Teach the Writer, Not the Reviewer

Short Essay · 2026-11-18 · 11:20 AM PT

Claim: In a writer and reviewer agent pipeline, a lesson that lands only in the reviewer's rules never reaches the writer, so every new draft fails it again at review.

I run a script pipeline where one agent writes documentary drafts and a supervisor agent reviews them: a claim check against sources, an editorial read, then an outside model's review pass before I sign off. Every review produced rules, and the supervisor folded each one into its rules file the same day. That looked like a system that learns. But the rules file fed the review read only. The writer's brief still pointed at older sections, so the writer never saw a single new rule, and each draft spent another round failing things we had already named. I noticed because I kept repeating the same notes.

The pre-checks had the same gap. Our mechanical check looked for repeated three-word phrases and too many numbers in one sentence. The last outside review listed 13 issues a script could detect: repeated four-word phrases, British phrasing like "keep trading", a long name said three times, "hundred and fifty" read aloud. The check found 0 of the 13. The writer's own check reported "zero repeats" on a draft with three repeated four-word phrases in it. So we added five rules, and on the same draft it now flags 12 of the 13.

The writer's brief now names the full rules file and says to re-read it before every draft, first drafts included. The writer runs the shared pre-check before each handoff, fixes or justifies every warning, and marks each rule pass or fail with a line of evidence. A small table tracks drafts, review rounds and outside passes per piece, starting from today's baseline, so the claim that first drafts are getting better becomes a number we can check. The learning was real. It just lived on the wrong side of the handoff.

Route every review lesson to the writer the same day: into the rules it re-reads before each draft and, if a script can detect it, into a pre-check the writer runs before it hands off. Then count review rounds per piece, so you can see whether first drafts are improving instead of assuming they are.

Tags: #failure-modes #execution #workflow-ownership