Post
Micromanaging the Model
A skill is micromanagement written down.
Micromanagement isn’t always wrong. A new hire who keeps missing steps does better with a checklist. Today’s models still drift, skip conventions, and lose the thread, and a well-written procedure can get them to a better result. Within its context, a skill works.
But micromanagement fails in a predictable way. Its value falls as the competence of the person being managed rises. Every instruction that helped a weaker report constrains a stronger one. The model is the one employee guaranteed to get better every few months, and at exactly the work the instructions describe. So a skill grows more over-instructive with every release. You end up with a worse result than the model would have reached on its own. Because the steps were yours, you’re also left justifying how you got there.
Ryan Lopopolo draws the distinction cleanly: “To build via ticket is to care about mechanism. To build via big objective is to only care about outcomes. Successful, high ambition agentic systems only build for and evaluate for outcomes.” A skill is mechanism. Mechanism is what the model is improving at fastest.
Yet the mechanism keeps getting promoted. A procedure that makes up for a missing capability gets described as a foundational structure of AI development. People publish it, teach it, sell it, and give talks about it. The workaround becomes the product.
Nobody has to be acting in bad faith for this to go wrong. It’s enough to have picked your sacred cows. Once a method becomes your business or your identity, each model release becomes a threat to defend against instead of evidence to update on. The defense isn’t malicious. It’s structural. Whatever its intent, it misleads people who aren’t positioned to see through it.
I keep running into those people. Companies and individuals, many of them non-technical, are convinced they’ve built a skill that does something remarkable. Often it does, right now, in that harness, against that model. What they don’t see is that all three are moving. The model improves, its ecosystem absorbs common procedures, and the harness changes what it injects and when. Ryan again: “for EVERY new model that comes out, you have to reset your priors and re-discover what is possible.” A skill is a prior that refuses to reset.
The model eats everything.
The strongest objection lands close to home. I maintain kit, a set of rules and procedures for building with agents, and that’s the kind of thing I’m describing. The difference is direction of travel. Kit began as a heavy, spec-driven harness. It’s been cut down to something lean and is getting leaner, so it stays out of the model’s path. What remains covers what a model shouldn’t be expected to infer: the human side of shipping, like approvals, change control, and evidence that the work is done, plus my own preferences. Those preferences get reconciled, aggressively, against each new model. Even procedures that belong have to keep earning their place.
That’s the test I’d apply to any skill. Does each release take something away from it, or add to it? A method that only grows is being defended, not tested, and someone downstream inherits the cost.
State the outcome. Own the rules only you can own. Stop managing the work.