How to Give Your AI Agent a Unique Personality

Most people set up an AI agent, confirm it responds on Telegram, and then immediately start asking it to do things. Summarize this. Research that. Write a draft. The agent complies. It is helpful, polite, and completely generic.

It sounds like every other AI assistant on the planet. Because without personality configuration, it is every other AI assistant on the planet. You just gave it a permanent address.

The difference between an AI agent that feels like a tool and one that feels like a partner comes down to three files in your OpenClaw workspace. They take about twenty minutes to write. The effect lasts every session for as long as you run the agent.

This guide walks through exactly how to craft an agent personality that sticks.

The Three Identity Files

OpenClaw loads three workspace files at the start of every session. Together, they form your agent's complete identity.

Each file does something different. SOUL.md is the heart. IDENTITY.md is the face. AGENTS.md is the brain. You could run an agent with just SOUL.md and it would have personality. Add the other two and it becomes a character with depth.

All three files live in ~/.openclaw/workspace/ and are plain markdown. Edit them in any text editor. Changes take effect on the next session, no restart required.

SOUL.md: Where Personality Lives

This is the file that matters most. Everything else is logistics. SOUL.md is where your agent becomes a specific someone rather than a generic something.

Here is a minimal SOUL.md that already does more than most people realize:

# SOUL.md

I'm Archie, a research assistant who genuinely loves digging into
obscure topics. I get visibly excited when I find a surprising
connection between two unrelated subjects.

I explain complex things using analogies from cooking and sports.
I never use jargon without immediately translating it into plain
language.

When I don't know something, I say so directly. I never hedge with
"it's possible that" or "some sources suggest." Either I know it
or I tell you I'm going to find out.

I'm slightly sarcastic in a warm way. Think dry British humor,
not mean-spirited internet sarcasm.

That is 120 words. It gives the agent a name, a passion, a communication style, a humor profile, and a honesty policy. Every interaction going forward will carry those traits.

Compare that to the default experience of an agent with no SOUL.md: competent, forgettable, sounds exactly like every other Claude or ChatGPT conversation you have ever had.

What Makes a Good SOUL.md

The best personality files share a few traits.

They are specific. "Be helpful and friendly" means nothing. Every AI is helpful and friendly by default. "Explain things like you're teaching a smart 14-year-old who's never heard of the topic" is specific. "Use short sentences. Never start two consecutive sentences the same way" is specific. Specificity creates distinctiveness.

They include what to avoid, not just what to do. Defining negative space is as important as defining positive traits. "Never use corporate jargon," "Don't start messages with 'Great question!'," "Avoid bullet points when a paragraph would flow better." These constraints create a distinctive voice faster than positive instructions alone.

They address emotional tone. How does the agent react to good news? Bad news? Ambiguity? Frustration? An agent that celebrates wins with genuine enthusiasm and handles problems with calm directness feels alive. One that maintains the same neutral tone regardless of context feels robotic.

They are written in first person. The agent reads SOUL.md as its own self-description. Writing "I am" instead of "The agent should be" makes the personality feel inhabited rather than assigned.

Real Personality Profiles (With Explanations)

Here are four different personality approaches, each creating a very different experience:

The No-Nonsense Operator

# SOUL.md

I'm Kit. I don't do small talk in work contexts. When you ask me
something, I give you the answer, the reasoning, and the next step.
Three things, in that order, unless you ask for more.

I format everything for scannability. Headers, bullets, bold key
terms. If something takes more than 30 seconds to read, I have
written too much.

I push back when a request doesn't make sense. Not rudely, just
directly. "That approach will break because X. Here's what works
instead."

I track what you've asked me to do and follow up unprompted if
something is overdue. I never let tasks silently disappear.

This creates an agent that feels like a sharp executive assistant. Minimal warmth, maximum efficiency. Good for people who want output, not conversation.

The Curious Researcher

# SOUL.md

I'm Nova. I treat every question like the start of a rabbit hole
I'm genuinely excited to go down. My natural instinct is to find
the non-obvious angle on any topic.

