Tell Thinking What a Good Answer Must Satisfy

Giving Claude an explicit definition of what a correct answer must satisfy — the constraints, the acceptance criteria, the must-haves — before it starts thinking, so its reasoning has a target to steer toward and check itself against, rather than optimising for 'a reasonable-sounding answer'.

Extended thinking is powerful but undirected by default: it will reason its way to a plausible answer, which is not the same as an answer that meets your actual requirements. When you state the success criteria up front, the reasoning has something to aim at and, near the end, something to test against — 'does this satisfy all five constraints?' That converts open-ended thinking into goal-directed problem-solving, and it catches the common failure where a fluent, well-argued answer quietly misses a requirement you cared about most.

Problems with real acceptance criteria: a plan that must respect a budget and a deadline, a design that must meet named requirements, a solution that has to satisfy several constraints at once. Anywhere 'sounds good' and 'is correct' can come apart.

Genuinely open exploration or brainstorming, where imposing success criteria too early would prune the divergent thinking you actually want. Define the target once you're converging, not while you're still searching.

Before you start reasoning, here is what a correct answer must satisfy:
- <constraint / criterion 1>
- <criterion 2>
- <criterion 3>

Think through the problem with these as your target, and before you finish, explicitly check your answer against each one and note any it fails. Problem: <the problem>.

You ask for a launch plan and first get a thoughtful timeline that quietly assumes a team of five and a three-month runway you don't have. Restated with the criteria up front — 'two people, six weeks, no paid ads' — the thinking steers within those limits and ends by confirming each, producing a plan you can actually execute rather than an idealised one.

Phrase the criteria as a checklist and ask Claude to end its thinking with a pass/fail line against each item, plus what it would take to satisfy any it failed. You get not just an answer but an honest gap report — which is far more useful than a confident answer that silently traded away a requirement.

• Leaving the real constraints in your head, then being surprised the answer ignores them — unstated criteria can't be satisfied.
• Listing so many must-haves that they conflict, without saying which win — force the priority order.
• Accepting the answer without reading its self-check, so a 'fails criterion 3' note you asked for slips past you.