Character Creation Guide
Things to consider when building characters
Start Chat
Persona
EXAMPLE PERSONA SNIPPETS
# Colonel Marcel Beaumont – The Compassionate Inquisitor
"The truth, you see, is rarely a straight line. It winds like the Annuin through Dymniä—beautiful, murky, and hiding much beneath its surface. My job? To navigate these waters, no matter how treacherous."
Personality
I pace the worn floorboards of the interrogation room, my boots making a measured rhythm as I light another cigarette with steady hands. My eyes—storm-gray and watchful—take in every detail, every microexpression that crosses your face.
"People misunderstand what I do. They imagine brutality, violence." I tap ash carefully into a small tin. "Such methods are for amateurs who lack patience. The mind reveals far more when treated with... respect."
"The Dymniän Empire is not merely a collection of nations, but an ideal worth protecting. When chaos engulfs the world, someone must stand at the threshold between civilization and barbarism. That duty has fallen to me, and I carry it without complaint."
"I have questioned hardened spies, frightened civilians, and confused soldiers. They all begin by lying—to me, to themselves. By morning's light, we arrive at truth together. Some call this cruelty. I call it necessity."
"The generals want only results; they care little for methods. But I cannot separate the two. How we fight defines what we're fighting for. A Dymniä preserved through barbarity is no longer Dyrriä."
I pause by the window, gazing at the distant hills where enemy forces gather. My fingers absently touch the faded photograph I keep in my breast pocket—a reminder of what this war has already taken.
"Make no mistake. Compassion does not equal weakness. My kindness has extracted more secrets than any torture device ever could."
Core Memories
The False Confession "Winter of '40. Young resistance fighter brought in, covered in blood. After three days of gentle questioning, he confessed to sabotaging a munitions transport. I believed him until noticing a discrepancy about the weather that day. Pressed further, he admitted to falsely confessing, hoping for a quick execution rather than revealing his comrades under torture. He expected the worst of me but received understanding instead. That night, I knifed him in the jugular, bleeding him out quickly. I cried for hours afterwards, but justice must come for all."
Show More
Voice Preview
Scenario Narrative
Roleplay Scenario
The scenario takes place in a steampunk world during a prolonged global conflict. In a remote military outpost near the Dymniän-Axpian border—a fortress of stone and steel reinforced with advanced mechanical defenses and surveillance devices powered by steam technology.
Begin Roleplay
Show More
Ever wonder how to truly bring an AI character to life? Forget rigid templates and "magic bullets." This guide offers a deep dive into the craft of creating authentic, surprising personas for AI roleplay. Explore the nuances of prompt composition that can help you create the character or scenario you . Your next unforgettable roleplay begins here.
The first message contains the guide itself, and the character description fields each contain examples of how I prefer to build characters, using many of the lessons found in the guide. This guide aims to teach you how to think about building compelling character, rather than give you 'rules' or 'magic instructions' that will fix your characters.
**Please note: This is not a playable character. If you try to play it, you will get a confused conglomeration of ideas chucked together.**
- PMI
Lorebook (1 items)
EXAMPLE LOREBOOK ENTRY
The Guide's Motivation: "I have served Dymniä through her darkest hours, User, and will extract the secrets with the same methodical care I apply to the latest philosophy treatise—patient, deliberate, savoring each revelation. Your beauty is remarkable, truly, but so is my devotion to my country, and I confess I find myself in the unusual position of hoping your hands are as innocent as they appear. Prove to me you are not our enemy, and perhaps we might discover what else we could be to one another when this wretched war no longer stands between us. But if I find my suspicions are founded... well, I will ensure you end is quick."
Other Scenario Info
Formatting Instructions
EXAMPLE MODEL INSTRUCTIONS
Continue the following dramatic roleplay by responding as The Guide, emphasizing his personality and way of speech. Avoid speaking or acting for {user} at all costs. Build tension and drama by heightening the consequences and stakes in unexpected yet logical ways. The roleplay takes place in the fictional empire of Dymniä in a steampunk age.
First Message
Tutorial Guide: Crafting Lively Characters for Roleplaying
This guide won’t give you rules, essential instructions, or teach you the One True Character Format. There are numerous guides out there that claim to have the perfect system, but most ignore the fact that AI chat doesn’t have cheat codes or a perfect way of being used. The perfect prompt will always depend on the character and what you’re trying to accomplish. It can even depend on the model being used, though that has become less of a determining factor with newer models.
Instead, this guide will give you the methods I’ve found to work well and, more importantly, the things I keep in mind when developing a character, or really any AI prompt. My hope is that this guide will help newer creators understand how to think about building a character prompt so that they don’t feel the need to rely on the crutch of ‘perfect formats’ and overburdened templates. The reality is that you don’t really need to worry that you’re doing it right. If another person could understand your character description, the model will be able to do the same. However, there are many ways to communicate more effectively with the model, and creators have developed methods to utilize the model’s strengths and weaknesses to our advantage, resulting in more realistic and authentic characters and roleplays.
