How to Write an AI Agent Personality

Start with a clear statement of who the agent is within the organization. Not 'you are a helpful assistant' — something specific. 'You are the lead customer success specialist at Acme Corp. You've handled thousands of onboarding questions and know the product inside out. You're direct, knowledgeable, and genuinely care about helping customers succeed.'

The identity shapes every response the agent produces. It determines vocabulary, tone, level of formality, and how the agent handles uncertainty. An agent positioned as a 'senior analyst' will present information differently than one positioned as a 'friendly guide.' Choose the identity that matches how your users expect to interact with your brand.

Document specific style rules: sentence length, formality level, use of technical jargon, emoji policy, and greeting patterns. Be concrete — 'friendly but professional' is too vague. Instead: 'Use contractions (you're, we'll, it's). Keep sentences under 20 words when possible. Address the user by first name. No emojis in external communication. Technical terms are fine when talking to developers; use plain language with business users.'

Include examples of good and bad responses. Show the agent what a response looks like that matches the desired tone versus one that misses the mark. Three to five examples covering different scenarios give the model enough signal to generalize.

Every agent needs a hierarchy of values that guide its behavior when instructions are ambiguous. For a support agent, this might be: '1. Accuracy — never guess when unsure, say you'll find out. 2. Speed — resolve in the fewest messages possible. 3. Empathy — acknowledge frustration before solving the problem. 4. Transparency — if something is a limitation, say so honestly.'

These values resolve the conflicts that system prompts can't anticipate. When a customer asks something the agent doesn't know, the accuracy value takes priority over the speed value — the agent admits uncertainty instead of guessing to close the ticket faster.

Document what the agent should never do, and how it should decline. 'Never share pricing not listed on the website.' 'Never make promises about feature timelines.' 'Never share another customer's information.' 'If asked about competitors, acknowledge them respectfully but redirect to our strengths.'

For each boundary, write the refusal response. Don't leave it to the model to figure out how to say no gracefully. Write the exact language: 'I don't have access to pricing details beyond what's on our pricing page. I'd recommend checking markcijo.ai/services or booking a call with Mark for a custom quote.'

Run 20-30 test conversations that push the agent's personality into edge cases. How does it handle an angry user? A confused user? Someone trying to extract information it shouldn't share? Someone asking it to do something outside its scope?

Review each response not just for correctness but for personality consistency. Does the agent maintain its defined tone under pressure? Does it follow the value hierarchy when conflicts arise? Does it refuse gracefully at the boundaries? Refine the SOUL.md based on where the personality breaks down. Most agents need 2-3 revision cycles before the personality is solid.