Verify Before You Send: Fact-Check an AI Draft

Workflow area: Verification · Tier: Foundation · Hands-on time: ~25 minutes

When an AI-drafted email quotes the wrong price to a customer, the customer does not blame the model. Every playbook in this track ends with "verify" — this one is the verify. The core move is separation: you cannot proofread for facts and flow at the same time, so you pull the checkable claims out of the prose into a list, verify the list, then fix the prose. The model can even build the list for you; what it cannot do is be the checker of its own work.

When to use — and when not to

• Use it on anything leaving your handsemails to customers or leadership, reports, slides, posts. External and upward communication gets the full ritual, always.
• Use it hardest on summaries and numberscondensed documents (pb-07), variance commentary, anything where the model computed or compressed. That is where inversions and drift live.
• Use it on high-consequence internal draftsthe memo announcing a date, the minutes recording a decision, the answer that will be forwarded.
• Skip the full ritual for zero-fact draftsa brainstorm or a tone rewrite of your own words has little to check; a read-through suffices.
• Never skip it because the model "sounds sure"confidence is a writing style, not an accuracy signal. That sentence is the whole discipline.

The recipe

1. 1.

   Extract the claims into a checklist. Use a fresh chat — not the one that wrote the draft, which will defend its own output:

   Copy

   Below is a draft I am about to send. Extract every checkable
   claim into a table: Claim | Type (number / date / name / quote /
   commitment / fact) | Where I should verify it.
   
   Include: all numbers and amounts, all dates and deadlines, all
   names and titles, anything presented as a past agreement or
   statement, and any commitment the draft makes on my behalf.
   
   Do NOT judge whether claims are true — just list them.
   
   Draft:
   [PASTE DRAFT]
   

2. 2.

   Verify each claim against its source — yourself. Good output gives you the worklist:

   Copy

   | Claim                                  | Type       | Verify against   |
   | Invoice 2214 total ₹1,80,000           | number     | invoice PDF      |
   | Payment due Aug 12                     | date       | contract terms   |
   | "Meera approved the revision"          | commitment | thread, msg 9    |
   | Delivery moved to 6 weeks              | fact       | supplier email   |
   

   Open the invoice; open message 9. The model listed the claims — only the source verifies them. Anything you cannot trace to a source gets deleted from the draft or rewritten as a question, never sent on vibes.

3. 3.

   Hunt the four classic hallucinations. Beyond the checklist, sweep for the errors AI drafts specifically breed — as bold-lead checks:

   • •
     Invented specifics — a meeting, conversation or "as discussed" that never happened; models add connective tissue that sounds like memory.
   • •
     Inverted conclusions — the source said "not recommended without changes," the draft says "recommended." Compression flips hedges first.
   • •
     Plausible numbers — a total that is nearly right, a date shifted by a day, a percentage from a different row. Close is the dangerous kind of wrong.
   • •
     Overwritten commitments — "we will have this fixed by Friday" when you said you would look into it by Friday. You are now on the hook.
4. 4.

   Recompute anything the model computed. If the draft contains arithmetic the model performed — totals, percentages, variances — redo it in a spreadsheet or calculator. Language models predict likely-looking numbers; a ₹4,80,000 total over figures that sum to ₹4,20,000 will sail through a prose read.

5. 5.

   Fix, then re-read once as the recipient. Correct the draft, then one final human read for the thing no checklist catches: does this say what I actually mean, to this person, in a voice I would sign? Then send. With practice, steps 1–5 genuinely compress to about a minute for a routine email — the ritual gets fast, it never gets optional.

Make it yours

• Two-model cross-check: for high-stakes drafts, have Claude extract claims from a ChatGPT draft, or vice versa. Different models have uncorrelated blind spots — and neither will defend the other's work.
• The reverse-summary trick: for a summary you produced with pb-07, ask a fresh chat to expand your summary back into its implied claims, then compare against the original document — inversions surface immediately.
• M365 Copilot / enterprise settings: drafts built from tenant context still hallucinate; citations tell you where Copilot looked, not that it read correctly. Click through to the cited email before repeating its contents.
• Team scale: turn step 3's four hunts into a shared one-page checklist pinned in your team channel. "Ran the pre-send check" should become a normal thing to say before anything customer-facing ships.

Quality check

• Every claim traced or removednothing survives on plausibility alone.
• Arithmetic recomputed outside the modelspreadsheet or calculator, not a second opinion from the same chat.
• Commitments are ones you choseevery promise in the draft is one you would make unprompted.
• The extraction ran in a fresh chatthe drafting session never grades its own homework.
• A human read the final version aloudthe last quality gate is still you.