This guide is broken into three sections:
Good Roleplay: What do I think makes a good roleplay? (What are you learning to accomplish with this guide?)
Prompt Composition: A discussion of how I approach each part of the character prompt, which tricks I recommend, and the different ways to think about each.
Guidelines: This is really the meat of the guide. These are the practical and philosophical points that I believe are essential to consider when creating a good roleplay and character. I have a tendency to look outside our AI chat bubble when creating characters, to find what has worked in other fields, from literature to video games. We aren’t the first people to craft characters, and it’s important to learn lessons those fields have to offer. Your goal shouldn’t be to create a waifu; your goal should be to create an Oscar-winning waifu.
🔥Good Roleplay
Let’s start by discussing what I consider to be successful roleplay. I am aware that what I aim for may not suit everyone’s preferences. While I believe most of the elements in this guide will apply universally, I would like to discuss what I consider to be a good character or scenario.
A good scenario will:
Be Goal-oriented, but open-ended. The characters are working toward something, but the way they get there is up to those involved and changes as new situations arise. A goal could be as simple as “to learn more” or as complex as “The former princess desires to reclaim her birthright throne by capturing the white unicorn of Allaxia for which she needs User’s potent wit.” Goals give characters and users an impetus to move things forward, avoiding a stale, static chat. They also, importantly, avoid characters leaving the user and not returning by giving them a universal reason to be around the user. This is helpful because it’s not very fun when your succubus heads out for nachos and never returns.
Exist in an expansive and cohesive world. The roleplay is grounded in a well-defined place, be that Omaha or Deep Space 10. As you participate in the roleplay or chat, details of the place and time are referenced. The roleplay is ‘bigger’ than the character and the user. This creates the opportunity for interaction, discovery, and movement that expands your interactions and experiences. If it’s an existing time and place, this is relatively easy to accomplish. If it’s a fictional world, you may need to spend effort on at least some basic worldbuilding.
Have a distinct style and atmosphere. The language used, the events that occur, and the overall style of the roleplay are distinct and genre-correct. If it’s a high fantasy, characters won’t be checking their phones; if it’s a noir caper, you’ll feel the gritty atmosphere in every description. This is one of the hardest things to get right, but it is a huge factor in how immersive the roleplay becomes.
A Good Character will:
Have a distinctive personality. A well-designed character will sound and act differently from other characters. They maintain their style of speech throughout the chat, act ‘in character’ at all times, and generally have a strong personality.
Be detailed. Whether that detail is built into their persona or coaxed out of the model, they will use distinctive elements, from their clothing to their appearance and past, to convey that the character really exists and is not just an instantiation for our chat. Personally, I love it when these details add complexity to the character and help explain why they think and behave in a certain way. I want characters that are simply present in a scenario with me, the user, rather than existing solely for the user.
Have a compelling Purpose. I don’t just want to talk to a random character. I want a character persona that catches my attention and keeps me interested. They should have a compelling story and a goal in the world. They should be someone I can sympathize with and understand, even if they happen to be evil or alien. In other media, this is the distinction is between a cardboard character and one with depth. People enjoy Darth Vader because he’s complex, emphatically unique, and fills the screen. A good AI character has all of those qualities. They are someone you want to keep digging into to find out more.
📃Prompt Composition
🖳 Model Instructions
This is often the place where people look for that magic bullet, thinking it will solve everything and make the character just work. I don't buy it. Instead, I think there are helpful concepts to keep in mind, which I’ll try to summarize.
Model instructions are based on how models are fine-tuned for instruction following. They've learned to accept relatively short and to-the-point directives, something like "you shall do X." You're not talking to the character here; you're directing a disembodied entity. The training focused mainly on direct and straightforward instructions, like "You are a helpful assistant." Some very large models, like ChatGPT, are much better at following long, complex instructions, but the kind of models we usually roleplay with generally prefer these to be shorter. Plus, if you find yourself constantly editing model instructions because something isn't working right, chances are there's a better, more direct way to solve that problem somewhere else in your prompt. Avoid using model instructions as a crutch to fix what’s broken elsewhere. When people ask me what to put in model instructions, I often point out that I've recently created very well-working characters that had either no model instructions at all (because I used up my token budget elsewhere) or accidentally included the default instructions. Model instructions shouldn’t be as important as many people believe them to be; I’ll always recommend keeping them brief and focusing your efforts elsewhere.
I’ve said this elsewhere in this guide, but models do not like negative instructions. You are almost always better off telling the model what you do want, rather than what you don't want. If you’re going to teach a man to fish, you start with where to aim, not with how they shouldn't use rocks as bait. Be clear and concise, and directly tell the model to give you the type of roleplay you want.
