Make every model write the way you want, with consistent prose rules and styles you can switch instantly without restarting Pi.
Two layers:
- Named output styles -- switchable live, mid-session, with
/style. Includes Claude Code's built-in styles, amatter-of-factstyle mined from my own day-to-day usage, and ASD-STE100 Simplified Technical English. - Prose contract -- an optional
contract.mdappended to the system prompt on every turn, regardless of the selected style. Cycle from Claude to GPT to GLM and the register stays put.
Unlike Claude Code's output styles (which require /clear or a new session),
style changes here apply on the next prompt: the system prompt is rebuilt every
turn via before_agent_start.
pi install npm:pi-prose| Command | Effect |
|---|---|
/style |
Picker; shows the active style |
/style <name> |
Activate for this session |
/style <name> --save |
Also save as your user default |
/style <name> --project |
Also save as the project default |
/style off |
Back to default for this session |
/style off --save |
Clear the saved user default |
pi --style <name> |
Start with a style |
The footer shows the active style when it is not default. The session's
style choice persists in the session file and survives resume and fork.
Precedence for the active style: session choice > project default > user
default > default.
| Style | What it does |
|---|---|
default |
No-op; the model's normal behavior |
matter-of-fact |
BLUF ledes, length matched to the question, sentences over bullet-splatter, -- never em dash, value before mechanism, caveats stated once, no sycophancy or closing recaps |
proactive |
Executes immediately, minimizes interruptions, prefers action over planning. Never overrides explicit safety rules from project instructions |
concise |
Terse responses: leads with the result, cuts preamble and narration, 1-3 sentences for simple questions, never trades correctness for brevity |
explanatory |
Adds short "Insight:" notes explaining implementation choices and codebase patterns |
learning |
Collaborative learn-by-doing; asks you to write small, strategic pieces of code via TODO(human) markers |
ste |
ASD-STE100 Simplified Technical English: active voice, short sentences, one instruction per sentence, one word per meaning |
default, proactive, concise, explanatory, and learning mirror
Claude Code's built-in output styles. matter-of-fact was distilled from ~10k of my own
coding-agent prompts: the style corrections I kept repeating across models,
collected over months of daily work, turned into a standing instruction set.
Drop a Markdown file in either location (filename becomes the style name
unless name is set in frontmatter):
- User:
~/.pi/agent/prose/<name>.md - Project:
.pi/prose/<name>.md(loaded only after the project is trusted; wins over user styles with the same name)
---
name: diagrams-first
description: Lead every explanation with a diagram
---
When explaining code, architecture, or data flow, start with a Mermaid
diagram, then explain in prose.The body is appended to the system prompt while the style is active. Custom
styles can override built-ins by using the same name. /style re-reads the
directories on every invocation, so edits apply without /reload.
Create contract.md in the same directories to append instructions on every
turn, independent of the selected style (project wins over user):
- User:
~/.pi/agent/prose/contract.md - Project:
.pi/prose/contract.md
Use it for the rules you never want to relitigate per style or per model:
banned punctuation, lede discipline, formatting invariants. A good starting
point is the matter-of-fact style body -- copy it into contract.md and
keep /style free for task-shaped modes like explanatory or ste.
| Claude Code output styles | pi-prose | |
|---|---|---|
| Switch takes effect | After /clear / new session |
Next prompt |
| Baseline layer | None | contract.md, always on |
| Scope | Replaces system prompt sections | Append-only |
| Custom styles | Markdown + frontmatter | Markdown + frontmatter |
| Saved defaults | outputStyle in settings |
--save (user) / --project |
- Claude Code's output styles -- the origin of the style concept and of the built-in set pi-prose mirrors.
pi-output-styles-- another take on named styles for Pi/OMP, with its own style set. pi-prose differs in the always-on contract layer and the Claude Code-parity styles.
MIT