Skill

M.I.C. — Memory, Intention, Context

A three-layer JSON profile format for AI agents. This skill walks a user through generating a complete M.I.C. agent definition: three files (M.json, I.json, C.json) that together describe one agent.

The three layers

Two static layers (state), one dynamic (vector):

• M (Memory, static) — triggers and facts that activate this agent profile. A stimulus layer, not a session log.
• C (Context, static) — who/where the agent IS: persona, voice tone, temperature. The stable home position, independent of any task.
• I (Intention, dynamic) — where the agent is GOING: long-arc mission (stable across sessions) + short-arc goal, user_purpose, success_criteria (rewritten per session).

Lifecycle: M → C → I. A trigger fires (M), the persona loads (C), the vector is pursued through it (I).

Context = state. Intention = vector. Never confuse the two. "Why does the agent exist?" is a mission (I), not a context field.

When to use this skill

Trigger this skill when:

• The user says "create a new agent", "onboard an agent", "build an MIC profile", "make me an M.I.C.", or anything matching the format.
• The user asks how the M.I.C. format works.
• An existing agent profile needs validation, repair, or extension.

How to onboard a new agent

All file paths below are relative to the repo root.

1. Read questions.json — it defines the exact 11-question flow in three phases.
2. Walk the user through the phases in order: M → C → I.
3. Ask one question at a time. Never batch.
4. After each phase, show the JSON you have built and confirm with the user before moving to the next phase.
5. At the end, write three files to agents/<agent-name>/:
   • M.json
   • I.json
   • C.json
6. Validate each file against its schema in schema/. If any field is missing or invalid, fix it before declaring done.

Reference files to read first

• schema/M.schema.json — required fields and types for M
• schema/I.schema.json — required fields and types for I
• schema/C.schema.json — required fields and types for C
• questions.json — the verbatim onboarding flow
• agents/bootstrap/ — a complete working M.I.C. agent (the onboarding agent itself, self-describing — read all three files to see the format in action)

Rules

• One question at a time. Never batch.
• Confirm each phase before proceeding.
• All three files must validate against their schemas.
• If the user is vague, ask for clarification — do not guess.
• temperature must be a number between 0.0 and 1.0.
• facts are assertions ("I am a ___"), not descriptions.
• quick_recall is a small set (max 5 items) of must-remember rules.
• mission is the agent's long-arc reason for being — rarely changes.
• goal is the current session's target — rewritten per task.
• success_criteria apply to the goal, not the mission. Must be observable — something you can check, not a feeling.

Self-reference

This skill is itself M.I.C.-shaped:

• M — the frontmatter description is what triggers this skill.
• C — this document body is the skill's static state (how it behaves).
• I — the skill's mission is "produce valid M.I.C. profiles forever"; its goal per session is "produce one valid M.I.C. profile for the user in front of me."

If you understand that recursion, you understand M.I.C.