How do I get more real work out of Claude with MCP?
I have Claude set up with a couple of MCP servers but I still end up doing most of the work myself. What setup actually makes it productive?
Productivity comes from Claude seeing your real material without you pasting it. Connect it to the place your work lives, usually a project folder and its git history.
Then give it a way of working, not just access. Skills are plain files that tell Claude how to do a recurring job: how to plan, how to review, how to test. They load only when the job comes up.
A good first pair is a planning skill and a review skill. Claude asks the questions it needs before it starts, and checks its own work before it says done.
Keep each skill readable. If you cannot explain what a skill does after reading it once, it will not behave predictably either.
Listings mentioned
- Brainstorming Ideas Into Designs · skill by obraMakes Claude ask questions and shape the idea before writing code.
- Code Review Reception · skill by obraA method for handling review feedback instead of agreeing with everything.
- mcp-server-git: A git MCP server · agent by modelcontextprotocolLets Claude read the history of the project it is working in.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.