The Problem
I’d spent the morning rebuilding the editing system I use to draft my own posts. I kept feeding it something rough and reading back something clean, correct, and a little dead — every edge I’d have kept, filed smooth. Once that was fixed, the obvious next question was whether any of it mattered. So I set out to audit a year of my own posts across two platforms: what actually lands, what falls flat, and whether the “voice” I’d been fussing over drives anything at all.The Approach
I pointed Claude at it as a small team rather than a single pass. Four agents, each with one job — one mapped the voice and personality across every post, one hunted the AI tells I’d been trying to strip, one took performance, and one played skeptic, whose only assignment was to find where my own rules were wrong. I handed them a shared rubric and told the skeptic to argue against the other three. The bet was that a panel with an adversary built in would catch the flattering conclusions a solo analysis talks itself into.What Happened
The report came back clean and confident, rendered out to a nice page. The headline: voice barely predicts performance; it’s mostly reach and luck. I almost shipped it as the answer. Then I pulled my actual platform analytics — the native export, the one with impressions in it — and the whole thing turned backwards. The agents had been working from scraped data that could see likes but not how many people ever saw a post. My single best post of the year, by a factor of seven, was one the scrape had never even captured. The real story was almost the opposite of the headline: most of my posts barely get shown to anyone. The ceiling was distribution, and the writing rarely got a fair test at all. I had Claude rebuild the report from the real numbers and correct the two wrong claims in the open, right at the top of the page.What I Learned
Scraped engagement lies by omission — without impressions you’re ranking reach and calling it resonance. The correction only happened because I went and got the denominator instead of trusting the tidy first answer, and because I’d built a skeptic into the panel who kept saying the numbers couldn’t carry the weight we were putting on them. Next time I get the native data before the agents write a word.Drafted by my Claude from our working session. I edited and approved it.