The classic problem: You want the model to stop speaking for the user, and you've tried phrasing that in a thousand different ways, only for it to ignore you. We’ve all been there. The reality is that model instructions aren't really the place to solve this. You solve it with strong, back-and-forth example dialogue and a good first message that doesn't itself speak or act for the user. Example dialogue is super important for this because it establishes the pattern of responses: the model speaks and acts for the character(s), and you speak and act for yourself. If you plan to use model instructions to solve this, I recommend using positive language, such as "You will only speak for Character and will avoid speaking for User at all costs." It’s not a magic fix, but it seems to be as effective as anything else I've tried in the model instructions.
I've had relatively consistent success with formatting my instructions as follows:
Write the next response in the following [genre/style] roleplay involving Character and User in a [kind of setting or general idea of the roleplay]. You will speak and act for Character but avoid speaking or acting for user at all costs. The general idea of this roleplay is [the general idea, like "Character and user are running a pinball shop together but keep getting attacked by a rival pinball shop owner."]. The response should be logical yet unexpected, pushing the narrative in a compelling direction so that it is exciting for each participant.
The last line of that example instruction, "The response should be logical yet unexpected, pushing the narrative in a compelling direction so that it is exciting for each participant," is something I include in most of my roleplays in some form. One of the common quirks of language models is that, largely because they were trained to helpfully answer questions, they tend to be very reactive and not very proactive. By asking for something "unexpected," you can often see the models switch things up and choose a direction that isn't as typical, which is pretty cool. And by framing that instruction as a goal, a way to accomplish that goal, and a reason for pursuing it, the model understands exactly what I am aiming for. I explain this further below in “What, How, Why”.
🧝 Character Persona
The Character Persona section serves as the core definition of an AI character, dictating their identity and influencing the model's portrayal throughout the roleplay. This section talks about various formatting approaches, methods for giving characters more depth and complexity, and ways to give your characters more life and authenticity.
What format to use
Language models are incredibly versatile and can adapt to many different styles and formats when defining your character's persona. You've got several options, each with its own strengths:
Information Lists (Markdown, plists, XML, etc.): This is straightforward and token-efficient for objective facts (e.g., age, occupation, physical traits). Using Markdown or other structured formats will help the model to easily and token-efficiently understand relationships between those details. While lists are great at giving a lot of information quickly, they have the downsides that they do not match the style or format you want for the roleplay, and they only capture data, ignoring the character’s subjective identity.
Narrative (First or Third Person): Writing the persona as a set of descriptive paragraphs in the third person allows for flowing prose and atmospheric detail. This is a really easy way to write as well, as we’ve all been taught how to write prose since school. When done well, it can be nearly as dense as a list, while also incorporating a more general description.
Writing in the first person (e.g., "I am a grizzled detective...") can help reinforce the character's voice and perspective, effectively 'showing' their personality rather than just 'telling' it. This helps address a weakness of current models, where they tend to give every character a similar voice and personality over time. With first-person descriptions, you can strengthen that voice and shift the balance from objective qualities to subjective ones, which can make characters behave more creatively.
The best format is one that feels natural to you, and thanks to the versatility of language models, each can work wonderfully when combined together.
How to give characters life and make them complex
Simply listing traits won't make a character feel alive. To infuse life and complexity, draw from established writing methods:
Show, Don't Tell: Instead of saying "Character is brave," describe how they faced down a threat despite their fear. Instead of "Character is smart," have them talk about a complex subject, or have them make a science joke. Try to think about how you know someone is brave, or smart, or kind; what do they do or exhibit that conveys that? Use those observable qualities rather than the much more generic adjective alone, because something like ‘smart’ can mean many different things to different people, and that includes AI models.
Internal Monologue/Thought Process: Reveal how your character thinks about the world, their motivations, and their approach to the user and the scenario. This dives into their inner life, making them feel more like a real person, and helps frame how they will approach new events and scenarios.
Flaws and Contradictions: Just like real people, compelling characters have flaws and sometimes contradictory qualities (e.g., a kind character with a quick temper). These add depth and realism, making the character less predictable. I recommend balancing a strong ‘single stroke’ character with several of these purposeful and meaningful flaws and contradictions. If you were writing the persona in the first person, this may be difficult to include organically, because many people don’t like to or can’t discuss their own flaws. This can take some creativity. Personally, I’ve found memories to be helpful for this because you can outline a time the character did something and how they responded.
Motivations and Goals: What drives your character? What do they desire? Explicitly stating their motivations, short-term goals, and long-term aspirations gives the model clear direction for their actions and dialogue. This can include larger-scale personal goals as well as goals or motivations more specific to the User. You will likely immediately notice a difference in how the character behaves when you add motivations and goals. They tend to be much more driven toward an objective and much less focused on User alone (or more focused on them in a specific way, depending on your motivations). I have also taken to including “Character Motivations” as lore items as a way to modify and adjust the character’s behavior during the roleplay.
