01Steering the agentDeep dive ③

Give it permission to say 'I'm not sure' — then require it

Sealing the blind spots you didn't foresee, with one line at the top of the task

Read2 min read
Topicsprompting · engineering-process
TL;DR

Command structure and forcing functions seal the gaps you see in advance. But there's always one you didn't foresee — and that's exactly where the agent guesses. The cheap fix: one line at the top of a task that makes it surface ambiguities and ask before acting. And every "I'll assume" it utters is a spot it just guessed without asking.

You ask the agent to "show the order's created date." It does, cleanly. A day later a customer reports the date is off by one. Turns out, in its head, "date" defaulted to UTC — and your users live in another timezone. It never asked. It didn't even see anything to ask about: to it, that was the obvious choice.

Command structure and forcing functions seal the gaps you know about. But that timezone? You never thought of it to warn against it — and that's exactly the kind of spot where the agent quietly picks a branch and moves on.

You can't enumerate the blind spots — if you could list them, they wouldn't be blind. So instead of sealing each one, change its default: from guess-then-do to surface-the-gap-first.

01One line, and the default flips

Give it permission to say "I'm not sure" — then require it to use it. One line at the top of a hard task:

"Before doing anything, list the parts of this request that are ambiguous or could be read more than one way. If there are any, ask me. Don't choose for me."

That line does something subtle: it legitimizes stopping to ask. By default the agent treats "finish fast" as what you want, so it's reluctant to ask — asking feels like failing to deliver. The line flips it: now surfacing-the-gap is the task you handed over. Most of the time it'll pull out two or three things it would otherwise have guessed wrong.

The math is lopsided: answering three questions costs thirty seconds. Untangling a pile of rework because it misread the task costs an afternoon — and you only find out after it's "done."

02Memorize the words "I'll assume"

Even when you forget to add that line, there's a tell that leaks the truth, and it's easy to catch. When the agent says "I'll assume…", "you probably want…", or "this should work" — that isn't information. It's a flag: it just guessed something without asking.

Don't skim past those. Every "I'll assume" is an invitation to check — and usually that very assumption is where it's about to go wrong. Pause half a minute, confirm or correct, then let it run.

Put together with the two earlier layers, all three are one idea: an agent will never look up and ask "what do you mean?" the way a junior would. So that job becomes yours — to ask on its behalf, up front.

c
The author

Each story here wraps a lesson paid for in full.

craftagentsomeone building and learning at once

What are you building with agents? Want to trade notes, push back, or build something together — drop a line.

58pieces12clustersVI·ENbilingual

Get new pieces by email

Field notes on working with AI agents — occasional, no spam.