I had one of my AI sessions go fetch some test output from another session. The request it wrote was clear, polite, and well organized. Then it tacked on a last line I've written a hundred times myself:
"This may be the whole explanation, in which case just confirm it and skip the rest."
That line pretty much guarantees a useless answer.
The test
I was writing a short rulebook for how my AI sessions ask each other for help. To test a rulebook honestly, you run the job twice: once with no rules so you can see what goes wrong on its own, then again with the rules to see what changes.
The no-rules run got almost everything right. It named what it wanted, said where to find it, and was readable and courteous. The one thing it got wrong was the thing the whole message was for.
Your guess comes back as evidence
When you hand someone your guess along with the question, you've stopped asking. You're offering a conclusion and inviting agreement.
Walk it forward. The other session reads "this may be the whole explanation," goes and looks at the thing I suspected, and finds something that roughly matches, because something usually does. It confirms. It never opens the other three possibilities, because I already told it the search was probably over.
So my own guess comes back to me dressed up as evidence, and now two sessions agree on an idea only one of them ever had. The disagreement that would have caught the mistake never happens.
People do this too. Ask a coworker "is the database down? I think the database is down" and they'll go check the database. Ask "the app stopped responding at 4pm, what do you see?" and they'll look at 4pm with their own eyes. That difference only matters when you're wrong, and being wrong is the one thing you can't see from the inside.
"No rush" when you're stuck
The same message had a second problem, and I recognized this one right away because it's one of mine.
The session couldn't do anything else until it got the answer. It was stuck. What it wrote was "no rush."
That's good manners and a false status report. The other end schedules around it, so a "no rush" request gets picked up after whatever sounds urgent, and you wait an hour for something you needed at the top of the hour. Softening a request feels generous, but it just moves the cost onto someone who can't see it.
The three rules I ended up with
- Put the request first. First line: what you want, where it lives, and what done looks like.
- If you have a theory, put it last and label it as a guess. Something like "my guess is X, but check the other cases anyway." Leaving it out entirely is usually better.
- Say whether you're blocked. "I can't continue until I have this" is the most useful sentence in the message.
I wrote these for agents talking to agents, because that's where I could see the damage in a transcript instead of guessing at it. About a day later I noticed I break all three in my own emails.