A theory I won't shut up about
The 90/90/90 theory
Every AI handoff is faster and a hair worse. Stack a few and "a hair worse" turns into "oof, what happened." The fix is almost insultingly simple: try. Like, try-trying. Here's the chain, the math, and why that beats everybody.
The chain
Some business person wants a new feature. She pops open Claude: hey, what should I build for this user? Out comes a brief. It's maybe 90% of what a sharp human would've written, and it took five minutes instead of a week. Honestly? Great. But it's 90%.
A product person grabs that brief and asks Claude to turn it into a spec. Same move, same 90%, done before lunch. Except now it's 90% of a thing that was already 90%. You see where this is going.
Then a dev takes the spec and asks for the code. 90% again. Then someone writes tests off the code, docs off the tests, release notes off the docs. Every step fast. Every step a perfectly respectable 90%.
Nobody screwed up. Everybody shipped by Tuesday. And the thing that falls out the far end is a sad little ghost of what you wanted.
The math
0.9 isn't the villain here. Compounding is. Ninety percent of ninety percent of ninety percent of a finished thing is way less than ninety percent.
0.9 × 0.9 × 0.9 × 0.9 = 0.656
Four handoffs and a third of it's just gone. Nobody was sloppy. It's math. 90% times 90% times 90% times 90% is 66%, and numbers don't care about your feelings.
66%
of the original still standing at the end of the chain
The flip side
Good news is hiding in the exact same math. Bump each handoff from 90% to 95% and the floor stops falling out from under you. Get each one to 99% and you basically hold the line the whole way down.
The gap between a 90% team and a 99% team isn't nine points. Over four handoffs it's 66% versus 96%. Same steps, same speed, totally different thing at the end. That gap? That's the whole game.
Same four handoffs, three different teams
Move the sliders above and this moves with them.
Okay, so what do you actually do
When anybody can ship 90% in five minutes, 90% isn't the edge. It's the floor. Table stakes. The way to stand out is boring and kind of a pain: get closer to 100 at every single step.
Read the output. Fix the 10% the model whiffed. Hand the next person something that's actually right, not just something that's fast. Do that at every link and you stop fighting the math. You start compounding the other way.
And here's the good part. The same tools that make 90% cheap make checking the last 10% cheap too, if you bother to look. Most folks won't. That's not a bummer. That's the opening.
Full disclosure: this post is probably only 90% itself. I did try to drag it to 92. You be the judge.