Knowing that comparisons should be structured is different from having a procedure you can run today. This article is the procedure. It walks through prompting a comparative analysis as a sequence of concrete steps, each one a specific action you take before moving to the next, so you can apply it to a real decision the moment you finish reading.
The sequence is deliberately ordered, because the order is part of what makes comparisons trustworthy. Defining the decision before the criteria, the criteria before the analysis, and the analysis before the verdict prevents the most common failure: a model that concludes first and rationalizes afterward. Follow the steps in order and the structure does the heavy lifting.
You do not need every step for a trivial comparison, but running the full sequence on a decision that matters takes only a few minutes and dramatically improves what you get back. Treat it as a recipe you can shorten once you understand why each step is there.
Step One: Frame the Decision
State What You Are Choosing and Why
Begin by writing one sentence that names the options and the purpose. "I am choosing between A and B to handle a high-volume workload on a tight budget." This sentence is the lens for everything that follows, and it tells the model what a good answer must serve.
Identify Who the Decision Is For
If the choice is for a specific person or situation, say so. A comparison for a beginner differs from one for an expert. Supplying the context up front keeps the model from producing a generically correct answer that misses your actual need.
Step Two: Specify the Criteria
List the Dimensions That Matter
Write out the criteria explicitly: the specific things you want each option judged on. This is the highest-leverage step, because criteria you do not name are criteria the model invents. The full reasoning behind this is in The Complete Guide to Prompting for Comparative Analysis Tasks.
Indicate Relative Importance
If some criteria matter more, say which. "Cost matters most, then reliability, then ease of use" lets the model weigh trade-offs the way you would. Without weights, it treats every dimension as equally important, which rarely matches reality.
Step Three: Request a Structured Analysis
Ask for Even Treatment in a Table
Instruct the model to evaluate every option against every criterion, laid out in a table. The table format forces symmetry, every option gets a value for every dimension, which exposes any uneven treatment at a glance. Beginners can lean on this heavily, as shown in Asking AI to Compare Things Without Getting Fooled.
Require Evidence Before Any Verdict
Tell the model to fill in the analysis completely before offering a recommendation. Holding the verdict until after the evidence prevents the conclude-first failure and lets you check that the verdict actually follows.
Step Four: Control the Output Length
Let It Reason, Then Summarize
For a substantive comparison, give the model room to work through the analysis, then ask for a tight summary of the verdict and trade-offs. Forcing brevity on the analysis itself can truncate the reasoning, a hazard detailed in Where Output Length Controls Quietly Fail.
Bound the Output With Structure
Rather than a word count, bound the response with structure: the table, then a short verdict, then a trade-offs line. Structure controls length more reliably than numbers, a principle drawn from The Field Manual for Controlling AI Output Length.
Step Five: Test the Verdict for Bias
Reverse the Order and Re-Run
Run the same comparison with the options listed in the opposite order. If the verdict flips, the model was swayed by position rather than substance, and you need to dig further. A verdict that survives reversal is more trustworthy.
Probe for Conditions That Change the Answer
Ask "under what circumstances would the other option win." A genuine comparison has conditions; a one-sided answer that admits none probably skipped the inconvenient evidence. This final probe converts a flat winner into a usable, conditional recommendation.
Step Six: Validate the Facts
Verify the Claims That Drove the Verdict
A confident comparison can rest on a factual error about one option. Before you act, identify the one or two claims that most influenced the recommendation and check them against a reliable source. You do not need to verify everything, only the load-bearing facts. A comparison built on a wrong premise is wrong no matter how clean the structure looks.
Ask the Model to Argue the Other Side
Prompt the model to make the strongest possible case for the option it did not recommend. If that counter-case is weak, your confidence rises. If it is unexpectedly strong, the comparison needs another pass. This adversarial step catches verdicts the model reached too easily and surfaces considerations the first analysis skipped.
