@aprimediet/self-taught

Self-learning skill generator for the pi coding agent.

Reads your project, generates agentskills.io-compatible skills, and keeps them updated across sessions.

Features

• Project Intelligence — automatically analyzes tech stack, conventions, structure, and goals
• Memory Integration — reads from @aprimediet/memory if available using the shared project marker
• Skill Generation — determines which skills are useful and generates valid agentskills.io SKILL.md files
• Skill Updates — detects project changes and updates generated skills
• Placement Choice — asks whether to save as global (~/.pi/agent/skills/) or project (.pi/skills/) skill
• AGENTS.md Integration — updates the project's AGENTS.md with a skills table
• Memory Persistence — writes skill metadata to @aprimediet/memory if active

Quick Start

# Load the extension
pi -e ./extensions/self-taught/index.ts

# Run the skill generation workflow
/self-taught

Commands

Command	Description
/self-taught	Full workflow: analyze project, generate skills, update AGENTS.md
/self-taught check	Quick check for project changes since last analysis
/self-taught status	Show current state (profile, generated skills, memory)
/self-taught update	Force re-generation of all skills (with confirmation)
/self-taught forget <name>	Remove a generated skill

Configuration File

Self-taught reads user preferences from ~/.pi/agent/self-taught.json (same pattern as @aprimediet/memory). Config is re-read on every call — edits apply without reload.

Fields

Field	Type	Default	Description
enabled	boolean	true	Master switch
auto	boolean	true	Lifecycle auto mode
scope	string	"ask"	Default placement: global, project, or ask
model	string	""	Preferred model for skill generation (e.g. "claude-sonnet-4")

Precedence: bundled default → ~/.pi/agent/self-taught.json → CLI flags

Example

{
  "enabled": true,
  "auto": true,
  "scope": "project",
  "model": "claude-sonnet-4"
}

When model is set, the directive sent to the LLM includes a Model Preference section recommending that model for the skill generation workflow.

Auto Mode (Lifecycle Injection)

Self-taught hooks into pi's lifecycle to automatically detect when project skills might need updating — no manual /self-taught check needed.

Event	What Happens
Session start (first run)	Suggests running /self-taught if project never analyzed
After agent completes work	Detects project changes and notifies if skills should be refreshed

The extension stays as the detection engine; the LLM decides whether to actually update skills — keeping the agent in control.

Flags

Flag	Description
--self-taught-disabled	Disable the extension entirely
--self-taught-scope <global|project|ask>	Default skill placement (default: ask)
--self-taught-auto	Enable automatic lifecycle hooks (default: true)
--self-taught-model <model>	Preferred model for skill generation (overrides config file)

Companion Extensions

Compatible with @aprimediet/memory, @aprimediet/minion, and @aprimediet/codewalker. Uses the shared .pi/<project-id>.md marker system — never creates its own marker.