Field Guide · Start here
Verify
Confidence is a style. Evidence is a process. Build a verification habit.
Why verification is the craft
Models can sound certain while being wrong. Your job is not to distrust everything. It is to match scrutiny to risk.
Fluency is a product feature. Evidence is your process. For why confident wrong answers feel so smooth, see the Challenge Why your AI sounds confident when it’s wrong.
Match scrutiny to risk
| Risk if wrong | What to do |
|---|---|
| Low (brainstorm title) | Skim and ship |
| Medium (internal howto) | Spot-check facts; run it once |
| High (customer, legal, money, health) | Independent source + human owner |
If you cannot place the task in a row, treat it as high until proven otherwise.
Five-check verification loop
Run these before you paste, send, merge, or publish:
- Assumptions - Ask: “List assumptions and the weakest claims.”
- Skeptic pass - Second prompt: “What would a skeptic challenge? What would falsify this?”
- Source you can open - Prefer a link, ticket, log, or doc you actually open over a citation that merely appears.
- Numbers and names - Re-check quantities, dates, people, and product names against a system of record.
- Paste privilege - If the prompt or the outbound message contains secrets or customer data, stop. Use the habit in Verify before you paste.
Low-risk work can skip to a skim. High-risk work needs all five, plus a human owner.
Habits that scale
- Ask the model to list assumptions and weakest claims.
- Re-run critical steps with a skeptic prompt.
- Prefer citations you can open over citations that merely appear.
- Save one redacted “good verify” example so the habit sticks on a busy Tuesday.
- When privacy is the risk, read Privacy and judgment and the Challenge Privacy is a habit.
Common failure modes
- Shipping because the prose sounded finished
- Checking only spelling while numbers stay unchecked
- Asking the same chat to prove itself without an outside source
- Skipping the paste check when you are late for a meeting
If you recognize one of these, slow down on the next send. Craft is mostly refusal of convenient shortcuts.
What good looks like
You can explain, in one sentence, what you checked and what remains uncertain. You are not required to trust nothing. You are required to own the residual risk.
See also
- Pulse: Verify before you paste
- Challenge: Why your AI sounds confident when it’s wrong
- Checklist: AI craft principles
- Next path: Build a workflow when the loop starts repeating