MCP Prompts and Resources: The Primitives You're Not Using

Your user asks for a weekly sales report. The LLM has four tools available: querying the database, aggregating data, calculating trends, and formatting the output. It chains them together. Steps 1 and 2 go fine. Step 3 goes wrong: the LLM tries to calculate week-over-week percentage changes itself, mixes up which week is the baseline, and produces a report showing 340% growth in a category that actually declined. The user gets a polished, confident, completely wrong report. This isn't a contrived scenario. It's the predictable outcome of asking an LLM to choreograph a multi-step workflow where some steps require symbolic computation. The LLM is good at language. It is bad at arithmetic. And when you give it tools for each individual step, you're asking it to be good at something else entirely: sequencing, data flow management, and knowing which steps it should delegate versus attempt itself. Now consider the alternative. The same user clicks a single prompt: "Weekly Sales Report." The