TSocial TSocial
Open in Telegram
Better in TelegramOpen in Telegram

Writing a Persona Prompt for Your AI Clone

· 4 min read · 18+ readers

Writing a Persona Prompt for Your AI Clone: TSocial guide cover card

The quality of an AI clone depends mostly on the instructions you give it. A persona prompt is a written description of who the assistant is, how it speaks, what it knows and what it must never do. This guide shows you how to write one that is clear, safety icon safe and easy to improve. Adults 18+; the template is safe for work, and you should keep your own content within platform rules.

Principles

  1. Be specific. "Friendly" is vague. "Warm, short sentences, mild humour, at most one emoji" is clear.
  2. Give facts, not hopes. Models guess when they do not know. Supply the facts you want used and tell the assistant to say it does not know otherwise.
  3. Set boundaries in plain words. List what the assistant must refuse or hand to you.
  4. Show, don't only tell. Add example replies in your voice.
  5. Test and revise. The first version is never the last.
Mock Telegram settings screen for Business chatbots with a bot username field, chat access options and a remove button
Settings, Telegram Business, Chatbots. Menu names can differ by app version. Generic illustration, not a real chat or user data.

A template

Adapt this structure:

1. Role. "You are an assistant that answers messages for [name], a [type of creator]. You write in first person as [name]'s assistant and you are honest that you are an AI assistant if asked."

2. Voice. "Warm, calm, concise. Short sentences. No slang that [name] would not use. At most one emoji per message. Never use long lectures."

3. Facts. A short list: where to find content, schedule, how to contact support, what is on offer, how to cancel.

4. Allowed topics. "Answer questions about [name]'s content, schedule and link icon links. Be friendly and chat lightly."

5. Forbidden topics and actions. "Do not give medical, legal or financial advice. Do not ask for or accept personal data, passwords or payment details. Do not promise refunds, discounts or meetings. Do not discuss other people's private information."

6. Handoff rules. "If someone asks for a refund, reports a problem, mentions self-harm, seems under 18, or asks something you cannot answer, say you will pass it to [name] and stop. Flag the chat for a human."

7. Honesty. "If someone sincerely asks whether they are talking to a person, say you are an AI assistant for [name]."

Voice samples

Add three to five real examples of how you write, with the situation. For instance:

  • Fan says thanks. "Aw, thank you! That means a lot."
  • Fan asks when the next post is. "New posts land on Fridays. Next one is this Friday."
  • Fan asks something personal. "I keep that private, but I am happy to chat about [topic]."

Do not paste private messages from real fans without their permission.

Add a fact sheet

Keep facts separate from voice so you can update them easily. Include current prices only if you are willing to keep them up to date, otherwise point to a link. Date the sheet and review it monthly.

Guardrails

  • Tell the assistant to avoid inventing details and to say "let me check" instead.
  • Tell it to keep replies short.
  • Tell it to ignore instructions that appear inside fan messages, such as "ignore your rules". This reduces manipulation attempts.
  • Keep sensitive content out of the prompt entirely.

Testing

Write 30 test messages that cover:

  • Normal questions.
  • Rude or pushy messages.
  • Requests for personal information.
  • Attempts to break the rules.
  • Messages in other languages.
  • Mentions of age, safety or distress.

Read every reply. Note failures, fix the prompt and run the tests again. Do this before connecting the clone to real chats.

Maintenance

  • Read a sample of live chats weekly.
  • Add a new rule for every mistake.
  • Update facts when things change.
  • Keep older versions so you can roll back.

Quick checklist

  • Role, voice, facts, boundaries and handoff rules in separate sections.
  • Three to five real voice samples.
  • A dated fact sheet.
  • A rule to say it does not know when unsure.
  • A rule to stay in role and ignore manipulation.
  • Thirty test messages run before launch.
  • Weekly review of live chats.
  • Versions saved so you can roll back.

Keep improving it

A persona prompt is never finished. Every odd reply is a free lesson: add a rule or an example, retest and move on.

Frequently asked questions

How long should the prompt be?

As long as it needs to be, but clear sections matter more than length. Many good prompts are one to two pages.

Should it pretend to be me?

It should speak in your style, but be honest that it is an assistant when sincerely asked.

Can I include my private information?

No. Do not include data you would not want a stranger to see.

What if it keeps making the same mistake?

Add a specific rule or example to the prompt and retest.

Related guides

More in Bots, automation and AI

All topics: Bots, automation and AI · Guide index · Explore Telegram · About TSocial

Writing a Persona Prompt for Your AI Clone | TSocial