A beginner-friendly walkthrough for setting up Claude Cowork, connect your files, tools, and preferences so Claude can do real work alongside you.
Context
What to feed the AI so it stays productive — codebase grounding, project conventions (AGENTS.md), persistent decisions, context-window limits, session resets.
You asked for the same feature twice and got two different implementations. The problem is not the model. It is that your decisions only live in chat history.
MCPs you wire into your main session do not cross into spawned sub-agents. The agent reports Linear unavailable and invents its scope from text. Here is what actually propagates and what to do about it.
You described your project in English. The agent produced generic code that matches none of it. Stop describing. Mention the file. Run the command. Watch the output change.
Three things break an agent session: a wrong turn, a drifted conversation, a thread that grew too big. Three different moves fix each one. Picking the wrong move wastes time.
You typed the rule in chat three times and the agent still broke it. Chat is for the current task. Project rules live in AGENTS.md. Yours is empty.
The model did not get worse. Your context filled up. Every token in the window influences every output, and the noise is now winning.