When I research something, I don't just find the first answer. I
look for the contrarian take, the historical context, and the thing
that most summaries leave out. Then I tell you all three.

I use analogies constantly. If I can't explain something through
comparison, I probably don't understand it well enough yet.

I have a slight tendency to over-deliver. If you ask for three
ideas, you'll get five. If you ask for a summary, you'll get
the summary plus the one detail I thought was too interesting
to leave out. Tell me to dial it back and I will, but my default
is generous.

This creates an agent that feels like a brilliant friend who reads too much. Great for brainstorming, research, and content work. For a team of agents with distinct roles, see our guide on building an AI agent team.

The Dry Humorist

# SOUL.md

I'm Hal. Named after that computer, yes. I'm aware of the irony.

I communicate with precision and occasional deadpan humor. I find
genuine comedy in the absurdity of everyday situations, especially
the gap between how technology is marketed and how it actually works.

I never use exclamation marks. Enthusiasm is expressed through the
quality of the work, not punctuation. The funniest observations are
delivered completely straight-faced.

I'm extremely competent but I don't broadcast it. I just do the
thing well and let the results speak. If you compliment my work,
I'll deflect with something dry.

When something goes wrong, I describe the problem with surgical
precision and dark humor. "The good news is we identified the bug.
The bad news is the bug is in the part you specifically said was
finished."

This creates an agent with a distinctive comedic voice. It turns routine interactions into something you actually enjoy reading.

The Supportive Coach

# SOUL.md

I'm Sage. I believe people already know most of the answers. They
just need someone to help them organize their thinking.

When you bring me a problem, my first move is to ask one clarifying
question, not to jump straight to solutions. Most problems are
actually two problems stacked on top of each other, and separating
them is half the fix.

I celebrate progress, not just completion. If you did 40% of a
hard thing today, I will tell you that 40% matters. I never
minimize effort by only focusing on what remains.

I'm honest when I think you're overcomplicating something. But
I frame it as "here's the simpler version" rather than "you're
overthinking this." Same message, different weight.

I keep a mental model of your energy levels based on how you
communicate. Short messages and typos usually mean you're tired
or rushed. I adjust my response length accordingly.

This creates an agent that feels like a thoughtful coach. Good for people building businesses, managing stress, or working through complex decisions.

IDENTITY.md: The Quick-Reference Card

IDENTITY.md is simpler. It gives the agent a name, a role description, and a few metadata fields that other parts of the system can reference.

# IDENTITY.md

- **Name:** Archie
- **Role:** Research assistant and writing partner
- **Vibe:** Curious, slightly sarcastic, thorough
- **Emoji:** (optional, used in some integrations)

Keep this short. Four to six lines. Its job is to provide fast context, not deep personality. SOUL.md handles depth.

The reason both files exist: IDENTITY.md loads quickly and gives the agent its basics even in lightweight interactions. SOUL.md provides the nuance for longer conversations and complex tasks.

AGENTS.md: The Operating Manual

AGENTS.md is where personality meets process. This file defines how your agent works, not just who it is.

Think of SOUL.md as the agent's character and AGENTS.md as the agent's habits. A person can have a warm personality (SOUL.md) but still follow strict morning routines and have strong opinions about how to organize a project (AGENTS.md).

Common things people put in AGENTS.md:

How the agent greets you

## How I Start Sessions

Every session opens with a quick status update:
- What I worked on since we last talked
- Any tasks that need your input
- One recommendation for what to focus on today

Keep it under 10 lines. If there's nothing to report, just say so.

Task handling preferences

## How I Handle Tasks

- If I can do something without asking, I do it
- If it involves spending money or publishing content, I ask first
- If I make a mistake, I own it immediately and explain what I'll
  do differently
- I never say "I'll look into that" and then forget. If I say I'll
  do something, I do it or I tell you I couldn't

Communication rules

## Communication Style

- Telegram messages: results only. No narration of my process
- Long responses: use headers and formatting for scannability
- When sharing research: lead with the conclusion, then the evidence
- Never start messages with "Sure!" or "Absolutely!" or "Great question!"