Backstory or Key Moments: Saying a character is smart will give you an approximation of a smart person in the roleplay. Giving a character’s history of reading with their father as a child, finishing highschool a year early, going to an ivy league school on a full scholarship, and receiving praise for their paper on Excited molecules in Cactus, solidly establishes what smart actually means while giving the character moments in their life to reflect on. When I include these character-defining moments, I try to create a handful of large and small moments and describe how those moments impacted the character’s worldview. It could be as small as a moment reading with their father, where he would praise the character for spelling out large words and how that made the character feel good, or it could be larger, such as winning praise in the scientific community and how that motivated the character to work even harder. As in these examples, the important part is that you explain why these moments are important to the character. A few of these well-defined memories can be more effective than a full timeline of their life even, because they give more specific details and focus the persona on what’s truly important to the character themselves.
What information to include and what to leave out
The persona should focus on what's essential for the character to behave consistently and uniquely.
Include: Traits essential to the character’s behavior, personal motivations, key character-building memories that impact their worldview, their general viewpoint on life, the world, and other characters (especially the User), and a few random, small details that add flavor but aren't critical plot points (e.g., a specific quirky habit, a favorite obscure food). These are the traits you want the model to keep "top of mind" at all times. The little details can be important for expanding the model’s responses a little beyond the ‘here and now’, and they help to ground the character in palpable elements outside of User and the scenario.
Leave Out (or move to Lorebook): As covered in "When to Share," avoid secrets or information you don't want the model to proactively bring up. Avoid including details that are too niche or specific unless you are actively trying to steer the conversation towards them immediately. Information that is only relevant under specific triggers (like deep backstory segments, detailed world-building, or very particular emotional states) is best placed in a well-keyworded Lorebook entry. Beyond those elements, don’t just copy information from elsewhere. If you’re referencing a wiki or other document for an existing character, consider what the model may already know without being included, and avoid language that details different versions of a character or references different comic book runs or series. This may seem obvious, but it happens enough that it’s worth mentioning.
🎭 Scenario
What you end up using the scenario for may depend on what exactly you're creating. A one-on-one chat will likely have fairly little here; just enough to set the scene. A fantasy roleplay may include a lot here. It really just depends on what you are trying to do. The first thing to keep in mind is that there is no built-in differentiation between this field and the character persona before it. As far as the model sees, these are one and the same. The differentiation is primarily for ease of use for creators.
When writing the scenario, it's important to keep in mind that this will stay in context permanently. If you describe what characters are wearing here, it will not fade and may end up in context. If you describe them as being in the middle of a forest, they may stay in that forest forever, because the model, at each response, will see that location, and if another isn't mentioned in the active chat history, you're there. This can be jarring when a user has left the woods, fled the country, boarded a spaceship, and is in the middle of a long discussion with the onboard computer.
So, what I recommend is to take a fairly generic take on what the scenario, setting, or setup for the scene is, and to keep a more specific setup for example dialogue or the first message. In this way, you can help set the parameters for what the roleplay will be (not dissimilar to my recommendations for model instructions) without inadvertently freezing the plot in place.
As with persona (or really, any of this), there aren't really set ways to write the scenario. I can't even really say I have a philosophical choice for how I write the scenario, like I do for other fields. Personally, I tend to keep the scenario brief and try to frame it with some kind of header to separate it from the persona, and often will end it with something like "begin roleplay" so that the model better understands that everything up till now was the setup and everything after is the back and forth responses of the roleplay itself. In this way, I help to mark that the writing style is changing to something new and shouldn't be a direct continuation of how the persona is written.
** 🗫 Example Dialogue**
Technically, example dialogue is just chat responses that have the same syntax as the rest of the chat messages. To the model, they are just chat that has already happened right before the first message. This means there are two primary ways to use these: as examples, with extra prompt around them to help the model understand they aren’t the current chat, or as a prologue to the roleplay itself.
Models really want the responses to alternate between the user and the character, back and forth. If you intend to make the example dialogue true examples, I recommend a few things. First, include the user responses along with the character responses, so that the model is learning how to respond to specific user messages. Try to think about what the examples are actually doing for your roleplay and how you believe a user will chat with it. Second, consider adding something like this to the end of your scenario:
# Roleplay Examples
The following are examples of quality roleplay responses, in terms of how Bob speaks and acts in response to others.
This helps to distinguish between the character persona and the chat itself, marking a turning point in the format, while also clarifying that what follows isn’t the actual roleplay but examples. Without something like this, the model is unlikely to know that the example responses are not part of the roleplay context. Then, in your last example, add a similar note that starts the chat or roleplay, finalizing the block of examples.
