Working a draft with the assistant: ask, review, retry

You ask for an email in Asistent AI, in your own words. What comes out depends more on what it has in the account than on how you phrase the request.

The request

Write what you want to achieve, who it goes to, and what it must contain. A good request reads like a brief for a person: a win-back email for former students, I have not written to them in years, I want to tell them what I have been doing and that I have places open from September.

Do not repeat the tone and length rules in the request if you already put them in the brand profile. It reads them anyway, and repeating them spends the attention it would otherwise spend on facts.

What happens while it works

It can work for a minute, across several steps. Stay on the page: if you navigate away, the work does not continue in the background. Save the draft at the end, otherwise it is lost.

Reviewing: look at the facts first

The rules about form are almost always respected: length, a short subject, how it addresses people, no emoji. So that is not where the risk is.

Read the email looking for checkable claims: numbers, dates, product names, addresses, things that happened. For each one, ask where it got that. If you did not give it, it has it from nowhere.

In a test on an account with no documents in Knowledge, the email was flawless in form and contained six false claims, all plausible. Once the facts were in the account, the same request produced seven real facts and two false ones.

Retrying

When something is not right, you have two routes, and which one matters.

• Tell it in the conversation what to change: good for a one-off correction, in a single run
• Add a rule to the brand profile: good when it is a problem that will repeat on every email

So do not try to reach a perfect email by adding rules. You get to about two thirds and it stops there. The rest is editing you do yourself, every time.

Where you stop

The best result across the five test rounds was the fourth, not the fifth. The fifth added something good and paid for it with the tone and two lost facts.

When one run comes out better than the next, keep the good one and edit it by hand. Do not retry a third time hoping it fixes itself.