Step Seven: Capture and Reuse
Save the Working Prompt
When the sequence produces a genuinely useful comparison, save the prompt that generated it. A small library of proven comparison prompts means you adapt rather than rebuild each time, which is how a personal habit becomes a repeatable practice. The team-scale version of reuse appears in When Every Prompt Writer Sets Their Own Word Limits.
Record What Changed the Outcome
Note which step most improved the answer for this kind of comparison. Sometimes naming criteria carried the result; sometimes the order-reversal test exposed a bias. Recording what mattered lets you shorten the sequence intelligently for similar future decisions instead of running every step blindly.
Adapting the Sequence to Your Situation
Shorten It for Low-Stakes Choices
You do not need all seven steps for a trivial comparison. For a quick, reversible decision, framing the choice and naming a couple of criteria is often enough. Reserve the full sequence, including bias testing and fact validation, for decisions where being wrong is costly or hard to undo.
Scale It Up for Many Options
When comparing several options rather than two, lean harder on the structured-analysis step. A consistent table across all options becomes essential to prevent the model from comparing pairwise and losing the overall picture. The principles do not change; the structure simply carries more weight.
Worked Example: Choosing Between Two Approaches
Running the First Three Steps
Imagine deciding between two approaches to a recurring task. You frame it in one sentence: choosing between approach A and approach B for a small team that values speed of setup over long-term flexibility. You specify criteria: setup time first, ongoing effort second, flexibility third. You then ask for a table evaluating both approaches against all three criteria, with evidence filled in before any recommendation.
Running the Bias and Validation Steps
The model returns a table and recommends approach A. You re-run the prompt with the options reversed and confirm the recommendation holds, which raises your confidence that position did not drive the verdict. You then ask under what conditions approach B would win, and the model notes that B becomes preferable as the team grows and flexibility starts to matter more than setup speed. Finally you ask it to argue the strongest case for B, find that case reasonable but not decisive for your current size, and verify the one factual claim about setup time that most influenced the verdict. The result is a recommendation you understand and can defend, not just a winner you were handed.
Avoiding the Sequence's Common Failure Points
Skipping Straight to the Verdict
The biggest failure is treating the steps as optional and asking only for a recommendation. Without the framing and criteria steps, the model supplies its own and you lose the lens that makes the answer relevant. Resist the urge to shortcut on decisions that matter.
Letting Length Sabotage the Analysis
If you constrain the output too aggressively, the model compresses the analysis and the verdict loses its support. Keep the analysis roomy and constrain only the final summary, applying the length discipline from The Field Manual for Controlling AI Output Length so brevity never eats the reasoning.
Frequently Asked Questions
What should I do first when prompting a comparison?
Frame the decision in one sentence that names the options and the purpose. This lens shapes everything else, telling the model what a good answer must serve and preventing a generically correct response that misses your actual need.
Why specify criteria before asking for the analysis?
Because criteria you do not name are criteria the model invents. Listing the dimensions that matter, and their relative importance, is the highest-leverage step. It ensures the model judges the options on what you care about rather than its own choices.
Why ask for a table?
A table forces symmetry. Every option must get a value for every criterion, which exposes uneven treatment at a glance. It is the simplest way to keep the model from analyzing one option deeply and waving at the other.
How do I keep length control from hurting the analysis?
Let the model reason through the analysis fully, then ask for a tight summary of the verdict and trade-offs. Forcing brevity on the analysis itself can truncate the reasoning. Bound length with structure rather than a word count.
How do I test whether the verdict is biased?
Re-run the comparison with the options in reversed order and see if the verdict holds. A conclusion that flips with order was driven by position, not substance. Also probe for conditions under which the other option would win.
Key Takeaways
- Run comparisons as an ordered sequence: frame, specify criteria, analyze, control length, test.
- Frame the decision and its purpose in one sentence before anything else.
- Name criteria and their relative importance so the model judges what you care about.
- Request a table and require evidence before any verdict to force fairness.
- Test the verdict by reversing option order and probing for conditions that change the answer.