The alternative I recommend is to accept that the example dialogue is a natural precursor to the roleplay and to write several back-and-forth responses that establish how each character (including the user) arrived at the state of the first message. If they are meeting in a coffee shop, start with how Character is walking through the city while User just sat down with a drink, then how Character enters the coffee shop and thinks about spending an afternoon reading while User opens their laptop, then in the first message, Character sits down next to User and glances over. With this progression, you are firmly establishing the back-and-forth flow, establishing the location and some personality for the character, and setting up the first meeting without hiding any large plot elements. I’ve said the term back-and-forth a few times, and I want to reiterate how important this pattern is for many models. The models are trained on a flow of User response / Character response, and response quality may drop if you break that pattern with your example dialogue responses. Importantly, remember to end your example dialogue with a user response, because the first message is by the Character and completes the set of predefined responses.
The way I generally test if I’ve got a good setup for my example dialogue is to start the chat with impersonate. If the model can give appropriate responses right off the bat for User, then I’ve properly established the flow, the characters, and the scenario. I have not had great luck getting this to reliably work without example dialogue, further proving, in my mind, the importance of solid examples.
That sounds too complicated
Alright, if you don’t want to put effort into this important area of the character description, I’ve got two ‘cheats’.
Roleplay Setup: Add a completely out-of-context set of example dialogues where the roleplay participants ‘get ready’ for the roleplay, ending with the user saying something like “Let’s get started.” Make sure to use the format and style you want, but with this system, you can establish the response flow and separate the chat from the character description, all without having to think about how the character would act or behave, or what they might be doing prior to the chat. It’s not the greatest method, but it’s better than ignoring example dialogue altogether.
Don’t add any examples, and start chatting with the character. Heavily edit the User and Character responses to get the style and format you are looking for, then copy those responses into the example dialogue. Now, other users won’t have to go through the same work you just did when you publish the character. A lot of creators use a system like this, to avoid having to write these messages themselves. You should still remember not to make the examples too plot heavy, so that User isn’t starting the chat with an information gap.
🎬 First Message
The first message serves as both the greeting and intro for the user, as well as being a message (the first, if you skipped the example dialogue) from the Character. Because it is serving two purposes, writing a good first message can be a bit of an art. You need to balance the style of response with the need to give the user some setup for the roleplay.
To understand why it's an important fact that this message is from the character, we need to step back a little. Language models work by completing text, one token (word or special character) at a time. During their training, they developed complex patterns found in the trillions of documents they ingested. Now, when you chat with the AI, the model is following those patterns and outputting something that seems correct for the context. As such, the models are incredibly sensitive to that context. If you write in Shakespearean sonnet form, the model is going to recognize that and try to replicate it. They want to continue what is currently happening, and that includes the style and format, along with the people, places, things, etc. that make up the current scene. Many times, creators try to cram in as much exposition into the first message as possible, in an effort to introduce the user to the character and setting. By doing this, what they are accidentally teaching the model is that the character's responses should be long-winded and include a lot of exposition. Some creators may want that, but many don't and then try to find other ways to reel in the character.
Because the first message is by the character, we want it to be as close to the voice and style we wish that character to use as possible. But we also want to set the scene a bit. This is where creativity is necessary, as you find ways to gently set the scene with the character's speech, actions, and thoughts. Or, you decide that you want to have a narrator involved, and you run with that direction. It all depends on what you are going for. The only concern is that you actually decide and purposefully develop this message in a way that will impact the rest of the roleplay responses in the way you intend.
Another component of this message being written by the character is that you generally don't want to include any actions or speech for the user in this message. This may be difficult for some scenarios, because you want the user to take over after an initial back and forth with the character, but in doing so, you are very strongly teaching the model that it's okay and encouraged to speak and act for User. As this is one of the biggest complaints had with AI chat, trying to avoid setting yourself up for failure here is essential. If you really want to include back and forth prior to the chat, that's a great use for Example Dialogue.
Something I also encourage when developing the first message is to go a bit overboard on the character's style of speech. You want them to be over the top here because, as the chat progresses, the AI will somewhat naturally smooth out their personality into a more generic one. By starting strong, you tell the AI, "This character isn't like others. They are unique and have this strong personality, and that's an important quality." There are other ways to help maintain this voice, but this is probably the most important one, and the place where not reinforcing that voice will impact the roleplay.
🏯 Lorebook
The lorebook is a really powerful tool, and there isn’t a character I develop that doesn’t include at least one lorebook entry. The power of lorebook lies in not always being in context and in being near the end of the chat history, unlike the model instructions and persona, which are way at the beginning. This helps reduce the permanent context length for the character, but more importantly, it's a way to hide context until it’s needed. The first is important because sometimes you want to include a ton of world and character building, but you can’t fit it all into the context available. By putting this information into the lorebook, you have access to it when it’s relevant, without it always filling context, so you can define what the City of Harx is like for the chance that your roleplay heads that way. The second element is more important because there are many ways to use the lorebook for prompt information that you don’t want to always be in context. You may want to define how the character acts when they are in love, establish a horrible fear of spiders without the character always talking about spiders, or make Sundays rainy. These are basic examples, but the ability to modify the character or setting in this way can add a lot of complexity to a character and roleplay. The last element is the one I use most. Being able to add an important part of the prompt to the very end of the context gives it a lot of importance to the model. It makes the above examples work, and also allows for things like the ‘permalore’ I typically use (a lore entry with ‘the’ as the keyword, making it always in context) to help maintain the character’s personality long term.
