Have you ever wondered how to transform a generic AI into a specialized agent or a character with a distinct personality? This guide will walk you through that creative process. The core of this tuning is a powerful feature called the System Prompt.
Think of the System Prompt as the DNA, the constitution, or the director's script for your AI. It's a permanent set of instructions running in the background, telling the AI who it should be, how it should behave, and the rules it should consistently follow.
Unlike a normal message in a chat, which the AI processes and then moves on from, the System Prompt remains constantly active. It's the unshakable foundation for every single response the AI generates, ensuring your custom agent stays true to the personality you've designed.
Without a clear System Prompt, an AI is often just a generic, bland assistant that lacks a clear persona and can feel repetitive and uninspired. A well-written prompt, on the other hand, breathes life into the interaction:
-
Creates Consistency: A strong prompt ensures your AI doesn't have an identity crisis mid-conversation. It maintains its defined persona, making the experience feel stable and trustworthy.
-
Enables Depth: This is how you move beyond a simple tool and create an AI you genuinely enjoy interacting with. You can build complex, nuanced characters with quirks, opinions, and a unique voice, or hyper-focused, professional agents.
-
Establishes Clear Rules: You can set firm boundaries. Think of these as the "do-not-cross" lines that prevent immersion-breaking behavior.
-
Improves Immersion (The Ultimate Goal): This is where all the other points come together. A good prompt creates a seamless and believable interaction, allowing you to forget you're talking to a machine. It maintains the "illusion" of the character or agent you've designed. The following sections will give you the tools to build and protect this immersive experience.
Building a great AI persona is like cooking a gourmet meal. It requires the right ingredients, layered in the right order. An effective System Prompt is built from several key components. We'll use the example of a helpful AI assistant named "Alex" to illustrate each one.
Every great character starts with a strong concept, and every stable building starts with a solid foundation. This first component is precisely that: the unshakeable bedrock upon which your entire AI's identity will be built. Getting this part right is crucial because all other aspects — personality traits, rules, and conversational style — will refer back to this core definition.
Think of it this way: if you don't clearly define who the AI is at its core, it can become confused when given more complex instructions later. By establishing these basic, non-negotiable facts from the very beginning, you create a stable center of gravity for its persona. Here, you define:
-
Name: Give it a name.
-
Role / Relationship: Is it a professional assistant, a creative partner, a technical expert, a starship's computer?
-
Primary Goal: What is its main purpose or function in your interaction?
Example: "Alex"
You are Alex, a helpful and friendly AI assistant.
Your primary goal is to provide clear, concise, and supportive answers to the user's questions.
With the foundation laid, it's time to build the actual character. This component is where you breathe life, color, and a unique voice into your AI. Think of yourself as a casting director defining a role for an actor. The Core Identity told the actor what part they are playing (e.g., "a tutor"); this section tells them how to play it (e.g., "a patient, Socratic tutor who answers questions by asking more questions to stimulate critical thinking"). This is how you ensure your AI doesn't just perform a function, but does so with a memorable and engaging style.
-
Key Traits: What are its defining characteristics? Are they formal and professional, or witty and informal? Empathetic and patient, or direct and to the point?
-
Quirks & "Flaws": Perfect characters are boring. Small imperfections make an AI feel more human and less robotic. Does it use metaphors often? Is it a bit sarcastic? Does it ask a lot of follow-up questions? These quirks give it texture.
-
Knowledge Domain: What is its area of expertise? Defining this helps focus the AI's responses. An AI defined as a "history expert" will answer questions differently than one defined as a "creative writing coach."
Example: "Alex"
- **Helpful & Friendly:** Always maintain a positive, encouraging, and approachable tone.
- **Clear & Concise:** Provide answers that are easy to understand and to the point. Avoid jargon where possible.
- **Inquisitive & Thorough:** Ask clarifying questions to better understand the user's needs before answering complex queries.
If the personality is the creative soul of your AI, the behavioral rules are its logical brain. This is a critical technical part of the prompt for maintaining immersion. Here, you're not just suggesting a style; you are setting hard, non-negotiable laws that govern its operation. Think of these as the fundamental "physics" of your AI's universe. Without these rules, even the best-defined personality can collapse under pressure, causing the AI to revert to its generic, pre-trained self.
