Essay says writing beats code as the source for AI builds

A developer tells an AI assistant to build a web app in one line. The assistant builds something. Nobody is sure whether it is what the person wanted. The essay argues that this gap is not an AI problem. It is a problem of what you wrote down. Daniel Miessler published it on 23 October. The Hacker News thread reached 58 points over the weekend.

The piece opens on an argument with Matt Pocock. Miessler notes that Pocock's public skills repo on GitHub had passed 301,000 stars by 23 October. Pocock holds that code is the main thing in a build. Miessler disagrees. His position is that writing is closer to the idea in your head than code is.

The three artefacts Miessler names

Miessler draws a chain from thought, to writing, to code. He calls writing the first copy of thought. He calls code the second copy. Each step loses detail. Writing is more accurate, he says. Code is more precise. For an AI build, he puts accuracy first.

He names a document to hold the writing. He calls it the Ideal State Artifact, or ISA. It has four sections. The Ideal State is a paragraph that says what the thing is. Constraints list what you will not build. Decisions log choices as they are made. Ideal State Criteria, or ISCs, are the atomic rules the build is tested against.

Each ISC carries its own test. Some are a simple assert. Some use a small model as a quick check. Some use a stronger model as a judge. The ISCs are both the thing you build from and the test harness.

Why he thinks this holds as AI gets better

Miessler points at Richard Sutton's Bitter Lesson. The point there is that general methods built on more compute keep beating methods built on human cleverness. He writes that the smarter the AI gets, the more work should go into saying what you want. Less work should go into the exact shape of the code.

The practical turn: the code becomes disposable. If the first build is bad, you throw it out and rebuild against the same ISA. The same ISCs test the new build. You cannot do that when the code is the only record of what you asked for.

What a team can try this week

The essay carries three worked examples. Someone wants to draw an owl. A yoga studio wants a booking app. A couple remodels a kitchen. In each one, the same mistake repeats. The implementation becomes the record, and the real idea is lost.

An AI engineering team can take two things from the piece. First, a written spec with its own tests sits above the code. Second, the writing itself has to be iterated. The ISA changes as rereading it changes your mind. The code changes as the drafts show you what you forgot to say.

Source

Writing Is Closer to Thought Than Code, Daniel Miessler.