ClaudeExtended thinkingLevel 3: Workflows

How Do You Get Claude to Work Through Math and Logic?

Give Claude the whole problem, raise effort, and ask it to check its result against the numbers you gave. Don't limit the answer's length or tell it to answer immediately. Then check the arithmetic yourself, or have it run in code, before you rely on it.

Prompt to copy 
Work this out carefully. Don't limit the length of your working.

Problem: [state it in full]
Known values: [every number, with units]
Rules: [each constraint, one per line]
What I want: [the exact quantity or yes/no]

Give the answer first, then a short explanation of the method. Before you finish, verify your answer against the known values and rules, and list any assumption you had to make because I didn't give you something.
Do
  • State the full problem
  • Raise effort
  • Ask for a check against your numbers
Don't
  • Ask for a one-line answer
  • Tell it to answer immediately
  • Size max_tokens for the answer alone
Rule of thumb, not an Anthropic chart: the settings and wording that leave Claude room to work a problem through.

Try it now

  1. 1Click the model name next to the send button, choose Effort, and pick High.
  2. 2Paste a problem whose answer you already know, using the prompt above.
  3. 3Expand the Thinking section above the response and look at which approach Claude took.
  4. 4You'll know it worked if the answer matches yours and the assumptions it lists are ones you'd accept. If it can run code in your account, ask it to recompute the result in code as a second check.

Use it when

Scheduling with constraints, unit economics, probability, proofs, spreadsheet logic, and word problems where one wrong step ruins everything after it. It helps most when the answer can be tested against something, such as a total that has to add up.

Skip it when

Don't use a chat model as your calculator for final figures on invoices, taxes or payroll. Use a spreadsheet or code you can inspect. For a quick mental check on one multiplication, this is more setup than the problem deserves.

Why it works

Anthropic's thinking docs describe the problem directly: a model that answers in a single pass has no scratch work and no chance to change approach. With adaptive thinking, a multi-step math problem typically triggers deeper reasoning, and effort sets how far that goes. Anthropic's prompting guide adds that appending "Before you finish, verify your answer against [test criteria]" catches errors reliably, especially in coding and math. It also says general instructions like "think thoroughly" often beat a hand-written step-by-step plan, so you don't need to dictate the method.

Example

A warehouse has 3 loading docks and 11 trucks with fixed arrival windows, and each truck takes 40 minutes to unload. You ask whether all 11 can be served by 2 p.m. and which trucks are at risk. You paste the windows as a table. The self-check matters here, because you can ask Claude to confirm that no dock has two trucks at the same minute, which you can then check by eye.

Going further

On the API, max_tokens is a hard cap on thinking and answer text combined, and Anthropic warns that a value sized for a response without thinking is often too small. If you see stop_reason: "max_tokens", raise the cap when the reasoning mattered and lower effort when it didn't. The SDKs require streaming above 21,333 max_tokens. One exception to the self-check advice: Anthropic says Opus 5 verifies its own work well, and carried-over verification instructions can cause over-verification, so remove them on that model.

Common mistakes

  • Asking for "just the number" on a hard problem, which leaves you nothing to check.
  • Telling Claude to write its reasoning in <thinking> tags. Anthropic warns that on Opus 5.5, Sonnet 5.5 and Fable 5.1 this may be declined. Use a self-check instead.
  • Leaving out units or a constraint, then trusting an answer that quietly assumed one.
  • Treating the summary in the Thinking section as proof. It shows the approach, not a verified calculation.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Best practice How Do You Tell Claude What a Right Answer Looks Like? Mistake Why Does Claude Sound Sure When It's Wrong? Best practice How Do You See Claude's Thinking, and Can You Trust It? Technique How Do You Generate a Typed API Client With Claude?

Beyond the course

Glossary Model Context Protocol (MCP) Pricing guide Claude pricing, plans and models