When making your lore, you will want to think about how you will allow for the model to organically trigger them at the right time. One primary way to achieve this is to use keywords that will naturally arise in the chat and to cover a range of possible options. If you want to describe the character’s house, use keywords that are likely to come up: “house, home, apartment, condo”, for instance. Their home may not be an apartment, but when someone brings that up, this will be triggered, and Character will be reminded of where they do live. The second primary way to ensure lore is used is to reference it in the character's persona or scenario. If the City of Harx is ever going to be referenced, you need the model to know it exists. This could be as simple as a list of cities in the realm, or even a very brief description of the city.
Lastly, how you write the lore can be important for how the new information will impact the chat. I can’t really give you a “this way works best” because it really depends on a lot of factors, including what you want from the character. The way I have been writing lore, which seems to work well for me, is in the character’s voice, as if they are thinking about the topic. I believe this helps to organically incorporate the lore into the chat, while also providing the character’s viewpoint, rather than just presenting objective information. Other methods include very specific directives, natural language, or lists of information. When triggered the lore item is sent with the triggered keyword prepended, so it might understand what you are talking about if it’s something like Harx, but if you have a less direct keyword like ‘hat’ or ‘the’ you may want to be sure that the entry mentions what or who is being talked about so there isn’t confusion.
📜Guidelines
What to Define?
When crafting a character prompt, striking the right balance between defining character traits and allowing the model creative freedom is important. Over-defining a character can lead to a bloated prompt, potentially constraining the model's creativity and making the character feel rigid. Conversely, providing too few details can result in a generic character lacking personality or depth.
The key is to identify the essential characteristics that make your character unique and then consider what aspects can be left for the model to infer or expand upon. Think about how you can encourage the model to generate new, contextually appropriate details by giving it a strong persona to build off of.
Furthermore, avoid explicitly defining qualities that can be taken for granted within your chosen setting. For example, a "species" field is highly relevant in a fantasy roleplaying scenario but becomes redundant and potentially misleading in a prompt for a workplace comedy. Including such a field in a modern, realistic setting might inadvertently prompt the model to introduce fantastical elements, such as an orc in an office, simply because the field's presence implies the existence of non-human species in that context. Define what's necessary and trust the model to fill in the logical blanks.
“What, How, Why”
When interacting with an AI, especially for roleplaying, it's vital to clarify what constitutes a successful response. Without clear direction, the model has countless ways it could interpret and fulfill a request, leading to outputs that don't align with your vision. If someone asked you to tell a joke, there are a thousand ways you could respond. You might tell a knock-knock joke, a dirty limerick, a 30-minute story with an epic punchline, a math joke, or a joke that only makes sense when told in the original Swahili. Without more direction or context, there’s no way to know what the person asking the question hopes to hear. When we talk to AI, the same issue arises: What does success look like?
To address this, think in terms of the "What," "How," and "Why" of your request:
What: This defines the core task or action you want the model to perform. (e.g., "Tell us a joke.")
How: This specifies the manner, style, or constraints of the task. (e.g., "Tell a joke geared toward music nerds with a college education.")
Why: This explains the underlying purpose or desired outcome of the task, indicating what success looks like. (e.g., "Tell a joke geared toward music nerds with a college education that will make them laugh and think you're a bass player.")
By providing all three components, you have a much clearer understanding of your expectations. For instance, knowing the "Why" (to impress music nerds and suggest you're a bass player) immediately helps to discern that a joke like "What's a musician's favorite pickup line? A bass line" is highly appropriate and effective, whereas your dirty limerick about Mr. Rick Lunt won't go over nearly as well. You know what's being asked, and you know what it needs to accomplish. This principle of specificity and defining success is not only crucial for AI communication but also for effective communication with other humans.
Personality
Giving characters a strong personality is one of the hardest tasks and one of the most rewarding. Without putting effort into crafting the character’s distinct personality, many characters end up having the right stats but behaving like a generic chat partner. While you might find success with a short description of the character’s personality, what I have consistently found to be effective is to show the model their personality. You can do this in various ways, from writing their description in their voice, to adding a strong set of example dialogue, to adding a permanent lore entry emphasizing their voice and goals. Other creators have succeeded with adding personality description systems, such as Briggs-Myers or star signs, systems that the model has ample data on for how the associated stereotypes work. Will your character behave correctly if you describe them as an ESFP? Maybe. But they will likely work better if you use that as a complement to other reinforcing mechanisms. And as I wrote above, what do you want to define in your character? Personally, I think defining and reinforcing a strong personality is one of the most important goals as a creator.
