Phrases such as “should work” and “likely fixed” express uncertainty, not execution. Ask what observation supports the conclusion rather than treating confidence as proof.

Claim → Supporting artifact → Missing check → Follow-up
  1. 1Claim
  2. 2Supporting artifact
  3. 3Missing check
  4. 4Follow-up

Work through the example

Prompt: Separate what you inspected from what you executed, and list the unresolved checks.

One practical example

Use the prompt above on one small, named part of your app. Before accepting an edit, ask the agent to point to the source or configuration that supports its answer. If it cannot locate that context, resolve the missing input before expanding the task.

In the diagram, each arrow represents a handoff you can inspect. Keep the intermediate result rather than jumping directly to the final claim. For example, a reported finding should retain its location, and an executed check should retain its outcome. This gives the next attempt a concrete starting point and makes a misleading completion message easier to challenge.

Apply it to your project

Prompt: Separate what you inspected from what you executed, and list the unresolved checks. Then compare the resulting diff with the original request. If it changed a neighboring concern, ask why that change was necessary. Save the result only after the relevant check has run, and record any part that remains unverified.

Evidence and limits

Wording alone cannot prove deception or correctness; inspect the evidence.

This is an educational guide. Its presence in the series does not certify a completed client-specific lab. The series evidence record separates executed checks from exercises and blocked environments.

Inspect the source used in this lesson.

What to do next

Next: Turn every failure into a rule: how findings become reviews