Tell the Model How to Work, Not What to Do

I’m not sure who needs to hear this, but I just don’t use skills at all. Zero skills in my repos.

The reasoning models inside coding agents are fundamentally trained to follow instructions. Skills are a way to supply instructions, but they’re privileged in the post-training sense: files on disk wired in as skills get very high instruction-following fidelity via rewards during RL. That can straightjacket the agent, especially when skills are bound to specific tasks and workflows.

The magic is their reasoning capability. The instructions I want to give agents are mostly theory-of-mind stuff: how to think about the world they’re spawned into, how to ground themselves, and how to find their own instructions to follow.

Broad knowledge of software architecture does not make a model locally correct. Pretraining gives it many possible ways to solve a problem, and post-training shapes which of those possibilities it reaches for by default. The model still has no way to know which choices this repo accepts as good until that context enters its attention. The harness supplies that context so I can get out of the way.

Daniel asks how I give agents long-term preferences around what good, done, and high quality look like. I put those preferences in files in the agent’s environment. The agent pulls them into context when they are relevant.

Those files explain how to think about the work, what good looks like, and what done means. This is why I keep pointing to AGENTS.md as a map of the environment. The model decides how to execute. I tell the model how to work, not what to do.

As Miguel Branco put it, “Teach them what ‘good’ looks like and let them work.” Typical skills prescribe specific workflows, which can make agents mechanically follow every step even when a step is unnecessary or the situation calls for an exception. Those instructions consume scarce context and bias downstream reasoning before the agent has considered the situation in front of it. No amount of making the model better will obsolete the need for in-context learning.

I use zero skills because the agents are already generally capable software engineers that can learn the job. I give them the local context in ordinary files and let them work.