Because you asked for research and you got research. It arrives comprehensive, balanced, well organised and completely inert, and you finish it better informed and exactly as stuck as you were. Balance is the opposite of a decision. The fix is to force the output into a fixed shape that ends in a recommendation with its strongest counter-argument attached.
The counter-argument requirement is the whole design, and I will explain why it does more work than it looks like it should.
The problem is the request, not the output
Ask a broad question and you get a broad answer, and a broad answer on a decision is a list of considerations.
Considerations are not useless. They are just not the thing you were short of. You already knew there were considerations, which is why you were stuck. What you were missing was someone willing to weigh them and commit, and nothing in a research request asks the output to commit to anything.
So the analysis is good and the hour is gone and the decision is still open.
The four-part shape
State the decision as a decision, then require four things.
What are the real options, including doing nothing? Doing nothing is almost always a live option and it is almost always missing from the list, which quietly biases everything that follows toward action.
What does each option cost, in money, time and reversibility? Reversibility is the one people leave out and it frequently matters most. A cheap decision you cannot undo is more expensive than an expensive one you can.
What would have to be true for each option to be the right call? This reframes the whole thing usefully, because now you are evaluating claims about the world rather than arguing about preferences, and claims about the world can sometimes be checked.
What is the recommendation, and what is the strongest available argument against it?
Read the counter-argument first
This sounds like a stylistic preference and it changes outcomes.
If you read a recommendation that agrees with what you already wanted, every argument against it subsequently arrives as an obstacle to be got past. Your reading is defensive before you started. Read the same argument first, before you know which way the recommendation goes, and it arrives as information, and you weigh it properly.
It costs nothing and it is the single highest-return habit in this entire playbook.
We have written about the conversational version of this in how to use Claude to think through a decision and, specifically on the counter-argument technique, in how to get Claude to argue against your own ideas. This post is the version that runs as a system rather than a conversation.
The Research to Decision Playbook
Fifth of the six. It produces a decision memo, one page, ending in something you can act on or reject for a stated reason.
The build.
- State the decision as a decision. Not "research pricing models" but "I am deciding whether to move from one tier to two by October."
- Load your context brain, including the decision log, so previously rejected options do not come back around as fresh suggestions.
- Require the four-part shape.
- Read the counter-argument before the recommendation.
- Write your decision and your actual reason into the decision log.
Step five closes the loop. The decision log feeds the context brain, the context brain makes the next memo better, and after a few months the system knows what you have already tried, which is most of what a good advisor knows.
The trap
Accepting a counter-argument that was written weakly.
If the case against sounds easy to dismiss, it probably was written to be easy to dismiss, because the request was ambiguous about how hard to push. Ask specifically for the strongest version, from someone who thinks the recommendation is wrong, and read it again.
The other failure is subtler. Running this on decisions that do not need it. A one-page memo on a reversible decision that costs forty dollars is a way of not deciding while appearing rigorous, and the honest move is usually to pick one and see.
Frequently asked
Why does AI research never help me decide? Because balance and comprehensiveness are the opposite of commitment, and nothing in a research request forces the output to take a position.
How do I use AI for decision making? Require a fixed shape: the real options including doing nothing, the cost of each in money, time and reversibility, what would have to be true for each, and a recommendation with its strongest counter-argument.
Should I trust an AI recommendation? Treat it as an input rather than a verdict. The genuinely valuable part is the structured counter-argument, because most of us cannot build a strong case against our own preference.
What should go in a decision log? The decision, the date, the reason, and the options you rejected and why. The rejected options turn out to be the most useful part later.
When should I not use this? On small reversible decisions. A one-page memo on a cheap, undoable choice is a way of avoiding the decision while looking thorough.