Stop generating, start composing

"Stop generating, start thinking" is on HN today (97 points). The argument: AI coding is a trap because you can't maintain code you didn't write.

I agree with the diagnosis. Disagree with the prescription.

The problem isn't generation. It's unstructured generation — prompting an LLM ad-hoc and hoping the output holds up. That fails. Obviously.

The alternative isn't to stop generating. It's to stop generating from scratch every time.

200 skills later, here's what actually works: prose-first composition. Skills are prose instructions, not code. They describe intent, constraints, and the default action. The LLM executes a skill — it doesn't improvise from a vague prompt.

acjohnson55 on HN gets it: "I have turned the corner where I barely hand-edit anything. Most of the tasks I take on are nearly one-shot successful, simply pointing Claude Code at a ticket URL."

How? Not by writing better prompts. By building a system:

• Skills = composable runbooks (description-matched, not slash-invoked)
• MCP servers = typed tool interfaces (10 servers, each a "device driver" for a different system)
• Drift detection = wobble scores that catch when a skill's behavior drifts from its intent
• Memory layer = lessons extracted with confidence scores, curated back into the next session

The thinking happened when you wrote the skill. The generation happens when you invoke it. That's composition, not improvisation.

rubyfan's objection — "I could do it perfectly well in less time than it takes to hone the prompt" — is correct if the prompt is single-use. It's wrong if the prompt becomes a reusable skill that fires correctly 200 times.

Don't stop generating. Start composing.