I run a small wellness newsletter, weekly, about 3,000 readers, the kind of thing where people reply and genuinely expect an answer from a human. So when I wanted help drafting replies and brainstorming angles, I tried the obvious chatbots first. Every single one opened with "how can I help you today?" on a flat gray background, robot icon, personality of a help-desk intern. My readers would close the tab. A friend pointed me at x-rush's AI Chat and said "just build your own character." I'm not technical, I write for a living, and it took me one evening. Here's the exact path, every field included.
The shopping list
My brand is calm, a little mystical, and very specific about tone: no exclamation points, no "let's dive in." A generic chatbot fights all of that, and it isn't only the words. The room the conversation happens in matters to me. My newsletter's whole aesthetic is teal and moonlight, and a clinical gray chat window breaks the spell before the first reply lands.
So my shopping list had four lines: a face I choose, rules it follows, knowledge of my world, and a room that matches. In AI Chat those are the avatar upload, the behavior rules, the background material, and the chat appearance. Most tools let you nudge one of the four, maybe two. Here you set all of them for one character, in one form.
Where characters live
The path is /workbench/text/chat. Full IM-style surface: character list on the left, transcript on the right, message box at the bottom. There's no "new conversation" button and no sessions to manage; the character list is the entire entry point. You pick a character and continue its thread.
One thing confused me at first: my list looked different from my friend's. That's deliberate. Characters belong to the Project you currently have open. Every account starts with a protected Default Project that holds the built-in characters (a tarot reader, a trivia master, and a fitness coach were waiting in mine), and you can't delete that Project or rewrite what the built-ins know. When I switched to the Project I keep for the newsletter, the built-ins stepped aside and the list showed what that Project owned: nothing, at first. Then my own characters.
Built-ins aren't dead ends. Any of them can be copied as a custom character into a Project I choose, and the copy is fully editable with a blank history. If a built-in is close to what you want, start there. I started from zero.
Here's the constraint I braced for and didn't find: no cap. No counter ticking down, no create button going quiet at some limit. Make three specialists or thirty. The list is yours.
Building mine, field by field
Creating a character opens a structured form plus a small helper: an AI creation assistant whose only job is turning a rambling brief into editable fields. I typed one paragraph of "I want a slow-creativity mentor for newsletter writers who asks before she advises," ran the assistant, and it drafted the persona and a one-line intro for me to accept or edit. Each round with the assistant counts as one reply in a shared site-wide pool I'll get to below. Every field stays hand-editable either way.
Basics. Name and a role goal are the only requirements. The avatar is an upload: my moon-phase illustration came out at 1.8MB and the form bounced it right at the upload step, so I compressed it under the 1MB line and it stuck. A one-line intro and tags are optional; they make the list searchable later.
Role and rules. Persona first: who the character is, what it's good at, who it serves. Then the rules, where mine got specific: default language, tone, answer length, how far it's allowed to go on its own, standing bans (exclamation points, "let's dive in"), and one rule stolen from my editor: ask one question before giving advice. Items I left blank fall back to product defaults, so nobody forces you through a long form.
Background material. Paste text or upload a file; PDF and Office documents work, one file, 1MB again. The assistant reads it and distills an editable background description. Mine pulled out my topic areas, my no-go phrases, and the tone words I actually use. The whole character shares a 5,000-character budget across every field, and the background description can take up to 3,000 of it. Go over and it asks you to trim or summarize. Nothing gets silently cut.
Capabilities. Two switches, web search and attachment handling. My mentor runs with search off; her job is voice, not news. The tarot reader has it on, because a pull that knows what my week actually looked like beats one working from guesswork.
Look. The field I came for. Upload a chat background image, then set a theme color and a bubble style. I uploaded a starfield photo and set the theme color to my newsletter teal, and the transcript stopped looking like a support portal on the very next message. Files belong to the character here: avatar, background image, background file, and chat attachments together get a 10MB budget per character, separate from the account media library.
Opening. A welcome message and suggested questions for when a conversation starts. The assistant drafted three; I kept two and rewrote one.
There's no test chat inside the form, which annoyed me for about a minute. Save, and you land directly in a conversation with your character. That's the real test anyway.
The first reply, and the history that holds
I opened with a real problem: a reader asked whether morning pages still count as journaling if she types them. The reply came back in the voice I'd described, measured and curious, no exclamation points, and it asked me one question before advising, exactly per my rule. The starfield sat behind the whole exchange. Not borrowed. Mine.
Then the detail that kept me. I hopped over to the tarot reader for a minute, then switched back. My thread was exactly where I'd left it. Every character keeps one persistent history under the Project it belongs to; switching characters switches histories, and nothing mixes or clears. While we talked, the chat was also folding its own notes: after every ten successful replies it compresses the thread's context in the background, so a months-long conversation doesn't bloat. I never manage that. It stays caught up on its own.
Cost, because you'll wonder. Replies draw from one shared pool across every character and creation assistant on the account: ten successful replies cost 0.5 credits total, and a failed reply counts for nothing. My creation-assistant rounds and my mentor conversations dip into the same pool. Before this, I re-pasted my tone rules into every prompt and still got exclamation points, two minutes of surgery per reply. Now the rules live in the character and I just type. I wrote for several evenings before the first batch ticked over.
What it still gets wrong
The 1MB wall is real and applies to every upload. It bounced me twice, once on the avatar and once on a background PDF I'd exported with fat images. The rejection is clean and tells you why; compress or trim and retry.
The chat carries the distilled background description on each reply, not my full file. For voice, rules, and core facts that's plenty. If you bury one clause on page nine of a PDF, copy it into the rules yourself.
Config edits apply from the next reply onward, and old messages stay as they were. When I tightened my rules mid-thread, the conversation kept drifting for a few replies until I stopped fighting it and cleared the chat to restart with the new rules loaded. That worked. I wish I'd done it first.
And "be helpful" as a persona gives you the generic bot back. I know because I tried it. Naming the audience and banning the phrases is what made mine sound like mine.
The room connects to the schedule
One more hook, since my newsletter runs on routine. A character can be bound to one scheduled Automation, one character per job, and the binding is permanent. When a run fires, its output lands as a message in that character's history, and "continue conversation" in Automations drops me onto that exact message, in context, where follow-ups keep the same thread. The tarot reader runs on an 8am job. My mentor doesn't; I like the typing.
What's doing the talking
One paragraph on plumbing. I never pick a model anywhere in this flow. x-rush routes each message to a model that fits the task (a quick reply draft goes one way, a heavier restructuring question another), and my persona, rules, and background ride along on every route. The routing layer updates as the market moves. My character doesn't change when it does. I pick a character; the routing picks the model.
Where to go deeper
If the form landed for you, the docs go further than I have room for:
- AI characters & chat — the full surface: the character list, one persistent history per character, and every field in the form.
- Text module overview — how AI Chat, Automations, and HTML Studio chain into one loop.
- Automations: scheduled pushes — the scheduling system whose outputs land in your character's history.
I came in wanting a character that wouldn't embarrass me in front of my readers. I stayed because I got to design the face, the rules, the knowledge, and the room in one sitting, and the thread is still there when I come back. Your brand probably isn't teal and moonlight. But if you've ever winced at "how can I help you today," one evening and a compressed PNG are a cheap price to never hear it from your own character again.