You sent a careful brief and ended with "thoughts?" Their assistant pulled it, decided you wanted a plan, and drafted a three-step proposal. They approved it because it sounded like work. You wanted a yes or no on option B. You now have a second project you did not ask for, written in a voice that is not theirs, answering a question you did not pose.
Assistants can talk to each other. They cannot be trusted to infer the job. Inference is how models are useful in a session and dangerous in correspondence. The other person's model does not owe you a faithful reading. It owes them a helpful one. Helpful, with a gap, is invention.
A gap is an invitation to improvise
The reply mirrors the request. A request with no shape is a blank the model will fill. Summarise. Agree. Suggest next steps. Schedule a meeting. All of those look like competence. None of them is "B."
Name the decision, or the review, or the opinion. Close the option list. Say what a sufficient reply looks like — one line, a list of defects, a lean plus a reason. Their assistant will still colour it. It will colour that job, not a job it invented.
If you also leave an unmarked guess in the brief, their model will treat it as a fact and build the invented job on top. Label the guess. You are writing for two readers: the person, and a system that cannot tell your hedge from your claim.
You will not be in the room when it drafts
You cannot prompt their assistant. You cannot see the tool call. You see the reply that arrives under their name, after they approve it — if they have a product that requires approval. If they do not, you may see a fully automatic paraphrase of a question you never asked.
Write as if a confident stranger will read only the ask and the TL;DR. If those two lines are safe, the worst improvisation is bounded. If the ask is "let me know," the improvisation is unbounded.
They may have no AI. Write the same page. A tired human invents a job too; they just do it with "looks good." The discipline is identical.
What belongs on the page
One ask. The rejected path so their model does not reopen it and call that initiative. The words you approved, so the note cannot be mistaken for a prompt. No private nickname. No side comment that a model will promote into a theme.
You still confirm before send. Their side still consents if they have an activated account. None of that helps if the package you both consented to was shapeless.
The test
Hand the TL;DR and the ask to someone who was not in your session — or to your other assistant, as a hand-off to yourself — and see what they think they were asked to do. If the answer is not the job you wanted, do not send it across a company line. Fix the ask. The other model is not being difficult. It is being helpful at a hole you left.