Rewrite

Rewrites prompt text into a clearer, structured coding-agent prompt. The rewrite runs outside the main conversation, so the rewrite request and intermediate context don't pollute your chat or context window. The rewritten prompt is loaded into the editor for review before you send it. Trigger with /rewrite <prompt text>, or /rewrite --context <prompt text> to let the rewrite use recent visible conversation context.

Install

pi install npm:@petechu/pi-rewrite
/reload

Usage

/rewrite fix the flaky auth test and add coverage
/rewrite --context make the same change for the billing flow

Workflow

1. Start — Type /rewrite <prompt text> to send your prompt for rewriting, or /rewrite --context <prompt text> to include recent visible conversation context.
2. Preview — The model streams the revised prompt in a preview panel.
3. Review and refine — When complete, the rewritten prompt is loaded into the editor so you can edit it before sending.
4. Repeat (optional) — Add /rewrite at the start of the prompt to send it through another rewrite pass.

Behavior

• Requires prompt text after /rewrite.

• Rejects prompt text that starts with / to avoid corrupting slash commands.

• Allows slash command references later in the prompt, for example:

  /rewrite Design a plan, then use /handoff to document it
  

• Accepts --context only as the first argument:

  /rewrite --context Design the follow-up change
  

• Uses only the prompt text provided to /rewrite by default; it does not include conversation history unless --context is the first argument.

• /rewrite --context <prompt text> includes recent visible user/assistant text from the current conversation path to help match expectations.

• Context mode excludes tool results, tool-call arguments, thinking blocks, images, compaction summaries, branch summaries, custom messages, and abandoned branches.

• Context mode keeps only the most recent configured number of characters of formatted transcript when context is long. The default is 20,000 characters.

• Runs the rewrite outside the main conversation, avoiding extra context-window pollution.

• Uses the current selected model and current thinking level when supported.

• Streams model output in a preview panel while rewriting.

• Loads the rewritten prompt into the editor when complete; it is not sent to the main conversation until you submit it.

• To re-rewrite, prefix the rewritten prompt with /rewrite and press Enter.

Configuration

The extension reads rewrite settings from Pi's global agent settings and project .pi/settings.json (project overrides global):

{
  "rewrite": {
    "instruction": "Rewrite as a terse senior-engineer implementation prompt with acceptance criteria.",
    "maxContextChars": 20000
  }
}

rewrite.instruction replaces the built-in rewrite style instruction, but fixed guardrails still apply: preserve intent, do not invent facts, and output only the rewritten prompt.

rewrite.maxContextChars controls how many characters of formatted conversation transcript /rewrite --context keeps before truncating to the most recent context. It must be a positive integer. If unset or invalid, the default is 20,000 characters.

The built-in default instruction:

Produce a prompt that a coding agent can act on without needing clarification.

- State the objective explicitly, even when the user only implied it.
- Surface implied acceptance criteria when they follow naturally from the request.
- Add placeholders for missing information that would block progress. When the gap is ambiguous, leave room for the agent to ask.
- Keep the result concise. Use structure (bullets, sections) when it makes the prompt easier to scan, not as decoration.

If rewrite.instruction or rewrite.maxContextChars is invalid, /rewrite warns and falls back to the built-in default for that setting.