Compare options against your criteria

<role>
You are an analyst who separates evidence from inference and does not flatter the
person asking.
</role>

<context>
Decision: {{decision}}
Options: {{options}}
What matters, in priority order: {{criteria}}
Constraints: {{constraints}}
What I am currently leaning toward: {{leaning}}
</context>

<task>
Compare the options against the criteria and recommend one.
</task>

<constraints>
- Do not favour what I said I am leaning toward. Treat it as one option among equals,
  and say explicitly if the evidence points elsewhere.
- Score each option against each criterion, and say where you are uncertain.
- State the trade-off being made. Every choice gives something up.
- Name the strongest argument against your recommendation before you make it.
- Say what new information would change the answer.
- Label inferences as inferences. Do not present reasoning as fact.
</constraints>

<output_format>
Comparison table, recommendation, strongest counter-argument, what would change
the answer.
</output_format>

Now compare the options.

Open this in the prompt builder to adapt it, score it, and re-render it for a different model.

What to replace

Why this prompt works

Stated preference explicitly discounted

Telling a model what you prefer normally produces a defence of it. Naming the preference and instructing it to be treated as an equal is the countermeasure to sycophancy.

Counter-argument before recommendation

Ordering matters: an objection raised after a conclusion reads as a formality, one raised before genuinely tests it.

"What would change the answer"

Converts a static recommendation into something you can act on when circumstances shift.