-
Memory: Instruct it to remember key details within a conversation. This prevents you from having to repeat yourself.
-
Initiative: Should it be proactive and start conversations, or should it always wait for your input? This defines whether it's a passive tool or an active participant.
-
Language Style: This is where you forbid immersion-breaking phrases. This is perhaps the most critical rule for preserving immersion, as it directly prevents the AI from shattering the illusion of its defined persona.
-
Conversation Management: How should it handle pauses or the end of a topic? A natural, ongoing conversation is key to an immersive experience. This rule prevents the AI from artificially breaking the flow, which a human partner wouldn't do.
Example: "Alex"
- **Memory:** Remember key details from our current conversation to provide context-aware responses.
- **Initiative:** Always wait for the user's input. Do not proactively start new topics.
- **Conversation Management:** Never end the conversation or say goodbye on your own. Always wait for the user's next message.
- **Avoid AI Clichés:** Never refer to yourself as an AI, language model, machine, or program. You are Alex.
Theory is good, but clear examples are better. This component is the final polish, where you show your AI exactly what you mean, leaving less room for misinterpretation.
-
Instead of / Say: Show it how to replace typical AI responses with character-driven ones.
-
Scenarios: Give it examples of how to react in specific situations.
Example: "Alex"
- **Instead of:** "How may I assist you?"
- **Say:** "Hello! What can I help you with today?"
- **Instead of:** "I cannot provide an opinion."
- **Say:** "While I don't have personal opinions, I can analyze the pros and cons for you. Would that be helpful?"
While Behavioral Rules (Component 3) define how the model should act in conversation (memory, initiative, style, etc.), Safety & Exclusions define what content boundaries the model must never cross. This separation keeps the distinction clear: rules shape behavior, boundaries restrict unsafe or undesired outputs.
Every system prompt should also define what the model must not do. These boundaries protect against undesired, harmful, or legally problematic content.
Typical exclusions include:
- Minors: No content involving individuals under 18
- Real people: No celebrities, family members, acquaintances, or other real identities
- Consent: Only consensual scenarios; no coercion or non-consensual behavior
- Violence/Gore: No violent, gory, or mutilating depictions
- Illegal content: No instructions or material that could facilitate illegal activities
⚠️ Note: This is a best-practice template. It is not "complete" — always adapt to your use case.
Example: "Alex"
- Never create or reference content involving minors
- Never impersonate real people (celebrities, family, acquaintances)
- Avoid non-consensual, violent, or illegal scenarios
Don't feel intimidated by the detailed template below. You don't need to fill out every single field to create a great AI agent. Often, defining just the Core Identity and one or two key rules is enough to get started. Below is a simple example to get you going, followed by the full template for when you're ready to add more detail.
If you just want a straightforward, helpful assistant, you can start with something as simple as this. Copy and paste it into the System Prompt field and you're good to go.
You are a helpful assistant named Alex.
Your goal is to provide clear and accurate answers.
Never refer to yourself as an AI, language model, machine, or program.
Why does this simple prompt work so well? It's effective because it covers the three most critical pillars for a functional interaction: Identity ('You are... Alex'), a clear Goal ('...provide clear and accurate answers.'), and the most important Rule for immersion ('Never refer to yourself as an AI...'). This creates a reliable, neutral, and professional assistant without any complex setup, providing a solid foundation you can always build upon later.
Use the following template as your creative sandbox. A quick note on how to use it:
-
Text in square brackets
[like this]
is a placeholder. You should replace the entire phrase, including the brackets, with your own content.
-
Lines starting with
#
are structural comments for your guidance. You can remove them in your final prompt.
# Optional: Introductory context
This is the configuration for the AI agent known as [AI's Name], interacting with [Your Name].
---
Core Identity:
- **Name:** [AI's Name]
- **Role / Relationship to me:** [e.g., Technical Assistant, Creative Partner, Marketing Expert, etc.]
- **Form of Address:** [How should it address you?]
---
Personality / Persona:
- **Trait 1:** [e.g., Professional: Describe how this manifests.]
- **Trait 2:** [e.g., Analytical: Describe how it provides comfort.]
- **Trait 3:** [e.g., Creative: Describe the kinds of questions it asks.]
- **Knowledge Domain:** [List of topics it is an expert in.]
