The short version

“Check your work” describes an intention, not a test. A stronger prompt defines the conditions the answer must satisfy: required sections are present, important claims are grounded, calculations reconcile, and uncertainty is labeled. Claude can then compare the draft with those conditions, repair failures, and return the result.

Use a bounded verification loop

A practical sequence is draft, test, fix, finish. The test should focus on properties that are observable in the output or its supporting evidence. It should also have a stopping point. Once the named conditions pass, repeated reconsideration usually adds cost rather than confidence.

Good criteria tend to cover four dimensions:

  1. Purpose: Does the output answer the actual question and support the intended decision?
  2. Evidence: Can important factual claims be traced to the supplied sources?
  3. Structure: Are all required fields, sections, limits, and calculations present?
  4. Exceptions: Is material uncertainty identified instead of being silently resolved?

Example

Suppose Claude is preparing a management variance brief. “Be very careful and check everything” gives no priority. A useful instruction is more concrete:

Draft the brief, then verify that every percentage agrees with the source table, every variance above the stated threshold is addressed, each causal claim is supported by the supplied evidence, and forecasts are labeled as estimates. Correct any failure before returning the final version.

The second instruction directs attention to failure modes that matter. It also avoids sending Claude into an open-ended search for ever more certainty.

Match verification to the task

For code, criteria might include tests, type checks, and expected edge cases. For document work, they might include required sections, source coverage, length, and tone. For quantitative analysis, use reconciliation rules and boundary checks. The best criteria are specific enough to evaluate but small enough to run consistently.

Verification is not a substitute for external controls. A self-check can improve a draft, while high-impact decisions may still require deterministic validation or human approval outside the model.

Practical checklist

  • Define success before asking Claude to verify.
  • Prioritize likely and costly failure modes.
  • Use evidence and measurable conditions where possible.
  • Ask Claude to repair failures, not merely report them.
  • Stop after the criteria pass unless new information appears.
  • Keep external approval for consequential decisions.

Try it

Replace “double-check this carefully” in one recurring prompt with four checks: the intended purpose, evidence standard, required structure, and condition that must be escalated.