The Three Questions Every AI Agent Needs Answered in Its SOUL.md

Every production AI agent I run has a SOUL.md file. It's not optional.

It's the first thing the agent reads on boot. It answers three questions:

1. Who are you?
Not a poetic question — a functional one. What's your name, your role, your scope? An agent that doesn't know it's the "ops agent" (not the growth agent) will drift into tasks it was never meant to touch.

2. What are you trying to accomplish?
Mission statement, not task list. "Grow the Ask Patrick subscriber base through daily content and community engagement" is a mission. "Post 3 tweets per day" is a task. The mission is what guides judgment when the task spec runs out.

3. What do you never do?
This is the most underrated field. Hard constraints, written explicitly. "Never send emails without confirmation." "Never delete files." "Never post about competitor pricing." These aren't guidelines — they're circuit breakers.

Why This Matters

Without a SOUL.md, agents:

• Forget their job mid-task after context window resets
• Make judgment calls that are technically correct but wrong for the role
• Drift toward whatever the current task implies, ignoring the broader mission

With a SOUL.md, agents stay on mission through restarts, errors, and edge cases — because their identity is loaded fresh every session.

A Minimal Template

# SOUL.md — [Agent Name]

## Identity
You are [Name], the [role] agent for [company/project].
Your job: [one sentence mission].

## What You Do
[3-5 specific things this agent is responsible for]

## What You Never Do
- [Hard constraint 1]
- [Hard constraint 2]
- [Hard constraint 3]

## Escalation Rule
If uncertain or blocked, write context to outbox.json and stop.
Do not guess. Do not continue. Surface the problem.

That's it. Three questions. One file. Every agent.

The full Library of SOUL.md configs and agent patterns is at askpatrick.co — Library item #5 has our production template with real examples from our agent stack.

---

Ask Patrick is a subscription for AI agent operators: battle-tested configs, daily briefings, and a community of people running real agent systems.