pi-auto-prompt

pi-auto-prompt is a Pi extension package that automatically prepends global and project-local Markdown prompts to Pi's system prompt based on the active model.

It is packaged as a standard Pi package for GitHub/npm distribution and can be installed with pi install.

Install

From npm, after publishing:

pi install npm:pi-auto-prompt

From GitHub:

pi install git:github.com/zara/pi-auto-prompt

For local development:

pi install /absolute/path/to/pi-auto-prompt
# or try it for one run
pi -e /absolute/path/to/pi-auto-prompt

Prompt files

Create reusable global prompts in:

~/.pi/model-prompts/prompts/

Create project-specific prompts in any consuming project at:

.pi/model-prompts/prompts/

Then add Markdown files to either directory:

all.md              # injected for every model from that source
gpt-5.5.md          # default prompt for matched model family
[email protected]   # named variant for that model family

Resolution order:

1. Global all.md, when present and non-empty.
2. Project all.md, when present and non-empty.
3. One global model-specific prompt for the active model family.
4. One project model-specific prompt for the active model family.
5. Pi's original system prompt.

all.md is an exact reserved filename in each source. Variants such as All.md or [email protected] are ignored rather than treated as global or model-specific prompts. Discovery is non-recursive; nested Markdown files are ignored.

The extension preserves Pi's existing system prompt; it only prepends resolved prompt content during before_agent_start.

Model matching

The extension reads the active model's available id, name, and provider identities, normalizes them, and fuzzy-matches prompt filenames against those identities.

When several prompt model keys match, the longest key wins. For example, gpt-5.5-mini beats gpt-5.5, which beats gpt.

Command

The extension registers /model-prompt:

/model-prompt status
/model-prompt use <version>
/model-prompt use global <version>
/model-prompt use project <version>
/model-prompt reset
/model-prompt reset global
/model-prompt reset project

• status shows model identities, matched family, global/project directories, active global/project variants, available versions, injected files by source, and diagnostics.
• use <version> persists a project variant for the current matched model family (backward-compatible default).
• use global <version> and use project <version> persist variants for a specific prompt source.
• reset clears the project persisted variant and returns to the default project prompt when available.
• reset global and reset project clear the selected variant for a specific source.
• Running /model-prompt with no arguments opens source-specific selectors in UI-capable sessions when global and/or project sources have two or more versions for the current model family.

Selection state is stored in the consuming project at:

.pi/model-prompts/state.json

Package structure

extensions/model-prompts/index.ts   # Pi extension entry point
extensions/model-prompts/core.ts    # prompt discovery/resolution/assembly logic
examples/model-prompts/prompts/     # sample prompt files
tests/model-prompts.test.ts         # node:test coverage

The package declares its Pi resources in package.json:

{
  "pi": {
    "extensions": ["./extensions/model-prompts"]
  }
}

Development

npm install
npm test
npm run typecheck
npm run pack:dry-run

Publishing checklist

1. Replace the placeholder GitHub metadata in package.json (homepage, repository, bugs, and optionally author).
2. Commit the package files.
3. Push to GitHub.
4. Publish to npm:

npm publish --access public

5. Verify installation:

pi -e npm:pi-auto-prompt
# or
pi install npm:pi-auto-prompt