Read the Reasoning, Not Just the Answer

A habit of actually reading Claude's extended-thinking trace — not just the final answer — to spot where its reasoning took a wrong turn: a misread requirement, a bad assumption, a step skipped. The visible reasoning is a debugging surface for the answer.

When a thought-out answer is wrong, it's usually wrong at a specific point — Claude misunderstood one constraint, or assumed a value you never gave. Extended thinking makes that point visible. Instead of rejecting the whole answer with 'that's not right, try again' and hoping the next attempt lands, you can find the exact fork and correct just that: 'in step 3 you assumed the list is sorted — it isn't.' That's a precise redirect, not a blind retry, and it fixes the real cause rather than rerolling for a different guess.

Whenever extended thinking is on and the stakes justify a look: complex analysis, multi-step math, a plan you're about to act on, or any answer that feels subtly off. Especially valuable when a first answer is wrong and you want to know why, not just try again.

Low-stakes or obviously-correct answers where reading the reasoning is more work than the task deserves. The trace is a tool for when the outcome matters or something looks off — not mandatory reading for every reply.

Work through this with extended thinking, and make your reasoning easy for me to audit: state your assumptions explicitly and number the key steps. <the problem>. After the answer, list the two assumptions that, if wrong, would most change your conclusion.

Claude computes a pricing tier and lands on a number that looks high. Reading the thinking, you see it assumed annual billing when you meant monthly — one wrong premise in step two. You correct that single assumption and the rest of the (sound) reasoning produces the right figure, without re-explaining the whole problem.

Ask Claude to end its thinking with the load-bearing assumptions it relied on and how confident it is in each. This turns the trace into a targeted checklist — you verify the few premises everything hangs on, rather than re-reading every step, and you catch the failure mode where a confident chain rests on one shaky foundation.

• Skimming only the final answer on high-stakes work, then discovering the flawed assumption after you've acted on it.
• Rejecting a whole answer and re-rolling when a one-line correction to a single step would have fixed it.
• Treating a long, fluent thinking trace as proof of correctness — length is not the same as soundness.