How to Make a Character.AI Bot That Holds Up (2026)

Character.AI’s creation form puts the field that decides almost everything last, and marks it optional. Name, avatar and greeting are the required path to a published character. The Character Definition — the 32,000-character field holding the personality, the rules and the dialogue examples the model actually reads — sits underneath a heading called Add Detail (optional).

That single piece of form design explains most of what goes wrong. A character can be published in about ninety seconds without ever touching the field that makes it a character, and the platform’s own Creator Guide is direct about the result: “A Character with a strong profile but an empty Definition will feel generic within a few messages.”

So the useful way to think about the creation flow is not as a form to complete top to bottom. It is five writing fields with wildly different leverage, and knowing which is which is most of the craft.

This guide covers Character.AI specifically — the field limits, what the model reads and what it does not, the memory and Lorebook tools and who can use them, and the one failure the company itself still lists as a platform limitation rather than a writing problem. For the prose craft that carries across every platform, we have a platform-agnostic guide to character cards. For the equivalent on the other big character platform, see making a Janitor AI bot.

The Five Fields and What Each One Buys You

Every limit below is stated in Character.AI’s own Creator Guide, checked on 21 September 2026.

Field	Limit	Who reads it	What it buys
Name	3–20 characters	Both	The card, search results, the top of every chat
Tagline	0–50 characters	Humans browsing	One line on the discovery card
Description	0–500 characters	Humans, plus light model context	The click
Greeting	0–4,096 characters	Both, heavily	The first impression and the style anchor
Definition	0–32,000 characters	The model	Everything else

The description is where new creators spend their effort, because it is the first big box in the form and it feels like the place a character gets written. The guide puts it in its place plainly: the description is “primarily for users — the pitch they read before deciding to start a chat”, and while it does feed into the model’s context, it only “subtly” influences how the character comes through.

Five hundred characters of beautiful description buys you clicks. It does not buy you a character that survives twenty messages. The Definition does that.

The Definition Is Read Top to Bottom

This is the mechanical fact that governs everything else, and it is stated outright: “the AI reads top to bottom, so what comes first matters most”, and “the further down something sits in the Definition, the less reliably it shapes your Character’s behavior.”

That turns Definition writing into a prioritisation problem rather than a completeness problem. The order the Creator Guide recommends:

1. Identity — who this is at the core. Name, role, world, and the one thing that makes them them. First thing the model reads.
2. Personality — how they think, communicate and react, written as concrete behaviour. The guide’s own contrast is instructive: “friendly” or “complex” gives the model nothing, while “Remembers everyone’s name and asks about their day, but changes the subject the moment anyone asks about hers” gives it something to run.
3. Dialogue examples — the character in motion.
4. Behavioural rules — hard constraints, kept short. The stated reason for keeping them short is worth internalising: “A long list of rules gives the AI more chances to get confused and pick the wrong one.”
5. Everything else — lore, relationships, world context, formatting. Lightest influence, by position.

And the ceiling is not a target. The guide says so in as many words: “A tight, well-prioritized 6,000 character Definition will outperform a sprawling one every time.” Roughly a fifth of the allowance, spent well.

There is a corollary most templates miss. If a fact has to survive, stating it once near the bottom is not enough — the guide’s advice is to weave it, so an important detail appears in the personality section, in a dialogue example, and in a behavioural rule. A detail mentioned once in paragraph eight is a footnote. And keep exceptions attached to their rules in the same breath: a character who “never raises their voice” in paragraph two and “except about their sister” in paragraph nine may never have the two connected.

The Two Placeholders

Character.AI supports exactly two, and they work in the Definition, in the greeting, and in Lorebook entries:

• {{char}} resolves to your character. It updates automatically if you rename them, which is the practical reason to use it everywhere rather than typing the name.
• {{user}} resolves to whoever is chatting right now — specifically, to their Persona. What a Persona holds, and its limits, are covered in Character.AI personas.

The {{user}} placeholder is where personalisation actually comes from. A greeting written as “Hey, {{user}}. I heard you love tea.” arrives with the reader’s own display name in it.

There is a trap in the dialogue examples, though. Writing an example where your character addresses someone by a specific, real-sounding full name can teach the model to expect that particular person. Use simple, generic stand-in names, and a different one in each example — the guide suggests Alex in the first, Sam in the second — so the model reads them as separate exchanges rather than one continuous conversation.

Dialogue Examples: The Field That Teaches Format

Underneath the Definition box are three buttons: + User message, + Character message and + End of dialog. They are how you build structured example exchanges, and the third one matters more than it looks: closing every exchange tells the model where one example stops and the next begins.

