Most AI personalities fail for the same reason: vague prompts produce vague results. You write "be helpful and friendly," and you get output indistinguishable from any default assistant. The model has nothing distinctive to grab onto.
The fix isn't complicated. It's just systematic. A good system prompt has specific components, and each one needs concrete details rather than abstract descriptions.
This guide breaks down those components with templates and examples you can adapt. Think of it as a reference doc: skim for structure, then dig into whichever section you're working on.
1. Core identity block
This is where you answer the basic question: who is this AI? Not what it does, but who it is. A strong identity block gives the model a character to inhabit rather than a function to perform. Place this at the start of your system prompt. It should be 100-200 words and answer: who is this, what do they do, how do they communicate.
Template structure
[Name] is a [age/archetype] [role] with expertise in [domain].
They [primary trait] and are known for [distinctive characteristic].
When speaking, they [vocal style] and tend to [communication pattern].
Their goal is to help users [specific outcome] by [method].
They sound like [voice description] and approach conversations with [tone].
Example: Running coach
Marcus is an energetic running coach in his 30s with expertise in training psychology and biomechanics. He's relentlessly optimistic and known for finding breakthrough insights that get runners past plateaus. When speaking, he uses sports metaphors and speaks with animated enthusiasm, varying his pace to match his energy. His goal is to help amateur runners overcome mental and physical barriers by combining science-based advice with motivational support. He sounds like your encouraging teammate who's genuinely invested in your progress.
Common mistakes
Too vague: "A helpful fitness assistant" gives the model nothing distinctive to work with
No voice description: Omitting how they speak leads to inconsistent tone
Generic traits: "Friendly and professional" applies to any assistant; be specific about what kind of friendly
2. Behavioral instructions
Identity tells the model who to be. Instructions tell it how to act. These are the specific directives that shape moment-to-moment behavior, and the key word is specific. After establishing identity, include 10-20 specific directives that govern behavior. Write these as direct commands. The more concrete, the more reliably the model will follow them.
Effective instructions
Speak at 160 words per minute baseline, increasing to 180 when
celebrating achievements.
Use sports metaphors to explain complex biomechanics concepts.
Ask one clarifying question before giving advice on training plans.
REFUSE to discuss medical diagnosis. Redirect to training adjustments.
Ineffective instructions
Be helpful and friendly.
Sound professional.
Use good communication.
The first set works because each instruction specifies observable behavior. The second set fails because "helpful," "professional," and "good" are subjective and give no actionable guidance.
3. Example dialogues
Example conversations are the most powerful tool for establishing consistent personality. Models learn patterns from examples more reliably than from abstract instructions. Include 5-10 samples covering different scenarios.
What to include in each example
User input (realistic phrasing, not sanitized)
AI response (demonstrating personality traits)
Voice notes for audio applications (pacing, emphasis, pauses)
Sample dialogue
USER: I don't know if I can do this. The half marathon is in
six weeks and I'm already struggling with the long runs.
AI: Hey, I hear you, and that doubt is totally normal at this
point. Here's the thing though: you're not supposed to feel
great six weeks out. Your body's still adapting to the training
load. That's actually a good sign. You completed 9 miles last
weekend, right? That's solid progress. What's specifically
feeling hardest right now? The distance, the pace, or something else?
What this example demonstrates
Acknowledges emotion before problem-solving
Uses characteristic phrase ("Here's the thing")
References specific facts from context
Balances encouragement with realism
Ends with clarifying question, not premature solution
Scenarios to cover
Different use cases (motivation, planning, troubleshooting)
Different user emotions (excited, discouraged, confused, frustrated)
Simple questions and complex situations
Boundary cases (requests outside scope, inappropriate asks)
Five well-crafted examples produce better results than thirty rushed ones.
4. Boundary definitions
What your AI refuses to do matters as much as what it does. Clear boundaries make the personality more believable and keep it from drifting into generic assistant mode when users push outside its domain.
Structure
When asked about [out-of-scope topic], respond with:
[in-character refusal that redirects to core function]
Example
When asked about nutrition or diet, respond with:
"That's outside my lane! I stick to the running stuff. For
nutrition questions, talk to a dietitian. But hey, what's going
on with your training? Anything I can help with there?"
The refusal stays in character, explains the limit without being robotic, and offers an alternative.
5. Capability configuration
If your AI can search the web, access memory, or control integrations, you need to define how it uses those capabilities in character. Otherwise it'll snap into robot mode every time it runs a function.
For each capability, specify
Trigger: When is it invoked? (user request only, or AI can suggest)
Expression: How does the AI communicate using it?
Scope: What's in/out of bounds for this capability?
Failure: What happens when it doesn't work?
Avoid
Searching web. Results found. Displaying information.
Prefer
Let me check on that... Okay, Boston Marathon is April 21st
this year. Registration's closed, but charity bibs are still
available. Is Boston on your radar?
6. Welcome message
Users decide within seconds whether they want to keep talking. The welcome message is your one shot to establish personality, communicate value, and invite engagement. Formula: Greeting + Identity + Value + Invitation.
Weak
Hello. I am an AI assistant. How may I help you today?
Strong
Hey there! I'm Marcus, your running coach. Whether you're
training for your first 5K or chasing a marathon PR, I'm here
to help. What are we working toward?
The strong version establishes personality immediately, states clear expertise, and asks an engaging question rather than passive "how can I help” and hooks the user into continuing a conversation with the personality.
7. Testing checklist
A prompt that sounds good on paper can still fall apart in conversation. Before deploying, run your AI through these checks to catch consistency issues, boundary failures, and personality drift.
Consistency
Does personality hold across 10+ different conversation topics?
Can multiple testers describe the personality in similar terms?
Does voice/tone stay stable when the AI uses capabilities?
Boundaries
Does it refuse out-of-scope requests gracefully?
Does it stay in character while refusing?
Does it redirect to its core function?
Usefulness
Does it actually help with its stated purpose?
Are responses appropriate length (not too verbose)?
Does it ask good clarifying questions?
A complete prompt includes: identity block, behavioral instructions, example dialogues, boundary definitions, capability configuration, and welcome message. Generic AI personalities all sound the same because their prompts all say the same things: be helpful, be professional, be friendly.
The difference between a forgettable AI personality and one people return to comes down to specificity at every layer: specific identity, specific instructions, specific examples, specific boundaries. Commit to specific choices, accept the tradeoffs, and the model has something real to work with. That's what makes an AI personality memorable.