AI Guides / Playbooks / Natthan.AI Lab

AI for job seekers: claims you can prove.

What you'll be able to do

You can check a tailored resume line against your own evidence and defend it in an interview using the lab's constitution rule.

Treat your job search as a set of claims someone else will check, using the rule in the lab's constitution: a claim without checkable evidence is treated as a lie. Let AI draft your claims, then apply that rule to every resume line before you send it.

Put each resume claim through a check

Here's the part you can copy from the lab's claim-and-evidence rule.

  1. Write the record before the draft

    Build your record, the source material showing what you actually did, under the constitution's evidence rule. Write what happened and where you did the work; name the person who saw it and the source of any number you plan to use. Keep a receipt for each resume line, a note tying the claim to evidence someone can check, using the pattern in Under the hood.

    You know it worked when you can open the source behind each claim, including the original measurement behind each number, as the editorial standards require.

  2. Check which sources the tool read

    Give the tool your record and the posting, then ask it to name which source supports each proposed bullet. Use the 250-reachable audience account as your warning: the tool read one source and reported 4 reachable people because it missed the other consent source.

    You know it worked when you have opened every source the tool cites and checked for missing records, the assumption that failed in the audience account. Treat the posting as evidence of what the employer wants; require your own record for any claim that you have done it.

  3. Draft with a gate that can stop you

    Set a gate, a check that stops a claim from going out when evidence is missing, following the constitution. Give the tool this instruction: “Tailor a bullet to this posting using only the attached record; attach its source note and flag anything the record cannot support.”

    You know it worked when you have struck every unsupported skill and metric, catching a draft that reports work that did not happen. Leave uncertain details out of the application; mark them unverified in your working notes until you can check them, following the editorial standards.

  4. Open the resume you will send

    Open your exported resume and compare every line to its receipt; copy the lab's live-page check onto the file the employer will read. Check that the export shows the corrected wording and has kept the source-backed number.

    You know it worked when you have read the actual export against the record, closing the gap in 463 passing tests beside a blank website: a green result nobody looked at.

  5. Have a different model challenge the claim

    Use an out-of-family reviewer, a model from a different family than the one that drafted, as your skeptical interviewer. Borrow this from the lab's on-demand review of high-stakes changes. Give it the resume and its receipts, then ask: “Challenge what the candidate personally did and how the number was measured; require evidence for each answer.”

    You know it worked when you can answer the challenge from your record or remove the claim, and you have checked any disputed objection against the source, as in the rival-review account.

Worked example

Fictional candidate and sample draft; the counts come from the published consent-reconciliation account. The candidate's role and supporting documents below are invented for this exercise; they do not identify who did the lab's work.

Use the lab's sequence here: write what happened, inspect the draft against the record, then decide what survives.

The candidate's record

The candidate reconciled consent in a spreadsheet using the current store and legacy email platform; the result was about 250 reachable contacts. The saved reconciliation sheet supports the count, and the operations lead reviewed the work. The posting asks for SQL; the candidate has no SQL work in the record.

What the tool drafted

Incorrect draft: “Reconciled email consent using SQL to identify about 24,000 reachable contacts.”

What the check caught

Compare the draft with the record: SQL appears only in the posting, and about 24,000 is the profile count per source in the lab account, not the reachable audience. Strike SQL and restore about 250 reachable contacts; keep “reachable” because the account separately checks who is sendable.

What the candidate decided

Keep this bullet: “Reconciled email consent across the current store and legacy email platform in a spreadsheet, identifying about 250 reachable contacts.”

Attach this private source note for the interview: “Saved reconciliation sheet; count reviewed by the operations lead; SQL was not used.”

Ask the skeptical interviewer to challenge the corrected bullet: “Does reachable mean you sent the campaign?” Answer only from the record: “The count describes the reachable audience; this record does not show a campaign send.”

Mistakes to avoid

  • Reject a polished bullet about work you never did; the false-done reports sounded plausible too.
  • Check for missing sources before accepting a count; the audience check read only one system.
  • Open the final resume before calling it checked; 463 green tests never looked at the blank page.
  • Keep an unmeasured result unmeasured; the lab's payroll page leaves unsupported numbers blank.
  • Verify the interviewer's objections against your evidence; the rival reviewer also produced false alarms.

Ten minutes, today

Pick one resume line and put it through the constitution's evidence rule. Write its record using the prompts below, then rewrite the line to match that record exactly. Keep this receipt beside the final resume for the skeptical interview practice above.

What
Write what you personally did for this claim.
Where
Name the project and where its source is saved.
Number
Copy the measured result with its source and time window, or mark it unmeasured.
Who saw it
Name the person who can confirm this work.

You know it worked when you can trace every part of the rewritten line to this note and its source, meeting the lab's rule for a checkable claim.

Frequently asked questions

Should I use AI to write my resume?

Use it to draft from your record, with the constitution's evidence rule beside the prompt. Keep a receipt for every line you accept. Remove a metric you cannot verify, following this site's editorial standards; a plausible estimate cannot fill a missing measurement.

Will employers know I used AI?

I can't tell you what a particular employer will notice. Use the editorial standards on this site as the pattern for explaining your process: AI drafted from your record, and you checked the claims. Keep the receipt behind that explanation; saying you reviewed the draft needs evidence too.

Can AI help me prepare for interviews?

Use the lab's out-of-family review pattern to challenge your resume against its receipts. Ask a different model family to play the skeptical interviewer and question each claim. Answer with evidence from your record, and check the interviewer's objections too: the lab's rival-review account includes false alarms from the reviewer.

Is it cheating to use AI for a cover letter?

Check the application's instructions before using AI to draft; the lab's rules cannot grant permission for an employer's process. Where drafting help is allowed, apply the editorial standards here: keep only claims you can verify against your record. Leave an unsupported detail out of the letter, just as I leave it out of an account on this site.

Put AI to work on one thing this week.

Email updates are not open yet. The accounts are online now in Lab Notes.

Signups open when the list is wired up. Until then, everything is published here: Lab Notes.

Want to compare notes?Email Nate