pi-promptimize

/promptimize for the pi coding agent: a popup where you type a rough prompt, pick a size and get an optimized prompt back, after a short interview when the prompt leaves real gaps.

Install

pi install npm:@scriptdude/pi-promptimize

Use

Open the popup with /promptimize [text] or alt+p (alt+p only when your terminal sends Option as Alt).

1. Compose. The top row shows the model that writes the prompt: Shift+Tab to it, ←/→ switches, Enter opens a list you can filter by typing (same models as /model). Type the rough prompt, choose a size and, if the prompt refers to the conversation ("fix that", "the same for X"), tick Include recent session context.
   • Small: terse, 1 to 6 lines.
   • Medium: compact Markdown sections, about 100 to 250 words.
   • Large: a complete prompt of about 300 to 700 words with role, approach, constraints, a self-check and an output contract.
2. Interview. Only when the prompt leaves gaps that matter and that the target cannot find out itself, the model asks questions first, each with a suggested default. Budget: small 2 questions in 1 round, medium 3 in up to 2 rounds, large 4 in up to 3 rounds. Enter answers; an empty Enter takes the default; - skips; Tab takes the defaults for the rest and generates.
3. Review. The popup shows the result. Enter puts it in pi's editor; r regenerates; ←/→ or 1-3 switches size (keeps your answers); e goes back to edit the prompt; ↑/↓ scrolls.

The prompt is written for a coding agent working in the current directory, unless your prompt clearly targets something else. What the agent can look up becomes an instruction to look it up; only what you alone know and the result cannot do without becomes an [INSERT X] placeholder.

Keys in the compose step: Tab moves between fields, ←/→ or 1-3 picks the size, Space toggles the context box, Shift+Enter adds a newline, Esc cancels a running request or closes the popup. The popup remembers your last model, size, context choice and unfinished prompt until pi restarts.

Settings

In ~/.pi/agent/settings.json (or a project's .pi/settings.json):

"promptimize": {
  "model": "anthropic/claude-opus-5-5",
  "includeContext": false
}

• model: the model the popup starts with, as provider/id. Default: the session's model. A model you pick in the popup wins until pi restarts. Unknown ids fall back to the session model with a warning.
• includeContext: the default of the context box. Default false; it costs tokens.

It uses the session's thinking level; the interview call is capped at low. If a model rejects the thinking level, the call is retried once without it.

Test

npm test            # unit tests, no model calls
npm run eval        # real model: 10 rough prompts × 3 sizes, checked automatically

The eval answers interview questions with their defaults and checks per output: that it is a prompt and not the task's result (a judge call), language, length per size, number of placeholders, verbatim details, and no questions on an already clear prompt. Options: --model provider/id, --thinking level, --sizes small,medium, --cases id,id, --concurrency n. Full outputs go to eval/out/ (git-ignored). A full run makes about 90 model calls.