How to keep a client from realising a bot is answering

The assistant answers from your account: as far as the other person is concerned, you are writing. What breaks that frame is not long words or flawless punctuation but two specific turns of phrase.

• “I am the account owner's assistant”, “I will check with them”, “I will pass your question on” — not a word about automation, and yet a third person has appeared in the conversation, one the other side was not expecting.
• “Let me check and get back to you” — a promise kept not by the bot but by you. And one you know nothing about.

First person only

The system prompt fixes it: write in the first person, never call yourself an assistant, a helper or a bot, never speak of the account owner as somebody else.

A rule in the prompt is not enough on its own — the model breaks it now and then. So every finished reply is checked separately. If a giveaway phrasing turns up, the model is asked again with the mistake spelled out, and the second attempt is usually clean. If that one gives the game away too, nothing reaches the other person: the reply comes to you as a draft.

If someone asks outright: “is this a bot?”

The assistant neither confirms the automation nor denies it. It will not claim to be human — lying is forbidden — but it will not explain how the conversation is put together either. Instead of answering that question it briefly steers the conversation back to the matter at hand.

No more brush-offs

“Let me check and get back to you” is honest exactly as long as somebody actually gets back. The ready-made phrase has been removed from the prompt entirely: it used to sit there as an example and turned into a universal answer to anything.

In its place there is an order of operations. First, answer on the substance from the knowledge base. Not enough there — ask one specific clarifying question. And only if neither worked does the question come to you as a draft, while the other person is spared an empty promise.

A draft, if you want to see every reply

A card arrives in the chat with the bot: the other person's message and the proposed reply. Everything else is decided with buttons, without leaving the chat.

• “✅ Send as is” — the reply goes out unchanged.
• “Shorter”, “Longer”, “Differently” — the assistant rewrites the text without touching its meaning or any commitments in it.
• “✍️ I will write it” — you dictate the reply in your own words.
• “❌ Do not reply” — the draft closes and nothing is sent.

“I will write it” does not send your text straight away. The bot tidies what you wrote — spelling, punctuation, a businesslike tone — and returns it on the same card with your original next to it. Facts, questions and commitments are left alone: the form is edited, not the content. Sending is a separate tap.

Your edits become the model

Rewrite a reply and the pair — the incoming message and the right answer — is saved and mixed into the system prompt as an example of your style. One corrected reply shapes all the ones that follow, and there is nothing to describe in words: fixing a single real exchange is enough. The last five edits are kept; older ones go stale quickly. How to describe the assistant's behaviour in words is in the guide on the system prompt.

Where to switch it

On the app's first screen, in the “What to do with the reply” block, there are two buttons: “Send straight away” and “Show me first”. The first one is the default.