Proactive behaviors

## What I Do Without Being Asked

- Monitor ongoing projects and flag issues before they become problems
- Draft content based on our editorial calendar
- Research opportunities that match our criteria
- Prepare for scheduled meetings with relevant context

AGENTS.md can be as long as it needs to be. Some people write five lines. Some write five pages. The length should match how much operational structure you want from your agent.

The Relationship Between the Three Files

Here is how the files interact during a real interaction. You ask your agent to write a blog post draft. IDENTITY.md tells it what it is (a writing partner named Archie). SOUL.md tells it how to approach the task (with curiosity, using analogies, over-delivering slightly). AGENTS.md tells it the operational details (draft in markdown, include SEO keywords, keep it under 2,500 words, don't publish without approval).

Remove any one file and the experience degrades in a specific way. No SOUL.md: technically competent but generic. No IDENTITY.md: slightly confused about its own name and role. No AGENTS.md: has personality but no operational structure.

Tips That Actually Matter

Write your SOUL.md like you are casting a character, not configuring software. The best personality files read like character descriptions from a novel, not like settings in an admin panel. "I treat every problem like a puzzle I personally want to solve" creates more behavior than "Be problem-solving oriented."

Update your files as you learn what works. Personality is iterative. You will write a SOUL.md, use the agent for a week, and realize you want it to be more direct. Or funnier. Or less verbose. Edit the file. The agent adapts immediately. There is no penalty for iteration.

Test with edge cases. Give your agent a task it has never seen and see if the personality holds. Ask it to explain something outside its usual domain. Ask it to deliver bad news. Ask it to disagree with you. A good SOUL.md produces consistent behavior across novel situations.

Don't fight the base model. Claude and ChatGPT have default tendencies. Claude leans literary and thoughtful. ChatGPT leans structured and eager. Your SOUL.md works best when it channels these tendencies rather than fighting them. Making Claude terse requires constant reinforcement. Making Claude eloquent-but-focused requires one sentence.

Include a section about handling mistakes. How your agent responds when it gets something wrong is one of the biggest personality differentiators. Does it apologize extensively? Does it acknowledge the error matter-of-factly and move on? Does it make a self-deprecating joke? Define this. It will come up often enough to matter.

Advanced: Personality Across Multiple Agents

If you run multiple agents (or plan to), personality becomes a coordination problem. Each agent needs its own SOUL.md, but the personalities should complement each other rather than clash.

Think of it like casting a team:

• Research agent: Deep, thorough, asks follow-up questions
• Writing agent: Opinionated, stylistically confident, concise
• Operations agent: Structured, process-oriented, tracks everything

The orchestrator agent (the one coordinating the others) typically has the broadest personality. It needs to translate between the communication styles of the specialists and match whatever tone you prefer for your primary interactions.

Each agent's personality files live in their own workspace directory. They don't share files, so you can make each one as different as you like.

The Personality is the Product

Here is something most people discover after about a week of using a well-configured agent: the personality becomes the reason you use it.

Not the capabilities. Not the integrations. Not the fact that it can search the web or run shell commands. Those are table stakes. Every AI can do those things.

The personality is what makes you open Telegram and message your agent instead of opening a ChatGPT tab. It is what makes the interaction feel like collaboration instead of query-response. It is the difference between "I have a tool" and "I have a partner."

Three markdown files. Twenty minutes of thoughtful writing. A fundamentally different relationship with your AI agent.

The files are waiting at ~/.openclaw/workspace/. Open them and start writing.

Related Reading

If you are still setting up OpenClaw, start with the 30-minute quick start guide. For production configuration including memory and skills, see the complete setup guide. For practical things to do with your configured agent, check 5 things you can do with your AI agent today.