What they are good at

Reading and summarising, classifying, extracting structured data from messy text (emails, PDFs, invoices), drafting replies, and deciding which tool to call next.

What they are bad at

Exact arithmetic without a tool, knowing your private data unless you give it to them, and being right without verification. Treat the model as a capable junior: give it context and tools, check its work, and never give it more authority than the task needs.

Patterns

1. Single call: prompt in, answer out (summaries, classification).
2. Tool use: the model asks your code to run a function (look up an order) and continues with the result.
3. Agent loop: repeat tool calls until the task is done, within limits.
4. MCP: a standard way to expose tools and data to any compatible AI client.