Why AI Gets You to 90%

Claude Code is gut-wrenchingly good. If you're not using it (or similar), stop reading these words right now and go try it. Bye.

Still here? Either you're converted or you're the kind of person who doesn't do what they're told. Either way, we can continue.

But how much are the productivity gains? Twitter / LinkedIn / everyone in San Francisco seems to have to have strong opinions that it's a large number followed by an "x".

Is it 1.1x, 2x, 10x or 100x?

Diminishing Returns of Progress

Imagine you're building a Lego tower. You place the second brick and it's satisfying - you've just doubled what you've built. By brick 100, you're adding 1%. Same effort, fraction of the dopamine. Early progress is intoxicating. Late progress is a grind.

What's more - you prioritise the impactful stuff first... Obviously! However, that implies that the more you work, the less important each next task is.

This doesn't mean the final 1% isn't important or even essential. If you're building a rocket, 99% of the engineering might be done. But if the last 1% is "the bit that stops it exploding," don't let Katy Perry on board just yet.

Prompt Inflation

Let's say you're using Claude to write some fantasy novels.

Your first prompt:

Write me a fantasy story. Main character's called Frodo - Small guy, big quest, evil ring. Throw in some walking trees and some really big battles. Three books, 400 pages each.

Congratulations, you've written a trilogy. Well, you haven't - you've described one and the AI has hallucinated something that structurally resembles one. But emotionally? You've written a trilogy. Your productivity was wild - 100x? 1000x? It doesn't have to be exact - just post about it on social media ASAP.

Many prompts later:

In chapter 5, introduce a character called Tom Bombadil. Make him as boring and pointless to the storyline as possible. Really lean into it. Ideally looking to lose about 30% of readers - the ones who were only reading for plot. They weren't your real fans anyway.

Many, many prompts later:

When Sam says "I can't carry the ring, but I can carry you" — can we try "I can't carry it for you, but I can carry you"? Actually that's worse. Go back. But maybe italicise "you"?

This is Prompt Inflation. As your project matures, you spend more prompts for smaller gains.

The 90% Problem

These concepts compound. The longer you work on a thing, the more time you're tweaking, the less value you're delivering.

And it applies at every scale. Netflix in 2016 was scaling to feed the demand of a global audience. In 2026 it's improving the "skip recap" button and keeping the lights on. This is no slight on Netflix. It's just that tech is no longer their edge as each new improvement is marginal. Content is their real edge.

At the other end of the spectrum, a small task adding some basic if/else logic? The AI probably does it first time... But you've still got to test, review, deploy, monitor, update docs and communicate the change with customers.

So What?

How can we prevent ourselves being bogged down by the the final 10%?

Lower your standards

Perhaps 90% is fine. Most things don't need to be perfect - get it out there and see if anyone notices the missing 10%.

The obvious counterargument: If everyone ships at 90%, how do you differentiate? The market for 90% done stuff could be about to get crowded.

Leverage yourself

AI allows you to leverage yourself. Wherever possible, make AI do everything - even if you could do it faster.

This allows you to work in parallel. Kick off a task, let it run and come back to it. Have 10 of them going. If you ever find yourself waiting for an AI to respond you're wasting time.

Working in parallel is going to feel strange. Software Engineering advice up until now has always been about preventing context switching and getting in "flow". Now I'm advocating the opposite.

It's a new uncomfortable way of working... But there's a lot of upside for those willing to embrace it and there'll be no mercy for those left behind.