Diagnosis: Solid bones, but 'options with trade-offs' is unbounded and '95% certain' is unmeasurable — the model will guess at both.
see the specimen they pasted
You are a top 0.5% expert in [field of expertise]. I want to [goal]. Provide options with trade offs for each. Here are the [background and context]. Ask me any questions until you are 95% certain you understand my asks. Do not generate the response until you do.
“Provide options with trade offs for each”
No count, no shape — the model returns 2 options or 9, with one-liners or essays, and all of it technically satisfies this instruction.
swap: 'options with trade offs' → 'exactly 3 options; for each: 2 pros, 2 cons, and who it suits best; max 120 words per option'
Pin every open-ended output: give a count, a shape per item, and a word cap — otherwise the model picks all three at random.
“Ask me any questions until you are 95% certain”
A model has no instrument for measuring its own certainty — this instruction produces either one question or an interrogation, depending on the run.
swap: 'until you are 95% certain' → 'ask up to 3 clarifying questions, one at a time, then proceed'
Replace self-measured certainty thresholds with a concrete stop rule: a question cap and a trigger to proceed.
“top 0.5% expert in [field of expertise]”
Percentile rankings are decorative — they add no behavioral instruction the model can act on, and the role still has no audience, stakes, or communication style.
swap: 'top 0.5% expert' → '[field of expertise] expert advising a [audience — e.g. "non-technical founder"] — be direct, skip jargon, flag risks first'
A role prompt does real work only when it specifies audience and communication style, not just a prestige label.
You are a [field of expertise — e.g. 'SaaS pricing strategist'] advising a [audience — e.g. 'non-technical founder']. Be direct, skip jargon, and flag the biggest risk of each option first. My goal: [goal — one concrete sentence, e.g. 'choose a pricing model before our public launch in 60 days'] Background: [background and context — key constraints, prior decisions, and what a bad outcome looks like] Before you respond, ask up to 3 clarifying questions, one at a time. After my last answer — or if you already have enough — deliver the following: **Options** (exactly 3) For each option: - Name + one-line summary - 2 pros - 2 cons - Best for: [who this suits] - Biggest risk Max 150 words per option. **Recommendation** (1 short paragraph): which option fits my situation best and why. Do not start generating options until the clarifying questions are complete.
https://getmeerkat.dev/roast/m6r7ncb3v2