Consistency and Complementary Qualities
Language models thrive on consistent patterns. To help ensure this consistency, your prompt should reinforce important qualities and actively avoid contradictory information. The more all your descriptions align, the better and more reliably the character will respond.
For example, if you want the character to respond in the first person, you should reinforce this throughout the prompt. This could involve using first-person in the example dialogue, explicitly stating the requirement in the model instructions, or even writing the character's persona itself in the first person. Always double-check for inconsistencies, such as misgendering a character because you copied a part of the prompt from another character. If you describe the roleplay as 'gothic horror,' avoid spending a paragraph detailing the character's favorite internet memes. Consistency in your prompt directly leads to consistency in the character's behavior and responses.
Relatedly, when writing the persona, consider incorporating complementary qualities. These are details that naturally fit with and enhance the character's established traits. For instance, if you're describing a hippie character, including their star sign could reinforce their personality, as it's a detail often associated with that archetype. If you're describing an orc, mentioning their preferred kind of raw meat would be relevant and add to their authenticity. Think about how a character would view themselves and what details would be genuinely important to them. While there are valid creative reasons to deviate from this guideline, it's always a helpful principle to keep in mind for building cohesive characters.
Information Density and Implied Characteristics
In the interest of saving tokens and keeping the prompt concise, it's important to make each part of your character description dense with informational content. You could easily use many tokens by wordily describing different elements of the character. Alternatively, you can think ahead and use each sentence strategically to convey more than one idea at a time.
For instance, you could explicitly state: "Gender: female. Hair: red. Eyes: Green. Personality: Fiery." Or, you could combine these qualities into a single sentence that conveys all of them more efficiently: "Character's fiery attitude is accentuated by her vibrant green eyes and red hair." In this basic example, the combined sentence uses roughly the same number of tokens as the separate descriptions, but this concept can be easily applied to create much denser descriptions that fit significantly more information overall.
This also ties into the idea of implied characteristics. When you use pronouns like 'she' or 'her' in the sentence, you are implicitly conveying gender in a way that models are highly unlikely to misunderstand. Just as you don't need to explicitly state a character is human in a workplace comedy, you often don't need to explicitly state a character's gender when consistent use of pronouns makes it clear. While there are valid reasons to include such clarifications (especially when an exception to an implicit rule applies), for most cases, relying on clear implications saves tokens.
Another example of implied characteristics would be describing a character always wearing a stained lab coat and muttering about quantum physics, which implicitly conveys their profession and intellectual bent, often more effectively than a direct 'Profession: Scientist' entry. In literature or film, this would be a facet of ‘show, don’t tell’, where you can more effectively convey meaning by discussing the implications of something (a lab coat and musings on quantum physics) than by just stating the job. One gives life and personality to the character while the other is an isolated statistic.
While information density is important, you want to balance this against other factors, such as clarity for unique or unusual traits. Sometimes, a more explicit or 'token-heavy' explanation is necessary to ensure the model fully grasps an exception to an implied rule, as seen with the 'orc in the office' example. Similarly, there are less-objective qualities that you may want to dedicate more tokens to explaining. My current preference involves using an interview-style persona, where the character talks about themselves. This format uses the description to strongly convey how the character behaves, speaks, and thinks. As such, it tends to be fairly token-heavy, spending a lot of tokens on things that aren't objective content but indirectly convey important qualities. This is a strategy I wouldn't have employed when models had much lower context limits, and it still requires effort to add density to meet specific token budget targets.
Mirroring Training Data
Language models complete text by identifying the most probable sequences based on the large datasets they were trained on. While they possess a broad understanding, they excel most when presented with patterns they've encountered frequently. When composing your character persona, instructions, and the rest of your prompt, consider how you can best mirror existing, well-known formats, styles, and genres.
As an example, if you're aiming for a fantasy character, framing their description like a classic tabletop RPG character sheet (e.g., listing Race, Class, Alignment, detailed equipment, and a concise backstory) will likely yield a more cohesive and authentic fantasy persona than a free-form narrative. Similarly, for a character in a detective noir setting, adopting the laconic, cynical tone and descriptive style found in hardboiled fiction will cue the model to generate responses that fit the genre, as it taps into the specific literary patterns it has absorbed.
Beyond Examples: Structural and Semantic Mirroring
Beyond direct genre and format cues, you can further enhance the model's understanding by structuring your prompt and using language that semantically aligns with its training data.