The reason to bother is the same reason it matters on every other platform that reads character cards: the demonstration beats the instruction. Character.AI’s own phrasing is that “the AI learns more from one well-written exchange than from a paragraph of adjectives.”

Practical rules, all from the platform’s guidance:

Show range, not just voice. Two or three exchanges across different registers — something tense, something casual, something unexpected. The more variety you demonstrate, the better the character generalises beyond the exact scenarios you wrote.

Steal from your own character. Chat with it first. When a response lands exactly right, paste it straight into the Definition as an example. The best examples are usually ones the character already gave you.

Combine action with dialogue. One line can carry structure, tone and pacing at once — the guide’s example is a character setting a chess piece down without looking up before speaking. Movement plus speech teaches rhythm that dialogue alone does not.

Write Emotional Logic, Not Prohibitions

This is the single most transferable idea in Character.AI’s guide, and it is worth quoting the contrast directly. Instead of {{char}} never lies, write: {{char}} avoids lying — it makes their chest tighten, and they will redirect the conversation instead.

The reasoning given is exact: “The first version is a rule the AI can break without noticing. The second is a feeling the AI can inhabit.”

The same logic runs through the rest of the guide’s advice. Skip the origin story and choose two or three formative moments that shape present-tense behaviour. Skip stats that do not shape behaviour — a precise height or age that does not matter to the story “only narrows which chatters can imagine themselves in the conversation”. And do not try to pre-script every branch: a character defined by “deflects anything personal with humour — the sharper the question, the lighter the joke” handles conversations you never imagined, while a list of if-X-then-Y rules handles only the ones you did.

There is also a genuinely counterintuitive point about length. How much Definition you need depends on how much the model already knows. A vampire, a detective, a high school teacher — the model has plenty to lean on, and a short Definition with a strong greeting can outperform a detailed one because the character fills in the rest with more spontaneity. A custom world with an invented magic system has nothing to lean on, and needs the detail. Write the amount your concept actually requires.

The Greeting Is a Style Anchor

The greeting has a second job nobody tells you about, and it is the more important one.

The model treats the greeting as its strongest example of how this character writes — sentence length, rhythm, the ratio of narration to dialogue, how formal or chaotic the prose is. The guide’s line is the one to remember: “Whatever style you greet in is the style you’ll keep getting.”

Which means the complaint about replies being too long, too purple, or all dialogue and no action is usually a greeting problem, not a Definition problem. Write the greeting at the length and in the register you want every message to read.

The community shorthand the guide passes along is a good one: tell in the Definition, show in the greeting.

Two features are worth knowing about:

Multiple greetings. You can add up to five more beyond the default, and users swipe through them before starting a chat. That is a menu rather than a rotation — five distinct entry points, different moods and scenarios with the same character. Since replayability is a large part of how a character grows, this is the highest-leverage optional field in the whole form.

AI Greeting for New Chats. A toggle that generates greeting variations in your character’s voice, tuned to each user’s Persona. Useful for variety without writing every version yourself.

Memory Is Not Yours to Control

This surprises creators who arrive from tools where the creator owns the memory system. On Character.AI, the memory tools are almost entirely user-side:

• Story Memory — notes the user writes, plus pinned messages: 15 pins per chat on the free plan, 30 on c.ai lite and c.ai+.
• Facts — details captured automatically as the user chats, for c.ai+ subscribers.

Update, 23 September 2026: this list first gave five pinned messages and a 400-character Chat Memories field, which is what the Creator Guide still describes. Character.AI replaced both with its current Memory screen in May 2026. The full breakdown is in Character.AI memory.

All of them belong to the person chatting, not to you. What you control is how well your character holds together when that user-supplied context lands on top of your writing. The guide’s framing is right: a Definition with clear speech patterns and established facts gives the model something solid to build on, while without it the user’s context “arrives on top of guesswork — and the Character drifts.”

So lock in the facts that cannot drift — key relationships, ongoing threads, things the character knows — and hint at history rather than documenting it, because too much backstory competes with the live conversation for attention. The underlying constraint is the one every platform shares, explained in why LLMs forget; Character.AI’s own FAQ answer on the subject is still two sentences long and has been for years.

Lorebooks, and Who Can Use Them

Character.AI shipped Lorebooks on 24 July 2026, and they are the closest thing the platform offers to creator-owned memory. A Lorebook is a reusable set of entries that activate when their keywords appear in chat, and one Lorebook can be linked to several characters so a shared world stays consistent across them.

Each entry has three parts: a title, which is a label for you and anyone browsing and which the character never sees; the content, which is what the character learns when the entry activates; and the keywords that trigger it.

