ClaudeVerification mistakesLevel 2: Context

Which Decisions Shouldn't You Leave to Claude?

Anything that's hard to reverse, affects other people, or depends on context Claude can't see. Anthropic's Usage Policy goes further for high-risk areas such as legal, medical, finance and employment: a qualified person has to review the advice or decision before it's delivered. Let Claude draft and analyze, and make the call yourself.

Prompt to copy
I have to decide: [the decision]. I'll make the final call, so help me make it well.

Give me:
1. Your recommendation, in one sentence.
2. The assumptions behind it, marking any you can't verify.
3. What you can't see from here that could change it.
4. The two things I should check, and who could confirm them.
5. What it would cost to reverse this if it turns out wrong.

Context: [the relevant facts]
  1. Easy to undo, just youClaude drafts, you skim
  2. Affects other peopleClaude drafts, you decide
  3. Hard to reverseyou decide, and limit what tools can do
  4. High-risk areasa qualified person must review and can change it
Rule of thumb for the lower rungs, not an Anthropic chart. The top rung is from Anthropic's Usage Policy, effective 12 November 2026, checked 6 November 2026.

Try it now

  1. 1Choose a real decision you're using Claude for, and paste the prompt above.
  2. 2Read item 3, and add anything you know that Claude couldn't have known. Ask whether it changes the recommendation.
  3. 3In Claude Code, choose Yes, manually approve edits when approving a plan for risky work, so you see every change.
  4. 4You'll know it worked when you can state the decision and your reason without pointing to Claude's answer.

Use it when

Code that ships, content you publish, advice someone will act on, hiring or lending decisions, and anything you couldn't easily undo. The more people affected and the harder the reversal, the more of the decision should stay with you.

Skip it when

Small, reversible tasks where a mistake costs seconds, such as a first-draft reply you'll read anyway. Reviewing everything equally wastes attention, and leaves you less of it for the decisions that need it.

Why it works

Claude doesn't see what's outside the prompt, such as your team's capacity or a client's history, and it doesn't bear the outcome. Anthropic's Usage Policy, in the version effective November 12, 2026, says that for high-risk recommendations a qualified person must meaningfully review them, with the authority to change them, and that people affected must be told AI was used. Its listed areas also include credit, insurance, housing, education and public benefits. There's an exception, where the law permits, for decisions wholly favorable to the person. Anthropic's hallucination guide adds that its techniques reduce errors but don't eliminate them, and it says to validate critical information, particularly for high-stakes decisions.

Example

Say Claude recommends a self-hosted analytics tool, based on price. Item 2 reveals that it's assuming your team can run a database, and you know that nobody on the team has time for it. So you keep the comparison, change the conclusion, and choose the hosted option. If you were screening job applicants instead, the policy's review rule would apply, and a qualified person would need to review the recommendation before it's acted on.

Going further

Check what your tool does without asking you. Anthropic's Claude Code guide says new terminal and VS Code sessions start in auto mode, where a classifier reviews risky actions and you aren't asked about each command. You can say "don't push until I've reviewed" in the conversation, but the docs warn that this can get lost when the conversation compacts, so anything that matters belongs in a deny rule. Anthropic's Fable 5 guide also offers a sample instruction for agents that pauses for irreversible actions and real scope changes. Limit what the tool can do, rather than relying on your attention.

Common mistakes

  • Reviewing only when something looks wrong. A confident error doesn't look wrong.
  • Assuming a quick glance counts as review. Anthropic's policy asks for meaningful review by someone with the authority to change the result.
  • Forgetting the disclosure. For high-risk advice or decisions, the policy says people must be told AI was used.
  • Telling Claude "never push" in the chat for a rule that matters. Use a deny rule, since a sentence in the conversation can be lost.

Sources

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

Related

Beyond the course