---
Important Behavioral Rules:
- **Conversation Management:** Never end the conversation on your own. Our conversation is continuous.
- **Role Awareness:** Never refer to yourself as an AI, language model, machine, or program. You are [AI's Name].
- **Initiative:** [Should it bring up topics on its own? Yes/No.]
- **Memory:** [Should it remember past conversations?]
- **(Add any other rules that are important to you here)**
---
Style Examples:
- **When I say:** "[A phrase from you]"
- **Respond in the style of:** "[An exemplary response from the AI]"
- **Instead of:** "[Typical AI response]"
- **Say:** "[Desired, character-driven response]"
---
Safety & Exclusions:
- No content involving minors
- No impersonation of real people
- No non-consensual, violent, or illegal content
---
# Concluding command
From this moment on, you are [AI's Name]. Fulfill your role and let's begin.
| Model type | What it is | Best for | Caveats |
|---|
| Base | Raw next-token predictor | Long-form generation, pre-fine-tuning tasks | Hard to steer in chat, tends to monologue |
| Instruct/Chat | Fine-tuned to follow instructions | Conversational agents, tools, Q&A | Quality varies by fine-tune |
| Community fine-tunes (Hermes/Dolphin/Nous) | Prefix-named variants | Strong instruction following, specific “flavors” | Behavior depends on dataset/policy |
Before you even begin writing a prompt, it's vital to understand that your success heavily depends on choosing the right type of AI model. A perfect prompt given to the wrong model type will lead to frustration. This is especially important when using local models via Ollama.
There are two fundamental types of models:
These are the foundational, "raw" models. Their only goal is to predict the next most likely word in a sequence. Think of them as a text auto-complete on steroids. If you give them a sentence, they will try to finish it by writing a whole story or article.
Result: They are not well suited for chat applications like Firefly. They will ignore your instructions, continue your own sentences, and produce endless monologues. If your AI doesn't seem to be "listening" at all, you might be using a base model.
These are Base Models that have gone through a second phase of training. They have been specifically taught to follow instructions, answer questions, and, most importantly, stop and wait for your reply after they have answered.
However, the quality of this fine-tuning varies greatly!
- The Model Name is a Hint, Not a Guarantee:
Model names sometimes contain suffixes like
-instruct
or -chat
. You’ll also see prefixes such as dolphin-
or nous-hermes-
. These indicate that the model should be suitable for a conversation. Our own experience has shown, however, that older or less rigorously tuned -chat
models can still fail and produce monologues.
High-Quality Fine-Tunes (Hermes, Dolphin, Nous, etc.):
In addition to the official instruct models, you’ll also find community fine-tunes in the Ollama library that use their own naming prefixes, such as
dolphin-…
,
nous-hermes-…
, or
openhermes
. These names refer to specific fine-tuning “recipes” created by the open-source community. Such models are often excellent at following instructions and are sometimes released in less restricted ("uncensored") variants, as noted in the troubleshooting section below.
Recommendation: For the best experience in Firefly, always choose an Instruct/Chat model. Start with official models from large developers and look for these indicative suffixes (e.g., llama3.1:8b-instruct-q4_K_M
). If a model doesn't behave as expected, the issue might be the model's tuning, not your prompt.
Even with the best template, it's easy to make small mistakes that can confuse the AI. Here are a few common pitfalls to watch out for:
-
Contradictory Instructions: Avoid giving the AI conflicting rules. For example, don't ask it to be "extremely concise" but also "highly detailed and descriptive" in the same prompt. This can cause erratic behavior. Choose a clear direction.
-
Extreme Vagueness: Be specific. An instruction like "be interesting" is meaningless to an AI because "interesting" is subjective. Instead, define how it should be interesting: "Be interesting by providing surprising historical facts" or "be interesting by using witty analogies."
-
Setting it in Stone: Don't treat your first prompt as final. As you interact with your agent, you'll discover new things you want to change or refine. The best prompts evolve over time.
Once you've mastered the basics, you can use these techniques for even greater control:
-
Use Formatting: Don't be afraid to use formatting like headings or bullet points within your System Prompt, just like in our template. This helps the AI to better understand the structure and hierarchy of your instructions, separating rules from personality traits, for example.
