Writing Workflow
The Core Loop
The fundamental workflow in Composez follows a cycle:
- Plan — outline what happens in your scenes (summaries)
- Write — generate prose from summaries
- Revise — refine the prose through conversation
- Summarize — update summaries to reflect changes
- Repeat — move to the next scene with accurate context
This loop keeps summaries in sync with prose, which is critical for context management. When writing Scene 5, the AI reads the summaries of Scenes 1–4 rather than their full prose, keeping the context window focused and efficient.
Step by Step
1. Set Up Your Structure
Before writing, create the skeleton of your narrative:
> /new act The Awakening
> /new 1 chapter Morning Light
> /new 1 1 scene The Alarm
> /new 1 1 scene The Discovery
> /new 1 1 scene The Decision
> /new 1 chapter The Journey
> /new 1 2 scene Departure
2. Write Summaries First
Add your scenes and describe what happens in each:
> /add 1 1 1
> Elena wakes to the lighthouse alarm at 3 AM. Dense fog.
> She checks the radar and sees something large moving offshore.
> She tells herself it's a whale.
The AI writes the SUMMARY.md. Repeat for each scene in the chapter.
3. Write Prose
Once you have summaries for a chapter’s scenes, write the prose:
> /write 1 1
This writes all scenes in Chapter 1. For each scene, the AI:
- Loads the scene’s
SUMMARY.mdas context - Loads sibling scene summaries for continuity
- Loads parent-level
.mdfiles (chapter outlines, etc.) - Loads
db/core/entries (style guide, etc.) - Generates
PROSE.md
You can also write individual scenes:
> /write 1 1 1
4. Get Feedback
Before diving into revisions, get a structured critique:
> /feedback 1 1 1
This gives you prioritized suggestions (high, medium, low) with specific passages and concrete improvements.
5. Revise Interactively
Load the prose into your chat and revise:
> /add prose 1 1 1
Then converse naturally:
> The opening paragraph is too slow. Start with the alarm sound.
> Make Elena's internal monologue more fragmented — short sentences.
> Add more sensory detail about the fog.
6. Update Summaries
After significant prose changes, regenerate the summary:
> /summarize 1 1 1
Or summarize an entire chapter:
> /summarize 1 1
This drills down to all scenes in the chapter and updates each SUMMARY.md from its PROSE.md.
Working at Different Levels
Scene-Level Work
For focused, detailed work on a single scene:
> /add 1 1 1
This loads both SUMMARY.md and PROSE.md. You can edit either file through conversation.
Chapter-Level Work
For working across a chapter:
> /add summaries 1 1 # Load all summaries
> /add prose 1 1 # Load all prose
> /add 1 1 # Load everything
Act-Level Work
Be mindful of context size when loading entire acts. Often, loading summaries is sufficient:
> /add summaries # All summaries across the entire novel
> /add summaries 1 # All summaries in act 1
The /edit Command
For targeted edits that aren’t full rewrites, use /edit:
> /add 1 1 1
> /edit Change Tom's dialogue to be more confrontational
This is similar to chatting, but the command makes intent explicit and works well in agent mode plans.
Auto-Context
By default, Composez automatically identifies relevant files before each LLM call. When you mention a character or scene by name, the AI pulls in the relevant database entries and narrative files as temporary read-only context—no need to /add them manually.
This works for conversational editing but not for /write and /summarize, which use their own focused context. See Configuration for details.
Context Management Tips
The AI produces better prose when it has focused context rather than everything at once. Load the scene you’re writing plus summaries of adjacent scenes, not the entire novel.
Edit db/core/style.md to match your voice. Be specific: instead of “write well,” say “use third-person limited POV, past tense, favor short sentences in action scenes.”
The AI’s prose is only as good as the context it receives. Invest time in clear, specific summaries that capture not just plot but tone, character dynamics, and sensory atmosphere.
Compose Mode
For complex revisions, use compose mode (architect mode) with two models—one plans the edits, the other executes:
> /compose
This is useful for large structural changes where you want a planning step before the AI starts editing files.
Instructions
For recurring directives that should apply to specific tasks, use the instructions system:
> /new instruction tone-shift
Then write the instruction:
> From this point forward, the tone should shift darker.
> Elena's inner voice becomes more paranoid.
> Dialogue should have more subtext and less directness.
Load it before writing:
> /instruct tone-shift
The instruction content is injected into the chat, guiding the AI for subsequent commands.