When people talk about prompting, they usually mean the words they type into the box. But the most powerful prompt is everything the model can already see: the file structure, the existing components, the names chosen for routes and variables, the comments left by earlier decisions.
A real codebase answers most of the questions a prompt would otherwise have to ask. It removes ambiguity about conventions, scope, and direction before the first word is typed.
Why this works
- Naming already encodes intent
- Existing files narrow the search space
- Real routes show what matters
- Less instruction, more continuation
Build the codebase like a prompt, and most edits become small instructions on top of an answer the model already has.
