The useful result
Find the assumptions that matter before using an AI-generated recommendation.
- Find the assumptions that matter before using an AI-generated recommendation.
A plan can look coherent because it quietly assumes away the hard parts. Asking for a counterexample makes those assumptions visible: under what realistic conditions would the recommendation fail?
This is an editorial reasoning exercise for everyday planning. The model's critique is not independent evidence, and agreement between two generated answers is not proof. Anthropic's prompting guidance informs the use of explicit context and constraints; the decision checklist is our own.
Review evidence
Illustrative example · 2026-10-07
Practice examples are fictional or illustrative. Only the spreadsheet fixture arithmetic is checked locally; no recorded external-model comparison is claimed.
State the decision and evidence
Write the decision, constraints and known facts. Keep forecasts and preferences in a separate list. Use a low-stakes example such as deciding whether to publish a weekly newsletter, not a medical or financial decision.
A counterexample is useful when it tests the actual recommendation rather than an exaggerated version of it.
Example / expected result
Illustrative decision: start a weekly newsletter. Known constraint: one editor has two hours per week. Unproven assumption: enough useful material will be available every week.
Check your result
Check that the facts are actually known and that forecasts are labeled as assumptions.
Ask for realistic failure conditions
Request a small number of plausible scenarios and the evidence that would support or rule out each one. Do not ask the assistant to invent proof for its own critique.
Review this plan using the stated facts and constraints.
Identify three important assumptions.
Give one realistic counterexample that would make the recommendation fail.
Explain what evidence would confirm or rule out that scenario.
Distinguish supplied facts from your speculation.
Do not invent market statistics or customer feedback.
Suggest a small reversible test instead of a guaranteed conclusion.Example / expected result
Illustrative counterexample: publishing quality drops because the editor cannot verify sources within two hours. Evidence needed: time spent on a small pilot.
Check your result
Verify that the scenario is plausible under your actual constraints and that the evidence has not been fabricated.
Turn the strongest assumption into a test
Choose the assumption most likely to change your decision. Design a small test, define the observation you need and decide what result would trigger a change.
For the newsletter example, draft one issue and record actual research and review time. That measurement is more useful than asking the model to estimate your productivity with confidence.
Example / expected result
Expected decision note: continue, revise or pause based on observed pilot effort. No revenue or growth prediction is established by this exercise.
Check your result
Make sure the test can produce an outcome that would genuinely change the decision.
When it goes wrong
The same model can repeat the same blind spot when criticizing itself. Use outside evidence and human judgment. A dramatic worst case is not automatically a realistic counterexample. High-stakes decisions need appropriate professional review.
Make it your own
Choose a low-stakes plan you are considering. Write one assumption, one counterexample and one observation that would change your mind.
Useful next resources
Sources and review
Last checked Oct 6, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
AI-assisted editorial lessons researched using linked primary sources. Worked examples are illustrative unless a live test is explicitly documented.