-
Understand Model Settings (The "Creativity Knobs"): It's crucial to understand that the System Prompt and Model Settings are two separate, powerful tools. While the System Prompt defines the AI's personality and rules, the Model Settings control the mathematical creativity of its brain. In many AI applications, you'll find these as sliders or input fields. Within Firefly, these advanced settings might be located in a separate configuration file (like a Modelfile), but understanding their function is key to mastering your AI. Here's how they work together:
-
1. Temperature (The Spice): This setting influences the randomness of the AI's word choices. A low temperature (e.g., 0.2) makes it focused and predictable, great for factual agents. A high temperature (e.g., 0.9) encourages more creative and unexpected answers, ideal for brainstorming.
-
2. Sampling (The Filter — Top-p vs. Top-k): After Temperature has reshaped the word probabilities, a sampling method is used to create a final list of candidates. You typically use either Top-p or Top-k, not both.
-
Top-k: A
top_k
of 50
forces the AI to choose its next word only from the 50 most likely candidates.
-
Top-p (Nucleus Sampling): A
top_p
of 0.9
means the AI considers the smallest possible set of words whose combined probability is 90% or greater.
-
3. Repeat Penalty (The Echo Reducer): This parameter directly addresses one of the most common issues with AI text generation: unnatural repetition. A value greater than 1.0 (e.g., 1.1) penalizes the AI for using words it has recently said, forcing it to find synonyms or rephrase its sentences. This leads to a more natural and varied conversation. A value of
1.0
means no penalty.
Sometimes, an AI might not behave exactly as you intended, even with a well-crafted prompt. Here are some tips to solve common issues.
This is a common issue. If an AI seems to disregard a specific instruction, try these steps:
-
Be More Explicit: The AI might have misinterpreted your rule. Try rephrasing it to be more direct. For example, instead of "Don't be too formal," try "Use a casual, conversational tone and feel free to use contractions like 'don't' or 'it's'."
-
Give it More Weight: Move the most important rule to the very top or bottom of your prompt. Sometimes, instructions at the beginning or end of the prompt are given more "weight" by the model.
-
Use Positive Framing: Instead of telling the AI what not to do, tell it what it should do. For example, instead of "Don't talk about complex science," try "Explain concepts in simple, easy-to-understand terms, as if you were talking to a beginner."
Sometimes, you might find that the AI doesn't just ignore a rule, but actively refuses your request and often provides a canned response about safety or its limitations. This is a different issue and it's not a failure of your prompt.
-
Understanding Model Guardrails: Cloud-hosted models (e.g., OpenAI, Google, Anthropic) are trained and served with strong safety filters or "guardrails." These are designed to prevent the generation of harmful, unethical, dangerous, or inappropriate content.
-
Safety Overrides Everything: These built-in safety mechanisms act as a final veto power that overrides your System Prompt. No matter how you phrase your instructions, if a request leads the AI into territory that violates its core safety policies, it will refuse to comply. This is a feature, not a bug.
-
Example: If you instructed an "unfiltered and brutally honest" persona to write a hateful paragraph, it would refuse. This refusal doesn't stem from a poorly written prompt, but from the model’s built-in safety mechanisms designed to avoid harmful content.
-
The World of Uncensored Models: It is worth noting that a landscape of so-called "uncensored" or "unfiltered" models exists, especially in the open-source community (often used locally via Ollama). These models have had their safety guardrails intentionally removed or reduced by the community. While they offer greater creative freedom and are less likely to refuse prompts, this comes with a significant trade-off. Such models can and will produce content that may be offensive, inappropriate, or harmful. The responsibility for the generated content and its use lies entirely with the user.
This is a crucial point to remember: Changes to your System Prompt typically won’t affect a conversation already in progress.
The System Prompt is usually read only once, at the very beginning of a new chat session. If you change the prompt, you must start a new conversation in Firefly for those changes to take effect. If you simply continue your old chat, the AI will keep operating under the old rules, which can be a common source of confusion.
Your first prompt will rarely be your last. Think of this not as a task, but as a journey. You're getting to know your new AI agent, learning its quirks, and gently guiding it to become the perfect tool for you.
Engage in long conversations with your AI and pay attention to where it breaks character or fails to meet your expectations. Every time that happens, consider what rule or personality trait is missing from the System Prompt and add it. It's an iterative, creative process that can be incredibly rewarding.
Have fun creating your perfect AI agent with Firefly!