How to Prompt AI for a Recommendation, Not Options
Models hedge because most questions are underspecified. Four forcing functions that produce one ranked recommendation with a named tradeoff you can actually check.
How to Prompt AI for a Recommendation, Not Options
To get a recommendation instead of options, you have to remove the model's escape routes: give it the constraints that make one answer better than another, require it to rank rather than list, make it name what it is trading away, and forbid the hedge explicitly. A model defaults to balanced surveys because balance is the safest response to an underspecified question, and most questions people ask are underspecified. The fix is mostly on your side of the prompt.
This is the mirror image of a problem worth knowing about: sometimes you genuinely want breadth, and there is a technique for getting multiple options instead of one answer. If you are early in a decision, use that one. This post is for when you have done the exploring and want somebody to call it.
Why models hedge
Three reasons, and knowing which one is biting tells you which fix to apply.
Your question has no single right answer as written. "Which database should I use?" has no answer without knowing the workload. The model is not dodging, it is correctly reporting that it lacks information you did not give it.
Hedging is trained behaviour. Balanced, caveated answers are rated as higher quality and lower risk across a huge range of questions, so the default leans that way even when you have supplied enough context.
The format invites a list. Asking "what are my options for X" gets options because that is what you asked for. People do this more often than they notice.
Four forcing functions
1. Supply the constraints that break the tie
A recommendation is only possible if something distinguishes the candidates. Give the model your actual constraints: budget, timeline, team size, existing stack, risk tolerance, what you have already ruled out and why. The last one is worth more than the rest combined, because it stops the model re-proposing what you already rejected.
A useful sentence to include verbatim: "If you need more information to make a single recommendation, ask me at most three questions first, then recommend." That converts the hedge into a question, which is the useful form of it.
2. Require a ranking, not a list
Ask for the candidates ranked from best to worst for your specific situation, with the top choice stated first as a single sentence. Ranking forces comparison on a shared axis, which is the cognitive work you actually want done. A list is just parallel descriptions, and it leaves the comparing to you.
3. Make it name the tradeoff
Add: "State what your recommendation gives up compared to the runner-up, and under what conditions you would change your answer." This is the highest-value line in the whole technique. A recommendation with a named cost is checkable. If the stated tradeoff does not apply to you, you have learned the recommendation is wrong for your case, which is more useful than a confident answer you cannot evaluate.
4. Forbid the hedge explicitly
Say it in the prompt: no "it depends", no "both are good choices", no restating the question back as a summary. Models comply with explicit format prohibitions more reliably than with a general request to be decisive, because a prohibition is checkable and "be decisive" is not. This is the same principle vendors put at the top of their own prompting guidance: clarity and explicit direction do more than any clever phrasing.
Before and after
The weak version, which produces a table of four databases and no decision:
What database should I use for my app?
The version that produces a recommendation:
I am building a booking app for a single business, roughly 500 bookings a month, one developer, deployed on a managed platform. I have ruled out self-hosting because there is nobody to maintain it. Rank the realistic options from best to worst for this situation. Lead with your single recommendation in one sentence. Then state what it gives up versus the runner-up and what would change your answer. Do not say it depends.
The second prompt is longer, and the length is entirely constraints rather than instructions about tone. That ratio is the tell for a good decision prompt.
How to sanity check the answer
A committed recommendation is easier to be wrong about, so check it rather than take it.
Ask for the strongest argument against its own recommendation. If that argument is weak or generic, the recommendation was probably shallow.
Run the same prompt twice in separate conversations. Different answers mean the constraints are not doing their job, which is a slightly different issue from models giving different answers to the same question in general.
Check that the stated tradeoff is real and applies to you. This catches plausible-sounding recommendations faster than checking the recommendation itself.
For anything consequential, treat it as one input. A model with your constraints is a good structured second opinion, not a decision maker.
If you find yourself doing this repeatedly for the same class of decision, write the constraints once and reuse them, which is the same economy that makes prompts that work across models worth the effort of writing properly.
FAQ
Why does AI keep saying it depends?
Usually because it genuinely does, given what you told it. Supply the constraints that resolve the dependency, or ask the model to tell you which three facts would change its answer and answer those. A hedge is often a correct response to a missing constraint.
Is a forced recommendation less accurate than a balanced answer?
Not if you supplied real constraints, because you have narrowed the question rather than removed the reasoning. It is less accurate if you forced a decision on a question the model has no basis to decide, which is why the named-tradeoff line matters: it exposes that case.
Does telling a model to be confident make it hallucinate more?
Suppressing hedging language does not make the underlying reasoning better or worse, but it does remove a signal you were using. That is the argument for requiring the tradeoff and the counter-argument in the same prompt rather than just demanding confidence.
What about high-stakes decisions?
Use the same structure and then discount the output further. The value is the comparison being made explicit, not the verdict. For decisions with legal, medical or financial consequences, the model is a way of organising the question for the human expert you should be asking, and general caution about where AI advice carries real risk applies.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
