Why your AI assistant keeps guessing

You ask for a follow-up email. You get a follow-up email. It is polite, well structured, and wrong in four small ways: the wrong tone for this client, a meeting date nobody agreed to, a sign-off you would never use, and a price in the wrong currency.

The usual reaction is to blame the model. Try a different one, reword the request, add "please be accurate" in capitals. None of that helps much, because the model was not being careless. It was guessing, and it was guessing because you left it no other choice.

The desk

Here is the picture that explains most of it. Each time you send a message, the assistant sits down at a desk, reads everything on it from top to bottom, writes a reply, and leaves. Next message, it sits down and reads the whole desk again.

On the desk:

• Your standing instructions, if you have written any.
• The project's instructions and files, if you are inside a project.
• Every message in this chat so far, including the drafts you rejected.
• Whatever you attached or pasted.
• Anything it looked up during this answer.

Not on the desk: your other chats, your inbox, your calendar, the file you meant to attach, the conversation you had with the client yesterday, and everything in your head that you did not type.

That last item is the big one. You know the client is formal. You know the meeting has not been booked. You know you sign off with "Thanks," and you know the invoice is in Canadian dollars. The assistant knows none of it, and it will not stop to ask unless you tell it to.

What it does with a gap

When a fact is missing, the assistant fills the hole with the most likely answer. Most likely across everything it has read, which means the average of everybody else's situation.

For a generic task, the average is fine. For your task, with your client and your prices, the average is the problem. "Write a follow-up email" produces the median follow-up email on the internet. Every gap got filled with the most typical choice, and every typical choice was a little bit wrong for you.

This is also why the same question can get different answers on different days. The model did not change. The desk did: a longer chat, a different attachment, a project you forgot you were in.

A worked example

Here is a typical request, with the names made up.

Before:

Write a follow-up email to Priya about the proposal.

What came back opened with "I hope this email finds you well," suggested "a quick call next Tuesday," summarized a proposal it had never seen by inventing three plausible deliverables, and closed with "Best regards." Four guesses, four misses. The assistant did exactly what it was asked with what was on the desk.

After:

Write a follow-up email to Priya Shah, operations lead at a regional dental group. We sent her a proposal on Monday for setting up their booking assistant; it is pasted below. She has not replied.

Goal: get a yes or no on the start date, not a meeting. Keep it under 100 words. Warm but direct. No "hope this finds you well". Do not restate the proposal; she has it. Do not suggest a call. Prices are in CAD if you mention any. Sign off with "Thanks, Damon".

If anything you need is missing, ask before writing.

[proposal pasted here]

The reply was usable on the first try. Nothing about the model changed between the two versions. What changed is that six facts moved from my head onto the desk: who she is, what was sent, what I want, the length, the things to avoid, and the sign-off. The last line gave it permission to ask instead of invent.

Notice that most of those lines are not about Priya. "Keep it under 100 words," "no hope this finds you well," "sign off with Thanks, Damon," "prices in CAD": I would want those in almost every email I ask for. Typing them every time is a chore, and forgetting one means another round of edits. That is the point where a message should become a rule, which is the subject of the next post.

The four gaps that cause most bad answers

Most misses fall into four buckets:

1. Finish line. You did not say what done looks like: length, format, audience, the one thing it must include.
2. Background. You did not say who the reader is, what has been decided, or what went wrong last time.
3. Source material. You described a document instead of pasting it. A summary is a lossy copy, and the assistant fills the holes with plausible fiction.
4. The unguessable. Budget, deadline, currency, the word your boss hates, the fact that your readers are on a phone. No amount of intelligence infers these.

A good request covers the first three. A good setup covers the fourth once, so you never type it again.

Two habits that fix most of it

Ask for assumptions first. For anything longer than a paragraph, start with: "Before you write anything, list the assumptions you would have to make. Then wait." You will spot the wrong one in ten seconds, and correcting an assumption costs one reply instead of a whole draft.

Start fresh when the chat goes stale. A long chat is a pile of everything said so far, including instructions you have since changed. When the assistant starts drifting back toward something you already corrected, open a new chat, paste a two-line summary of what was decided, and carry on. The desk gets cleaner and the answers get sharper.

What this is not

This is not about clever prompt wording, magic phrases, or flattering the model with a grand job title. Those tricks move results a little. Putting the missing facts on the desk moves them a lot, and it works the same way in Claude, ChatGPT, Cursor, and Codex, because every one of them reads the same kind of desk.

Start here

The free primer, 10 rules that make your AI assistant stop guessing, turns this post into ten short habits you can use in your next chat. No code, no signup wall beyond an email.

If you already know the fourth gap is your problem, the Rules Builder asks you eight plain questions and hands you a rules file to paste into your assistant. It runs in your browser and keeps your answers there.

When you want the whole picture (how the desk fills up, how to read what the model did, where each kind of instruction belongs, and setup steps for each assistant), that is the AI Field Guide, $39 USD, one-time.