Open ten AI sales assistants and nine personas begin the same way: professional, friendly, knowledgeable, always helpful. Those words do almost nothing. Every model already believes it is professional, friendly and helpful, so the instruction adds no information and the replies come out as generic as the brief.
A persona earns its place when it makes the assistant answer differently from the default. That means constraints, examples and refusals, not adjectives.
Start with length and shape
The single change with the largest effect on how human a bot reads is message length. Left alone, models write in paragraphs with headings. People on Telegram write two lines, wait, then write two more. Say so explicitly: short messages, one idea each, no bullet lists in a DM, no headings, and a question at the end when the conversation needs to move.
Say what to do with the greeting too. A returning contact should not be greeted like a stranger, and a lead who has already given their name should never be asked for it again. The CRM record knows this contact has been seen before, and the persona should tell the assistant to write accordingly.
Name the things it must never do
Refusals are the part of a persona most people skip. Write them anyway, and write them as concrete situations: do not quote a price that is not in the fact store, do not agree to a call time without checking, do not answer questions about someone else's account, do not discuss competitors by name.
Some rules are not yours to write. Flirting, romantic framing and promises the business cannot keep are blocked by the engine on every reply regardless of persona, which is covered in the hard rules. Your refusals sit on top of that floor and handle the things specific to your business.
A persona is not a knowledge base. Do not paste your price list into it. Prices, policies and product detail belong in the fact store, where they can be updated once and used by every bot without rewriting anyone's voice.
Voice comes from examples, not descriptions
Instead of writing "our tone is direct and warm", paste three real replies you have sent and mark them as the target. Models copy structure from examples far more reliably than they follow abstract tone instructions. Pick examples that show the range: one that answers a simple question, one that handles a price objection, one that says no.
Two more details worth pinning. Decide whether the assistant uses the business name in the first person, and decide how it refers to itself if asked directly. Both come up in real conversations within the first week, and an assistant that improvises the answer will improvise a different one each time.
Language is not part of the voice
Scalegram answers in the language the incoming message arrived in. That happens outside the persona, so do not write "always reply in English" unless you genuinely mean it, and be aware that a customer writing in their own language expects an answer in it. Voice and language are separate settings for a reason, and the mechanics of that are covered in voice and language.
What the persona should carry is register: whether the business is formal or plain-spoken. That survives translation. Slang usually does not, and an assistant reaching for idiom in a language you do not read is a risk with no upside.
The handoff sentence
Every persona needs one sentence about giving up. Something like: when the answer is not in the facts, when the customer is upset, or when the request involves money or access, stop and hand over rather than guess.
That sentence does more for perceived quality than any amount of tone work. Customers forgive a bot that says "let me get a person on this" almost instantly. They do not forgive a confident wrong answer. The routing side of it, who receives the conversation and what they see, is worth setting up properly and is covered in handing a conversation to a human.
Test it the boring way
Write the persona, then send the bot twenty messages you have actually received. Not invented test cases, real ones from your Telegram. Most people find three problems in the first ten: the assistant is too long, it is too eager, and it answers something it should have escalated.
Fix by subtraction. Cut the persona rather than adding to it. Every extra paragraph competes with every other paragraph for the model's attention, and a persona that runs past a page usually performs worse than the tighter version of itself. The teams that get the best results run something closer to fifteen lines than fifty.
One persona per audience, not per whim
If you sell two products to two different audiences, run two bots with two personas rather than one persona full of conditionals. Conditionals are where models get confused, and the failure is silent: the assistant quietly applies the wrong branch and nobody notices until a customer quotes it back.
Keeping them separate also makes measurement possible. Two bots produce two sets of numbers, and when one clearly outperforms the other you can read the personas side by side and see which decision caused it. That is a slower way to improve a bot than rewriting it on instinct, and it is the only one that compounds. More on the product itself is at scalegram.io.
"Write the persona the way you would brief a new hire on their first morning. Not adjectives. What to say, what to skip, and when to fetch someone else."
— Alex Onta, Executive Director, SINGUARD
Key Takeaways
- Adjectives like professional and friendly change nothing; constraints, examples and refusals do.
- Message length and shape is the highest-impact setting: short Telegram-style messages, one idea each, no headings.
- Keep prices and product detail in the fact store, not in the persona, so one update reaches every bot.
- Give the persona one explicit handoff sentence, then test it against twenty real messages and cut rather than add.
Frequently Asked Questions
How long should a bot persona be?
Shorter than most people expect. Around fifteen lines of concrete instruction usually outperforms a page of description, because every extra paragraph competes for the model's attention.
Should the persona say which language to reply in?
Normally no. Scalegram answers in the language the message arrived in as a separate setting, so the persona should carry register and refusals instead.
Can one persona cover two different products?
It can, but conditionals are where assistants quietly pick the wrong branch. Two bots with two personas is easier to control and lets you compare the numbers side by side.
About the Author
Alex Onta is an Executive Director at SINGUARD. He built eTrader, the terminal, the mobile apps, eTrader Broker, Copytrading, Business and Community, along with the worldwide clustered-server infrastructure it all runs on, with his brother Roman Onta helping on the design, and he leads that division today. Together with Roman he builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals, and the two of them carry worldwide compliance, payment processing and international business structuring side by side. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.