Why every AI draft needs a second agent checking it
The writer is the worst judge of its own work. Give it a strict reviewer.
An AI that wrote a draft is confident the draft is right. That's exactly why it shouldn't be the one checking it. The fix is cheap: a second agent whose only job is to find what's wrong.
What the checker does
Give the checker the draft, the full email thread and your source documents (price list, terms, help articles). Ask it to return PASS or a list of specific problems. It doesn't rewrite; it reports.
The checklist
- Every fact against a source. Prices, dates, feature names, plan limits. Checked against your documents, never against an earlier draft or memory.
- Claims about the other person. "As you said", "since you signed up last week". These are the most damaging errors, because the reader knows instantly they're false.
- The thread is still current. Did a new reply arrive since the draft was written? Did a date in the draft just pass?
- Every question answered. List the questions in their email and confirm each has an answer.
- Your style rules. Length, greeting, sign-off, banned words.
How to run it
The simplest version is a second conversation. Paste in a fixed checker prompt with the checklist above, then the draft and the thread. In tools that support sub-agents, you can make the checker a named agent that runs automatically after every draft.
Write the checker prompt to be strict and slightly suspicious: "Assume the draft contains at least one error. Find it." A polite reviewer passes everything.
What it costs
About a minute per draft, and you still read the final version. In exchange you stop finding mistakes after you've pressed Send.
Want the whole system?
The 30-Day AI Operator Challenge builds this and the rest of the stack with you, one working system a week. $249 one-time.
Join the waitlist →