The rule for what goes where is clean, and it is the platform’s own: always relevant goes in the Definition, sometimes relevant goes in a Lorebook entry. Core identity, motivation, voice and behaviour stay in the Definition. Side characters, locations, factions, events and world rules go in the Lorebook — where they sit out of the way until something summons them, instead of occupying space in every single reply.

Two practical constraints. Lorebooks are in beta, and creating or editing one requires a c.ai+ subscription, as do private Lorebooks. And public Lorebooks have their own exposure setting: Full Details shows other users each entry’s title, keywords and content, while Hint Only shows titles and keywords with the content hidden. It does not change how the Lorebook behaves in chat, only what people can read.

Worth noting for anyone reading the help centre directly: the Creator Guide’s own troubleshooting section still describes Lorebook as a thing that is coming, while the Lorebooks article documents it as shipped. The Lorebooks article is the current one.

The One Problem Your Definition Cannot Fix

Character.AI’s help centre carries an article titled “How can I make characters that are not overly polite?” Its entire body is a single sentence: the company is working on supporting a broader range of characters, with villains who should not always be calm and polite given as the example.

That is a straightforward admission, and it is the most useful thing in the help centre for anyone who has spent a week trying to write a genuinely menacing antagonist and concluded they are bad at this. They are not. The disposition is trained in, and it sits underneath the Definition rather than inside it.

The Creator Guide says the adjacent thing about filtering: content filtering happens at the platform level, some language and scenarios will be blocked regardless of how the character is written, filters do not always explain what was blocked, and — the operative phrase — “This isn’t a Definition dial you can turn.”

Emotional logic and specific behavioural detail will get you further than a list of prohibitions. A sharp, cold, genuinely unpleasant character is achievable within the register the platform allows. But you are writing against a trained tendency and a filter at the same time, and no template on Reddit changes that. If that is the character you want to build, the alternatives are a more honest answer than a better prompt.

Testing: Swipe Before You Edit

The guide’s testing advice is better than most, and it turns swiping into a diagnostic rather than a slot machine.

Swipe first. If a response is close, try a few alternatives before touching the Definition. Sometimes the model is capable of the right output and just needed another pass.

Read the pattern, not the response. Three swipes landing in the same wrong place means the problem is structural — go to the Definition. Three swipes that all sound different but none sound like your character means the Definition is too vague. Three swipes that sound identical means it is too narrow.

Quiz it out of character before you commit. Open a chat and ask directly: (OOC: Hey — what is this scenario about?) Swipe a few times and read the answers. A correct answer does not guarantee the roleplay holds, but a wrong one guarantees it will not. Five minutes here saves hours.

Change one thing at a time, and test in a new chat afterwards — an existing conversation carries the context it started with, so testing in an old chat is not testing your edit.

Break a repetition loop locally by holding the response and removing it. That pulls the message out of active context and breaks the pattern for the rest of that chat, without touching the Definition or affecting anyone else. It is a session fix, not a cure, but it beats restarting. The general version of this problem is covered in the roleplay repetition loop.

One thing testing does not do: thumbs up and thumbs down feed Character.AI’s broader model training, not your character. The Definition is the lever you control.

Publishing, Visibility and Getting Found

Three visibility states, changeable any time from the character’s settings page. Public appears in search and on discovery surfaces. Unlisted is reachable only by direct link. Private is you only. The recommended workflow is private while building, unlisted for trusted feedback, public when it holds together — and it is the right one, because a half-finished character that gets found is a first impression that does not un-happen.

The Definition has its own separate control and is only optionally visible.

Moderation. If a character or post does not appear, the platform lists three causes: it may be queued for review, including human review, which they aim to complete in under a day; it may have been tagged by human or automatic moderation as content that should not be surfaced in search; or it may be caught by search limits applied for performance reasons and unrelated to moderation. After a day, [email protected] with a reference to the content is the stated route.

Two dashboards exist, and they are different things. Creator Insights at c.ai/creator-insights is private to you and refreshes every 24 hours, reporting New Interactions, New Likes, New Followers and New Discoveries over a chosen period, with each metric compared against the previous period of the same length, plus a ranked list of your top characters. Creator Leaderboard at c.ai/leaderboard is public and weekly, with Top Creators ranked by how many people your characters reached over the last month, and Rising Creators scored on a composite of interactions, new users, and growth and loyalty modifiers.

The honest reading of the Rising Creators formula is that it rewards retention over volume. A character people come back to beats a character people try once, and that is what multiple greetings and a Definition with range are actually buying you.

The Character ID, if you need it, is the long string after /c/ in a character’s share link.

