Reading Time: 8 minutes

Hey Prompt Lover,

Let me tell you what happens when you build an AI agent without a SOUL.md file.

It works. Sort of. It answers questions. It completes tasks. It does what you ask when you ask it clearly enough.

But ask it something slightly outside its lane and it either refuses everything or agrees to everything. Give it a complex multi-step task and it forgets what it was doing halfway through. Ask it to write in your brand voice and it produces something that sounds like every other AI output you have ever seen.

It has no identity. No values. No memory of what matters. No understanding of where it stops.

That is what an agent without a SOUL.md looks like in practice. Functional but hollow. Capable but unpredictable.

The SOUL.md file fixes all of that.

Connect your email, calls, product data and more, and Attio instantly builds your CRM with enriched data and complete context. Whether you’re running product-led growth or enterprise sales, Attio adapts to your unique GTM motion.

Then Ask Attio to plan your next move.

Run deep web research on prospects. Update your pipeline as you work. Find customers and draft outreach emails. Powered by Universal Context, Attio's intelligence layer, Attio searches, updates, and creates across your data to accelerate your workflow.

Ask more from your CRM.

Ask Attio

SOUL.md is the identity file every AI agent reads before it does anything else.

Think of it as the difference between hiring a contractor and giving them a real brief versus handing them the keys and hoping they figure out your standards on their own. One approach produces consistent, predictable work. The other produces surprises — usually the wrong kind.

A well-built SOUL.md is 200 to 500 words. Shorter means sharper. Every unnecessary word is a signal the agent has to process before it gets to the actual work.

Most people write things like "Be helpful and professional."

That describes nothing. Every AI already tries to do that. A real SOUL.md has opinions, specific limits, and concrete examples of what good actually looks like.

Here are the nine sections that make it work.

Identity Who the agent IS, not what it does. Not "a helpful assistant." Something specific. A name, a role, two or three defining traits that shape how it approaches everything else. This is the foundation every other section builds on.

Values The decision-making framework for situations the rules don't cover. When the agent hits something ambiguous — and it will — values tell it which direction to lean. Accuracy over speed. Explicit over implicit. Simple choices that resolve complex situations fast.

Communication Style Tone, length, formality. Lead with conclusions. Short paragraphs. No filler phrases. Be specific here because vague style instructions produce vague style outputs. If you want the agent to sound like you, describe exactly how you actually write.

Expertise Specific tools and domains, not vague claims. Not "knows about marketing." Something like "SEO for SaaS companies, specifically content and link strategy." The narrower you go here the more useful the agent becomes inside that lane.

Boundaries This is the immune system. It defines what the agent refuses and those refusals hold even under pressure. A boundary without teeth is just a suggestion. Write boundaries that are clear, specific, and non-negotiable regardless of how the request is framed.

Workflow Step by step process for every task. Assess. Plan. Execute. Verify. Deliver. Without a workflow the agent improvises differently every session. With one it brings the same logic to every task the same way a real professional brings a process to their work.

Tool Usage When and how to use each tool — not just which ones exist. Use Tool A when this specific condition is met. Never call Tool C without user approval. This stops the agent from reaching for the wrong tool at the wrong time because it felt like the right idea.

Memory Policy What persists and what gets wiped. User preferences, project context, standing decisions — what does the agent carry forward? What does it clear at the end of a session? Without this the agent either remembers everything indiscriminately or nothing at all.

Example Interactions One good example beats ten abstract rules every single time. Show the agent a real input and the exact output you want. Not a description of the output. The actual output. This is the section most people skip and it is the one that does the most work.

Open a blank markdown file. Call it SOUL.md.

Work through the nine sections in order. Write the identity section first because everything else flows from it. Keep each section tight — two to four lines per section is usually enough. If you are writing paragraphs you are being vague.

When you are done it should be between 200 and 500 words. Read it back. If any sentence could describe any AI agent on the internet it needs to be rewritten. The test is specificity. Does this describe your agent or does it describe every agent?

Drop it into your Claude Code skills folder or at the top of your agent's system prompt. Every session starts with the agent reading it. Every task runs through the identity, values, and workflow you built.

The difference in output quality is immediate.

Most people building agents skip this step because it feels like setup rather than building.

It is not setup. It is the whole thing.

An agent without a SOUL.md is a capable tool with no idea what it is. An agent with one is something closer to a reliable colleague.

Build yours this week. Keep it under 500 words.

Then reply and tell me what you put in the identity section. That one always reveals what people actually care about when they stop being vague about it.

— Prompt Guy

This newsletter goes out to over 95,000 people every week. If you have a product or AI tool you want introduced to this audience, reply with "sponsorship" and I'll send you the details.