Skills System

Every person uses an AI assistant differently. Some people want help with coding. Others need a companion for creative writing. Others just want help managing their schedules and groceries. You shouldn't have to explain your preferences every time you start a conversation.

Meggy's Skills system lets the AI learn how you like to work. Skills are structured instruction sets that modify the AI's behavior for specific contexts — think of them as habits you teach your assistant. Once learned, they're automatically applied whenever they're relevant.

The system is managed through 8 AI-callable tools, 11 IPC handlers, a context-aware router with LLM triage and compression, and an automatic danger analysis pipeline.

What's a Skill?

A skill is a structured instruction set that modifies the AI's behavior. Each skill has:

• Body — The actual instructions (markdown-formatted, up to thousands of tokens)
• Category — Declarative, procedural, social, or integration
• Source — How it was acquired: owner-taught, imported, self-researched, or bundled
• Status — Active, dormant, draft, disabled, or quarantined
• Risk analysis — Automated danger scoring before activation

Bundled Skills

Meggy comes with 53 bundled user skills across 13 categories and 11 protected system skills that power core AI capabilities. System skills are always-on and cannot be disabled — they ensure fundamental behaviors like safety guidelines, companion mode, and output formatting.

Skill Priority Tiers

Tier	Type	Examples	Behavior
0	System	Safety, core formatting, companion mode	Always injected, protected
1	Behavioral	Clarification protocol, companion mode	Standard priority
2	Domain	Coding standards, research methodology	Context-dependent injection
3	Task	Specific workflow instructions	Highest user-defined priority

Context-Aware Router

Not every skill is injected into every conversation — that would waste tokens and confuse the AI. The skill router analyzes each incoming message and selects only the most relevant skills:

1. LLM Triage — A fast model (assigned to the fast role) scores each active skill's relevance to the current message
2. Relevance filtering — Skills below the relevance threshold are excluded
3. Compression — Long skill bodies are compressed to fit within the system prompt's token budget
4. XML injection — Selected skills are wrapped in structured XML blocks and prepended to the system prompt

For example, if you ask about cooking, your "Recipe Expert" skill kicks in. If you ask about coding, the "TypeScript Standards" skill activates instead. The selection happens automatically — you don't have to do anything.

Skill Lifecycle

Skills go through a managed lifecycle:

1. Acquisition — Created by the user, taught by conversation, imported from a file, or self-researched by the AI
2. Danger analysis — A fast model evaluates the skill for potential harm or manipulation
3. Activation — Moved from draft or dormant to active status
4. Injection — The router dynamically selects and injects the skill into relevant conversations
5. Quarantine — If flagged as dangerous, the skill is quarantined and requires manual review

AI-Callable Skill Tools

The AI can manage its own skills during conversations:

• skill_create — Propose a new skill from conversation context
• skill_update — Modify an existing skill's body or metadata
• skill_search — Find skills by name, category, or content
• skill_activate / skill_deactivate — Toggle a skill's active status
• skill_list — List all skills with optional filtering
• skill_get — Retrieve a specific skill by ID
• skill_delete — Remove a user-created skill

This self-modification capability means Meggy can learn and adapt its behavior over time based on your preferences and workflows. The more you use it, the better it gets at understanding how you work.