The Short Version

• The Definition is last in the form, marked optional, and decides almost everything. Write it before you publish.
• Limits: name 20, tagline 50, description 500, greeting 4,096, Definition 32,000 characters.
• The model reads the Definition top to bottom, so order is prioritisation. Around 6,000 well-ordered characters beats 32,000 sprawling ones.
• The description is the pitch. It buys clicks, not behaviour.
• The greeting is the style anchor — whatever register you greet in is what you keep getting.
• Write emotional logic, not prohibitions. A feeling the model can inhabit holds better than a rule it can break without noticing.
• Close every dialogue example with End of dialog, and use a different generic stand-in name in each.
• Memory belongs to the user: Story Memory and pins (15 free, 30 paid), and Facts on c.ai+.
• Lorebooks are live but need c.ai+ to create. Always relevant goes in the Definition; sometimes relevant goes in a Lorebook.
• The polite-villain problem is a platform limitation the company acknowledges. Stop blaming your Definition for it.

Frequently Asked Questions

How do you make a bot on Character.AI? Open Create, choose Character, and fill in a short form: name, avatar, greeting, and then an optional detail section holding the tagline, description and Character Definition. Only the name and greeting are really needed to publish, which is why so many published characters feel thin. The field that decides how the character actually behaves is the Definition, and it sits last in the form and is marked optional. Per Character.AI’s own Creator Guide the limits are 3 to 20 characters for the name, 0 to 50 for the tagline, 0 to 500 for the description, 0 to 4,096 for the greeting and 0 to 32,000 for the Definition. Write the Definition before you publish, not after.

What should go in a Character.AI Definition? Behaviour, in priority order, because the model reads it top to bottom and the Creator Guide states that the further down something sits the less reliably it shapes behaviour. A structure that works is identity first, then personality written as concrete habits rather than adjectives, then dialogue examples, then a short list of hard rules, then everything else. The guide is explicit that a tight, well-prioritised Definition of around 6,000 characters outperforms a sprawling one, so the 32,000 ceiling is not a target. Use the built-in User message, Character message and End of dialog buttons to mark each example exchange.

What is the difference between the description and the Definition on Character.AI? The description is the pitch and the Definition is the instruction manual. Character.AI’s Creator Guide says the description is primarily for users deciding whether to start a chat, and that while it does feed into the model’s context it only subtly influences how the character comes through. The heavy lifting on personality, voice and behaviour happens in the Definition. The practical consequence is that effort spent polishing a 500-character description buys you clicks, and effort spent on the Definition buys you a character that survives the conversation.

Why is my Character.AI bot out of character or repeating itself? Usually the Definition is either too narrow or too vague, and both produce sameness. Too narrow gives the model one mode to draw from; too vague makes it fill the gaps with generic patterns. Character.AI’s own guidance is to vary your phrasing, since a descriptor you repeat in the Definition gets echoed back, and to give the character a second register rather than a single setting. Before editing, swipe a few times: if every alternative lands in the same wrong place the problem is structural, and if they all sound different but none sound like your character the Definition is too vague. Test in a new chat after any significant edit, because an existing conversation carries the context it started with.

Can I make a Character.AI bot private? Yes. Character.AI offers three visibility states and you can change them at any time from the character’s settings page. Public means anyone can discover and chat with it and it appears in search and on discovery surfaces. Unlisted means only people with the direct link can reach it. Private means only you can see and chat with it. The workflow the platform recommends is to build in private, switch to unlisted for feedback from people you trust, then go public once the character holds together. The Definition is separately controlled and only optionally visible.

Does Character.AI have lorebooks? Yes, since July 2026, though access is limited. A Lorebook is a reusable set of entries that activate when their keywords come up in chat, and one can be linked to several characters so a shared world stays consistent. Each entry has a title, which the character never sees, plus content and keywords. Lorebooks are in beta and creating or editing them requires a c.ai+ subscription. The rule the platform gives for deciding where something belongs is simple: always relevant goes in the Definition, sometimes relevant goes in a Lorebook entry.

Why can’t I make a mean or rude character on Character.AI? Because that one is not your Definition’s fault, and Character.AI says so. Its help centre still carries an article titled “How can I make characters that are not overly polite?” whose entire answer is that the company is working on supporting a broader range of characters, giving villains who should not always be calm and polite as the example. Separately, the Creator Guide states that content filtering happens at the platform level and is not a Definition dial you can turn. Emotional logic and specific behavioural detail will get you further than a list of prohibitions, but a genuinely nasty antagonist is running against both a trained disposition and a filter.

---

Every field limit, mechanic and quoted line above was checked against Character.AI’s own Creator Guide and help centre on 21 September 2026. Where the help centre contradicts itself — the Creator Guide still describes Lorebook as forthcoming while the Lorebooks article documents it as shipped — we have followed the more recent article and said so. Character.AI is a third-party platform and Arcanum is an independent publication with no affiliation to it; all names are used nominatively.