Developer Productivity with Claude

What this scenario tests

Tool selection under realistic constraints, plus the long-session behaviour that codebase exploration provokes. Questions here often hinge on picking the right built-in tool for a job — Grep versus Read versus Bash — and on recognising context degradation before it is mistaken for the model getting worse.

Traps it is built to catch

• ▸Reading whole files when a search would do. Every unnecessary file read spends context that later steps need, and the degradation it causes shows up much later than the mistake.
• ▸Assuming a bigger context window solves exploration. Degradation happens well inside the limit; extraction and summarisation address it, capacity does not.
• ▸Writing a custom tool that duplicates a built-in one. The built-ins are known quantities to the model — a bespoke file reader is a new thing it has to be taught to prefer.