Build a Custom AI Agent: Avatar, Personality, Chat Background, and Backstory (2026)

The problem, stated plainly

My brand is calm, a little mystical, and very specific about tone - I don't do exclamation points, I don't do "let's dive in." A generic chatbot fights all of that. And it's not just the words. The space the conversation happens in matters to me. My newsletter's whole aesthetic is teal and moonlight. A clinical gray chat window breaks the spell before the first reply lands.

What I wanted was an agent I could tune on four things at once: what it looks like, how it sounds, what room it lives in, and what it knows about its job. That's avatar, personality, chat background, and backstory. x-rush's AI Chat hands you all four in one form. Most tools give you one - maybe two - and call it customization. This one doesn't.

Step 1: Open AI Chat, where custom agents live

The path is /workbench/text/chat. This is the deep-conversation room, not the little floating bubble you see on other pages. Full IM interface: a 240px agent list on the left, a scrollable transcript on the right, a message box at the bottom. The whole surface is keyed to one color - #00D4AA, a teal that marks every active tab and selected agent. Sounds cosmetic, but it's how I confirm I'm in the right place.

The sidebar shows the built-in seven first - Tarot Reader 🔮, Trivia Master 💡, Horoscope ♐, Fitness Coach 💪, Study Assistant 📚, MBTI Analyst 📊, Nutritionist 🍎. Below them, a header reads Custom Agents (0/10). That counter is doing real work: it tells you how much room is left, and when you hit ten the "Create Custom Agent" entry goes quiet until you delete one. Ten is enough for a recurring cast; it stops the sidebar from becoming a 40-item dump.

I clicked Create Custom Agent. The form modal opened with five fields. Four of them are the dials I came for.

Step 2: The four dials

Here's where most tutorials wave their hands and say "set your preferences." I won't. Here's exactly what each field does, because I learned the hard way that they're not interchangeable.

Avatar. This is one input that takes either an emoji or an image URL. I typed 🌙 - done. You could also paste a hosted logo URL if you want your brand mark instead. Here's the part I didn't expect: there's a priority chain in the code. The helper getAgentEmoji() checks your custom avatar first, then falls back to a built-in emoji table, then to the first letter of the agent's name. Translation: your override always wins, but built-in agents still get a sensible icon for free. I didn't have to touch the seven defaults - they already had their emojis.

Personality (Description). A one-liner that sits under the name in the sidebar and shapes the vibe when you're picking. Mine: "calm, encouraging, asks a question before giving advice." This is the flavor field. It's short on purpose - the heavy lifting lives next door.

Chat Background. This is the dial I came for and the one almost no other tool offers. One input, takes either a hex color or an image URL, and it applies to the actual chat area behind the messages. I pasted #00D4AA - my brand teal - and the transcript's backdrop flipped to match my newsletter instantly. You could drop in a soft gradient image URL instead. The catch, and yes there's a catch: it has to be a valid hex or a real image URL. A typo buys you nothing - the field just ignores it and you're back to default. I tested a bad hex to be sure. Nothing broke, but nothing changed either.

Backstory (System Prompt). A textarea, and this is where the agent actually becomes yours. The description is the flavor; the system prompt is the rulebook. Mine runs about six lines: who the agent is (a slow-creativity mentor for newsletter writers), who the audience is, the tone rules (no exclamation points, no "let's dive in," ask before advising), and a standing instruction to ground suggestions in the reader's actual previous topics. Here's the honest part: "be helpful" in this field produces a generic bot. I know, I tried it first. The specificity is the whole point. Spend the ten minutes.

Name was the fifth field - just identity, not a dial. Save, and the new agent is selected automatically. You're dropped straight into a fresh thread with it. No extra click, no hunting in the sidebar.

Step 3: Talk to it and watch the room hold

I typed my first message. The reply came back in the voice I'd described - measured, curious, no exclamation points. The teal background I'd set sat behind the whole exchange. The moon emoji headed the thread. For the first time the agent didn't feel borrowed.

Then the detail that sold me on staying. I switched to the Tarot Reader to grab a daily card, then switched back to my custom agent. My thread was exactly where I left it - nothing cleared. Each agent keeps its own history, stored in a map keyed by agent id. Switching is free. Most chat tools reset the conversation the moment you change contacts; this one treats every agent like a separate contact with a separate thread, the way your phone does. That's the difference between "I'm chatting with a bot" and "I have a cast of characters who each remember me."

The four dimensions, and why they matter together

Here's the sell, stated as a tradeoff. Plenty of platforms let you tweak a system prompt. Almost none let you also set the room the conversation happens in. The four dials map to four layers of "feels like mine": the avatar is who's there, the personality is how they sound at a glance, the chat background is the room they're standing in, and the backstory is what they actually know and do. Skip the background and you've got a well-written bot in a fluorescent office. Skip the backstory and you've got a pretty room with nobody home. x-rush is the first place I've used where I didn't have to skip any of them - and where tuning all four took one form, one evening, no code.

There's a nice overflow valve too. When I'm reading a daily push from Auto Content and want to go deeper, the "continue chatting" link drops me here with the agent pre-selected and a ContextCard quoting what I came from. The custom agent inherits that context. It's not a separate world - it's the room my whole text workflow empties into.

Pro tips I picked up the hard way

• Spend your time on the system prompt, not the avatar. The emoji takes five seconds. The backstory is the difference between a bot and a character. Write it like a brief to a freelancer: audience, tone rules, standing instructions.
• "Be helpful" is a trap. I tried it. You get the generic bot back. Specificity is the entire mechanism - name the audience, ban the phrases you hate, tell it to ask first.
• Test your chat background with intent. A valid hex like #00D4AA works; a malformed one silently does nothing. If the room didn't change, you typoed - there's no error message, just no change.
• Make specialists, not one generalist. Because each agent keeps its own history, you lose nothing by having a "reply drafter" and a separate "topic brainstormer" instead of one agent doing both. Threads stay clean.
• Mind the 10-agent cap. The counter is real - hit ten and Create goes quiet. Delete to add. Better to retire a stale agent than to merge two roles into a muddled one.
• The global bubble is hidden here on purpose. Inside /workbench/* the floating chat bubble doesn't show, so you're never fighting two chat surfaces at once. This page is the chat.

What's actually doing the talking

One paragraph, because it matters. x-rush is wired into the industry's top text models - Claude Fable 5, GPT-5.6, Gemini 3.5, Qwen3.7 Max - and smart-routes each message to whichever fits the task. A quick "draft a reply" hits a fast model; a longer "redesign my onboarding angle" gets the heavier one. My custom agent's personality stays mine regardless of which model answers - the system prompt rides along on every route. The pool refreshes as the world moves; whatever ships next, the router uses it where it wins. I don't pick a model. I pick a character.

Where to go deeper

If the four dials landed, the docs go further than I have room for:

• AI characters & chat - the full T-3 surface: the agent sidebar, per-agent history, session reuse, and every field in the form.
• Text module overview - how T-1 (HTML Pages), T-2 (Auto Content), and T-3 (AI Chat) chain together into one loop.
• Auto content generator - the daily-push system that feeds context into your custom agents.

I came in wanting an agent that didn't embarrass me in front of my readers. I stayed because I got to design the room, the voice, and the rules in one sitting - and the thread's 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 is a cheap way to never hear it from your own agent again.