Verdict: '95% certain' sounds rigorous until you remember the model has no certainty meter — it'll ask one question or twelve, and neither feels wrong to it.
see the specimen they pasted
Act as my research assistant. Ask me questions until you are 95% certain you understand what I need, then give me the best possible answer.
“Ask me questions until you are 95% certain”
The model can't measure its own epistemic confidence — '95% certain' is a number you invented to sound precise, but it's as actionable as 'ask until you feel good about it.' You've given the model an unbounded loop with no exit condition.
swap: 'until you are 95% certain' → 'ask me up to 3 clarifying questions, one at a time, then proceed'
Give the model a concrete stop rule — a count, a checklist, or a hard gate — never a self-assessed confidence threshold it cannot compute.
“give me the best possible answer”
'Best possible' is doing the same nothing that 'world-class' does on a job title — it's a superlative with no spec. Best in what shape? What length? What format? The model will just default to its most average output and call it best.
swap: 'give me the best possible answer' → 'deliver your answer as: a 2–3 sentence summary, then bullet-point details, under 300 words'
Always give the model a format contract: length, structure, and what to avoid — 'best' is not a format.
“Act as my research assistant”
'Research assistant' is a mood board, not a brief — it tells the model to vibe like a helper without specifying what kind of research, what depth, what sources to prioritize, or what the output is actually for.
swap: 'Act as my research assistant' → 'You are helping me research [topic area]. Prioritize accuracy over comprehensiveness; flag anything you're uncertain about.'
A role earns its place only when it narrows behavior — pair any persona with a concrete domain and a behavior cue, or drop it.
You are helping me research [topic — one sentence describing the subject and purpose, e.g. 'the competitive landscape for AI writing tools for SMBs']. Before answering, ask me up to 3 clarifying questions — one at a time, wait for my reply before asking the next. Stop after 3 questions or when you have enough to proceed, whichever comes first. Then deliver your answer in this format: - **Summary** (2–3 sentences: the core finding) - **Key details** (bullet points, max 8) - **Gaps or caveats** (anything you're uncertain about or couldn't verify — be explicit, don't omit) Rules: - Prioritize accuracy over comprehensiveness. If you don't know something, say so. - Stay under 400 words unless I ask for more. - No filler phrases like 'Great question' or 'Certainly'.
/roast/example/unmeasurable-instruction