control which severity levels /lint reports.

Usage

/lint-level                      # Show current level
/lint-level error                # Errors only
/lint-level warning              # Errors + warnings (default)
/lint-level suggestion           # Everything

/analyze-style

Add prose files to the chat as read-only context for style analysis.

Usage

/analyze-style act 1 chapter 2 scene 3
/analyze-style 1 2 3             # Shorthand
/analyze-style act 1 chapter 2   # All prose in chapter 2
/analyze-style 1 2               # Shorthand
/analyze-style act 1             # All prose in act 1
/analyze-style path/to/file.md   # A specific file

How it works

  1. Collects all PROSE.md files from the specified location
  2. Adds them to the chat as read-only context
  3. You then chat with the AI about the style, asking it to write a style guide or answer specific questions

This is a context-loading command, not an automatic generator. After loading the prose, follow up with a message like:

> /analyze-style 1
> Write a comprehensive style guide to db/core/style.md based on this prose.

This approach gives you control over what the AI focuses on and lets you iterate on the style guide interactively.

/feedback

Get a structured, prioritized critique of a manuscript section.

Usage

/feedback act 1 chapter 2 scene 3
/feedback 1 2 3             # Shorthand
/feedback act 1 chapter 2   # All files in chapter 2
/feedback 1 2               # Shorthand
/feedback act 1             # All files in act 1
/feedback path/to/file.md   # A specific file

How it works

  1. Loads the specified files as read-only context
  2. Spins up a temporary query-mode session with the style guide loaded
  3. Returns a structured critique with prioritized suggestions

Output format

The feedback is organized into three priority tiers:

  • High Priority (2–3 items) — critical issues like plot holes, character inconsistencies, pacing problems, unclear prose
  • Medium Priority (2–3 items) — important improvements like dialogue refinement, scene transitions, deepening characterization
  • Low Priority (2–3 items) — polish-level suggestions like word choice, rhythm, atmospheric details

Each suggestion quotes the relevant passage, explains the issue, and proposes a concrete improvement. If a style guide is loaded, feedback references it.

/auto-context

Toggle automatic context identification before each LLM call.

Usage

/auto-context              # Toggle on/off
/auto-context on           # Enable
/auto-context off          # Disable

How it works

When enabled (the default), Composez runs a lightweight pre-pass before each user message to identify relevant files, adding them as temporary read-only context. See Auto-Context for details.

Changes from this command only last for the current session. Use the Settings tab (in the web UI) or .composez to persist the setting.