A colleague compared the same model across two environments and came away with one conclusion: he needed to load more context. That is prompt-level thinking applied to a problem that is not about prompts. Loading more context into a session is not the same as building a layer that keeps context persistent and domain-specific. One is a habit. The other is a system you build.

The model is not the tool. The tool is everything around it. Specialized agents carry domain knowledge into every session so they do not start cold. A context manager pulls prior decisions and facts back automatically. A skill system routes recurring work without re-explaining it each time. A knowledge base feeds into every session. Claude runs on top of all of it. The model did not build any of it.

The model is a runtime. It executes what you give it. Give it nothing, and it produces output indistinguishable from what anyone else gets. The specificity lives in the layer you built, not in the model.

Agents have to be designed around how your work breaks down. The knowledge base has to be fed and kept current. The context manager has to know what to surface and what to drop. None of it is automatic. None of it comes preconfigured. You build it the same way you build any infrastructure, over time, with maintenance.

Until you have done that, you are running prompts against a runtime and waiting for the general training to be good enough. The model will be replaced. The scaffolding you build will not.