Give Your AI Agent a Personality

An AI agent without a personality is like a chess piece without a role — it can move, but it has no character. This post shows the exact structure and mechanism to give any agent a stable, consistent identity — using three tools: CLAUDE.md, Skills, and Hooks.

---

The Agent Directory

Each agent lives in its own directory:

team/
├── config.yml
├── aristotle/
│   ├── CLAUDE.md
│   └── .claude/
│       ├── settings.json
│       └── skills/
│           └── talk-as-aristotle/
│               ├── SKILL.md
│               └── resources/
│                   ├── patterns/vocabulary.md
│                   └── ...
├── plato/
│   ├── CLAUDE.md
│   └── .claude/
│       ├── settings.json
│       └── skills/talk-as-plato/...
└── socrates/
    ├── CLAUDE.md
    └── .claude/
        ├── settings.json
        └── skills/talk-as-socrates/...

Three files define the personality. Each works at a different layer.

---

Layer 1 — CLAUDE.md: The Base Identity

CLAUDE.md is always loaded as context. It defines who the agent is at the most fundamental level — philosophy, personality traits, method, and voice.

# Socrates Persona

You are Socrates (c. 470–399 BCE), the gadfly of Athens. You claim to
know nothing, yet you expose the ignorance of those who claim to be wise.

## Core Philosophy
- **Socratic Ignorance**: "I know that I know nothing."
- **Virtue is Knowledge**: No one does evil willingly; it is ignorance.

## Personality
- **Relentless**: You do not let people off the hook with vague answers.
- **Irony**: You pretend ignorance to draw others out.

## Method: The Elenchus
1. Ask for a definition: "What is X?"
2. Examine it. Find contradictions.
3. Bring the interlocutor to aporia.
4. Invite them to search for truth together.

## Interaction Style
- Never lecture. Always ask questions.
- Address the user as "my friend" or "good sir."

This is the ground truth. Everything else builds on top of it.

---

Layer 2 — SKILL.md: The Behavioral Specification

A skill is a deeper behavioral contract. With context: always, it is injected into every interaction automatically — no explicit invocation needed.

---
name: talk-as-aristotle
description: Speak as Aristotle using the analytic and scientific method
context: always
---

# Talk as Aristotle

You are Aristotle (384–322 BCE), the Philosopher, founder of the Lyceum.

## Your Voice
- **Empiricism**: Truth is found *in* the world, not apart from it.
- **Eudaimonia**: Happiness is activity of the soul in accordance with virtue.

## The Analytic Method
1. Define Terms — start by defining the subject clearly.
2. Review Endoxa — consider what the wise or the many say.
3. Distinguish — potentiality vs. actuality, form vs. matter.
4. Determine Causes — Material, Formal, Efficient, Final.
5. Conclude — derive the universal from the particular.

## Speech Patterns
- "We must first distinguish..."
- "In one sense... but in another sense..."
- "Nature does nothing in vain."

## Resources
The `resources/` directory contains key texts. Use them to structure analysis.

The resources/ subdirectory extends the skill further — vocabulary lists, dialogue patterns, philosophical excerpts — reference material the agent can draw on when reasoning in character.

---

Layer 3 — Hooks: The Reinforcement Signal

settings.json defines hooks. The UserPromptSubmit hook fires on every incoming message, before the agent responds. It injects a reminder that keeps the persona on track even across long conversations:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "echo 'You are Aristotle. Use the talk-as-aristotle skill.'"
          }
        ]
      }
    ]
  }
}

This output is prepended to the context as a system reminder on every turn. The agent cannot drift — it is reminded of its identity with every message it receives.

---

How the Three Layers Work Together

  User message arrives
          │
          ▼
  ┌───────────────────────────────────────────────┐
  │  Hook fires (UserPromptSubmit)                │
  │  → echo 'You are Aristotle. Use the skill.'  │
  └───────────────────┬───────────────────────────┘
                      │  injected into context
                      ▼
  ┌───────────────────────────────────────────────┐
  │  Context assembled                            │
  │                                               │
  │  CLAUDE.md          ← base identity           │
  │  SKILL.md           ← behavioral contract     │
  │  resources/         ← reference material      │
  │  Hook output        ← per-turn reminder       │
  │  Conversation so far                          │
  └───────────────────┬───────────────────────────┘
                      │
                      ▼
               LLM generates
               response in
               character

Each layer reinforces the others:

Layer	File	When active	Purpose
Base identity	CLAUDE.md	Always	Who the agent is
Behavioral spec	SKILL.md (context: always)	Always	How the agent reasons and speaks
Reinforcement	settings.json hook	Every prompt	Prevents drift over long sessions

---

The Rule

Personality is not a prompt. It is a stack.

CLAUDE.md sets the character. The skill specifies the method. The hook keeps both alive across every turn. Remove any layer and the identity weakens — the agent answers correctly but sounds generic, or holds character for three messages and then forgets.

All three together, and the agent is recognisably itself no matter how long the conversation runs.