Paul TakisakiTHE AI FIELD GUIDE / 01

FIELD NOTE 06 / EVIDENCE OVER REASSURANCE

Make “done” come with a receipt.

Tell the difference between a confident answer and a checked result.

617prompts asking for verification or evidenceSource & definition ↗

Decide what would count as proof before the work starts. Then ask for that artifact at the end: a test result, a source row, a working screen, or a completed dry run.

“Done means the command was run and the output is in front of me.”

Paul’s working notes · September 2026

What it looks like

THE FRICTION

“Everything looks good. The totals are correct.”

A BETTER STARTING POINT

Source: the supplied export. Method: count unique session IDs, not message rows. Result: show the total and a small sample. Check: compare with a second calculation. Limit: identify missing dates.

Illustrative examples you can adapt.

Put it to work

  1. 01

    Choose the check that fits the claim.

    A code change needs the relevant tests and often a browser check. A research claim needs a source you can inspect. A count needs the unit, date range, and deduplication rule. A screenshot of a dashboard does not establish how the number was calculated.

  2. 02

    Ask for the actual result.

    Get the command, result, and the important output. If the output is long, keep the raw artifact linked and the summary short. “I would test this by…” describes a plan; it does not mean the test ran.

  3. 03

    Use a different check where the risk is high.

    Recompute an important total from the original rows or ask a fresh reviewer to inspect the method. In my working notes, a repeated log line inflated a number by 2.3 times on the first pass. The counting unit mattered more than the polish of the page.

Give this to your AI

Adapt the brackets to your task.
Before you call this complete, show the evidence for each important claim. For code, report the exact checks run and their actual results. For numbers, show the source, unit of counting, date range, filters, and duplicate handling. Link the raw output when it is long. Mark anything untested or inferred. Do not rerun expensive or destructive operations just to create a receipt; ask how to verify them safely.
WHAT TO WATCH FOR

A green build does not prove that a button works, and a citation does not prove that a source supports the sentence. Read enough of the evidence to check the claim being made.

THE HABIT TO KEEP

Define the proof first. Ask for it at the end.

Sources & further reading

Personal counts are from Paul’s prompt-history analysis. Tool guidance was checked against these sources in September 2026. Confirm current behavior in your installed version.

A USEFUL NEXT READ

Automate the check before the job.