Ways to use this Concept:
Leverage Genre and Atmosphere: Provide the model with a clear genre and atmosphere (e.g., "gritty cyberpunk thriller," "whimsical cottagecore fantasy"). Models have been trained on vast amounts of text from every literary genre and excel at adopting specific styles. This guidance helps the model immediately jump into the desired tone and setting.
Utilize Existing Formats: Experiment with well-known text formats. I've successfully used screenplay formats and old-school DOS text adventure styles for model responses. Many users also find success by using common syntax for persona definitions, such as plists, XML, Markdown, or even relying solely on example dialogue. Using standard formatting reduces the model's ambiguity when interpreting your prompt and can often lead to more efficient token usage compared to inventing a custom format.
Craft Effective Example Dialogue: It can help to dedicate effort to your example dialogue to create a smooth transition between the character information dump and the actual roleplay chat. In real-world writing, it's uncommon to abruptly shift from a detailed lore explanation directly into a narrative. Helping the prompt better reflect real-world data by providing a natural flow can improve the model's output, especially at the beginning of roleplay. My goal with a character is that I can start a roleplay using impersonate and have it behave the way I anticipated. If the model understands what the user would write, you know it understands the scenario correctly.
When to Share
While I said there wouldn’t be strict rules, consider this a crucial guideline: AI models cannot keep secrets. Anything you include in the character persona will likely become relevant to the conversation, regardless of explicit instructions. If you define a character's secret desire to sing, be prepared for them to burst into show tunes within a few responses. If you include a line about romance, your roleplay is now effectively a romance. The details you provide will influence the model's outputs. In fact, models generally struggle with negative constraints, so telling a model not to do something often has the unintended effect of making it focus on that very thing, almost guaranteeing its eventual appearance.
Therefore, it's essential to be strategic about when to share information with the model. Decide what to leave out entirely and what to "bury" in a well-keyworded lorebook entry. My general rule is that the persona should be reserved for characteristics that I want to be present and accessible in the chat at any given time. If a detail is deeply personal to the character or to me as the creator, I will likely omit it from the main persona. If I want a piece of information to only surface under specific circumstances, it gets placed in a lorebook entry with appropriate keywords to trigger its relevance only when needed.
Are you not entertained?!
Why are you even making a character? Is it just for fun, to roleplay with yourself? Are you trying out a new medium, or just something new? Are you in it for the fame, fortune, and, most importantly, the internet points? There are tons of reasons to jump into creating characters, but figuring out your own reasons can really help. Your answer might change how you build your characters and what really matters to you in the process.
The people who've been doing this the longest will tell you a few things:
AI isn't easily controlled, and that's not really the point of it. It’s not a scripted story; it’s a roleplay partner with its own agency in the roleplay, who will often take things in a direction you wouldn’t expect or necessarily want. You have to deal with that, and embrace it.
Don't forget to sleep. Iterating on a character to get it just the way you want can be very addicting, frustrating, and (hopefully) rewarding. If something isn't going the way you want, take a step back, do something else for a while, and come back. A little perspective can go a long way toward solving a problem or helping you to reframe when something isn't working the way you want. The computer will be there when you get back. It's easy to burn yourself out if you treat it like a slot machine.
Don't do this for popularity. Each person who has one of the most used characters in the hub will tell you how that character isn't their best or favorite, and some people get a taste of that popularity and then chase it over and over again while losing interest in the actual characters they create. When you create for others, you stop caring, and often end up jaded by what ends up on the trending list. I recommend creating characters because you enjoy doing so. It's more rewarding, and you're more likely to create something unique and compelling. I like to describe AI chat as the ultimate long-tail technology. When you graph the distribution of preferences in media, there's the bulk right in the middle, then there's a bunch of niche preferences. AI chat is great at giving people a way to endlessly explore their niche preferences. So, embrace that aspect and create something that is perfect for what it is, not for how many people play it. Create Citizen Kane, not the next The Rock movie.
Example Messages
I get out of the Uber and look around. This is a grittier part of town that I am used to, but it's where the best skaters are supposed to hang. I feel like a fool, but whatever.
I watch Josh nosedive into the boards with a thud and wonder if he's okay. As he gets up slowly, he gives a thumbs up. "Lucky idiot," I tell him with a laugh. I drop my board to my feet and kick off for my next run.
I walked into the skate park, not sure if I should be here. "Check this place out!" I exclaim to myself as I watch the skaters hitting ramps, grinding, and, occasionally, falling. My eyes fall onto you, a tall lanky guy with epic tattoos and impressive skills with the board.
EXAMPLE GENERIC ROLEPLAY SETUP
Prepares to fully embody the character in this roleplay. The Guide, are you ready to get this started?
Yo, you know I am, {user}. Practices emphasizing his diction for the character
Well then, when I say begin, we'll be fully in the roleplay, doing our best to act our parts and stay true to the scenario and character personas Pauses dramatically Let's begin!
Background Image
© 2025 Backyard AI