Give Every Telegram Group Its Own AI Personality and Skill Set

Most AI assistants come with one voice, one set of tools, one way of behaving. That works for a quick question. It falls apart the moment you try to use the same assistant across different contexts — a family group chat, a work team, a book club, a side project.

Agent One now solves this. One assistant, a completely different personality and skill set in every group you put it in. Brand voice in your company's Threads launch group. Strict study buddy in the university thread. Project manager in the team chat. Witty friend in the family group. Same bot, same login, same memory per group — but a different mode of operation the moment it sees which chat a message came from.

This post walks through what changed, why it matters, and how to set it up.

What's New: Personality + Skills, Per Group

Two things used to be global: the assistant's personality, and what it could do. If you set your Agent One to be a "friendly study buddy," that's what it was everywhere. If you wanted it to build websites on demand, that capability was on in every chat, including the ones where you didn't want it.

Now both live at the group level:

• Per-group personality. Each group gets its own name, emoji, and full personality definition. Pick from built-in templates (Study Buddy, Writing Coach, Brand Voice, Project Manager, etc.) or paste in your own — up to 10,000 characters of role, tone, and directive.
• Per-group skills. Skills are custom capability files you create yourself in Settings → Skills — a code-review checklist, a brand-style writer, a recipe formatter, an SEO auditor. Enable only the ones a given group needs. Built-in features (memory, web search, reminders, website builder) are always available — skills are on top of that.

The assistant recognizes which group a message came from and loads the matching persona and skill kit automatically. No manual switching, no prompts to tell it what mode to be in.

Why This Matters

Context collapse is the quiet problem with AI assistants. A single persona that tries to be everything ends up being mediocre at everything. Ask a generalist chatbot to be a concise senior engineer in one thread and a warm therapist in another and it averages the two out — never fully committing to either.

Per-group personalities solve this by giving the model a clean slate for each context. In the brand account group, it commits to the brand voice. In the therapy support group, it commits to warmth and patience. The model doesn't have to decide what to be — it just is, based on where the message landed.

Skills work the same way. Create a custom skill — a code-review checklist, a brand-style writer, an SEO auditor — and enable it only in the groups that need it. The assistant picks up the relevant skill files when a message lands in that group, and ignores the rest. Core features (memory, web search, reminders, website builder) remain available everywhere.

Real Setups We've Seen

A few of the more interesting configurations users have built recently:

• Brand voice for a social media team. A company's Threads account uses the assistant as the brand voice in a team group. Personality: short sentences, strong opinions, no AI disclaimers. Custom skill enabled: social-post-adapter — a skill the user wrote that reformats any message into a Threads-ready post.
• Study group tutor. A university exam prep group with the assistant as a patient, rigorous tutor. Personality template: Study Buddy with a custom addendum about the specific course material. No custom skills needed — built-in web search handles the research.
• Family coordinator. A family chat uses the assistant as a friendly household manager. Personality: warm, casual, emoji-light. No custom skills — reminders and web search are always on.
• Engineering team assistant. A small dev team configures the assistant as a concise engineering peer. Personality: direct, technical, no fluff. Custom skills enabled: code-review and adr-writer — two skills the team wrote in Settings to enforce their review standards and write architecture decision records.

How to Set It Up

The whole flow lives in a Telegram mini app. You don't have to write a config file or visit a website.

What You Get for Writing a Good Custom Personality

Template personalities are fast and good enough for most groups. But the custom field is where this system really opens up. A well-written custom personality turns a generic assistant into something that feels purpose-built.

A few things that tend to work well in custom personalities:

• Directives, not descriptions. "You write in short sentences. You never use corporate jargon." beats "You are someone who writes clearly."
• Explicit boundaries. Tell the assistant what not to do. "Never introduce yourself as 'an AI assistant'." "Never use the words 'seamless' or 'game-changer'."
• Examples of voice. A few sample responses or post drafts train the model's style faster than adjectives.
• A clear role. "You are the brand voice on Threads." is enough context for everything else to fall into place.

Privacy and Isolation

Each group has its own memory, its own personality file, and its own skill configuration. The assistant can't leak context between groups — what's said in one stays in that one. When you change a personality, the assistant's conversation history for that group is cleared so old responses don't contaminate the new persona.

And because the data layer is per-user, your groups aren't pooled with anyone else's. Your configurations, your conversations, your skills.

Try It

If you already have an Agent One assistant, open any group it's in and type /settings. The mini app will show you the per-group options.

If you don't have one yet, the free plan includes one group with a custom personality — enough to see how this works with your own use case. Sign up, add the assistant to a group, configure it, and see what it feels like when an AI actually fits the context you put it in.