When an AI gives you a hedging answer, full of "it could be this, but it might also be that", the usual cause is that it is missing something you know and have not told it. It is rarely being careful. The fix is not to rephrase the question or switch tools. It is to add one missing fact, one example or one rule at a time, and run it again.
Is a hedging answer a sign the AI is being careful?
It can look that way. A cautious, balanced answer reads as responsible, so most people accept it and move on, or decide the tool is not up to the job.
In practice, hedging tends to track how much the AI has to go on. Give it very little and it covers every possibility. Give it the context a person would need and it commits. The caution is a symptom of missing information, not a virtue.
What does that look like in practice?
Anthropic ran a short prompting workshop that shows it well. The task was to work out who was at fault from a car accident claim form. The first attempt, with almost no instructions, came back saying the form described a skiing accident. Nobody had told it what the form was.
What happened next is the useful part. The presenters did not rewrite everything from scratch. They added one thing at a time:
- what the job was, and who it was for
- what the boxes on the form meant, since people mark them in messy ways
- a few tricky past cases, with the right answers
- a rule: only call fault when you are sure
Each addition made the answers firmer. The hedging faded as the information went in. The confidence was earned with information, not demanded in the instructions.
Why does one-at-a-time work better than rewriting?
Because you learn what matters. If you rewrite the whole request and the answer improves, you do not know which change did it. If you add one fact and the answer firms up, you have found out something about how the AI is reading your situation, and you can keep that.
It also keeps you from the other common mistake, which is to bury the AI in a long, anxious wall of instructions. A small, well-chosen addition usually beats a big one.
What should I add first?
Think about what a new starter would need to know before they could commit to an answer. That is usually one of four things.
The situation: what is this, who is it for, what is it used for. The meaning of the data: what do the columns, boxes or codes actually stand for. A tricky example: a past case where the answer is not obvious, with the correct answer attached. And a rule: the thing it must or must not do when it is unsure.
Start with whichever the hedging suggests is missing. If the answer is vague about what something is, add the situation. If it is wrong on the odd case, add an example.
How is this different from writing a better prompt?
It is a loop rather than a document. You are not trying to write the perfect instruction on the first go. You are running the request, looking at where the answer goes soft, and adding the one thing that would settle it.
At Handiwork, when an answer comes back wishy-washy, the first thing we ask is what a new starter would need to know before they could commit. Then we add exactly that, one thing at a time, and run it again. It works for people who will never write a technical prompt, which is most of the owners we work with. If you want the longer version of the examples step, see how to train AI with your own examples.
What to do this week
Next time you get a hedgy answer, do not rephrase the question. Add one missing fact, one example of a tricky case, or one rule, and see what changes. Keep the version that worked, because it is the start of a reusable setup. If you want to turn that setup into something that runs reliably rather than a one-off chat, a prototype is not a system explains the gap, and our practical AI workflows guide covers what to build first. Real examples of this kind of work are on our use cases page.
Ready to find out where you stand?
If you want help working out what your AI setup is missing, our free AI Readiness Check takes about five minutes: AI Readiness Check.
Frequently asked questions
Why does my AI keep saying "it depends"?
Usually because it does not have enough context to commit. Add the one fact that settles the "depends" and ask again.
Should I tell the AI to be more confident?
Not by itself. Asking for confidence without giving it the information tends to produce confident guesses. Confidence follows information.
How many examples should I include?
A few. Pick the tricky cases, not the easy ones, and attach the correct answer to each. Three or four well-chosen cases do more than twenty obvious ones.
Does this work with any AI tool?
The loop does. The principle that a vague answer points to missing information holds across tools, though each has its own way of letting you save the context so you do not retype it.
What if adding facts does not help?
Then the task may genuinely be ambiguous, or the AI may not be the right tool for it. At that point the answer is a useful finding, not a failure.
Ready to find out where you stand?
Take the free five-minute AI Readiness Check. There is no pitch at the end of it.
Take the AI Readiness Check
Sources
- Anthropic's Claude prompt engineering workshop (insurance claim form example), as shared by Jonathan Peslar on LinkedIn - https://www.linkedin.com/posts/jonathan-peslar_anthropic-just-showed-a-24-minute-workshop-ugcPost-7474612325057187840-mCc4/
