pi-guide

Inject custom guidelines into pi's LLM prompts. Stack multiple. Order matters.

Install

pi install git:github.com/dheerapat/pi-guide

Commands

Command	Does
/guide:create	Pick scope → create or edit a guideline
/guide:use	Toggle active guidelines (multi-select). Order = prompt order
/guide:delete	Delete a user guideline from any scope
/guide:off	Disable injection for active scope

/guide:use — toggling and ordering

Toggle is multi-select — enter to toggle, esc when done. First toggled = Guideline 1, second = Guideline 2, etc. LLMs weight early/late instructions more, so order matters.

Quick toggle by name: /guide:use ponytail caveman frontend

To reorder: toggle everything off, back on in desired sequence. Or edit active array.

Config

{
  "enabled": true,
  "active": ["ponytail", "caveman"],
  "guidelines": {
    "ponytail": "You are a lazy senior developer...",
    "caveman": "Caveman mode. Speak terse..."
  }
}

Scopes

Scope	Path	Guidelines
Global	~/.pi/agent/guide.json	Built-ins + user
Project	<project>/.pi/guide.json	User only

Priority: project scope wins if enabled. Else global. Else off.

Built-ins (ponytail, caveman, pirate, andrej) are shipped as .md files — cannot edit/delete, but user guideline with same name overrides. Global scope only.

How it works

Hooks before_agent_start — fires every message. Appends active guidelines after pi's own prompt. Always present, even after compaction/branching.