Context
The model sees what we curate.
An LLM is a context-bound function. Same prompt, different context, different output, by orders of magnitude. Senior engineers treat the context window as a designed surface, not a chat scratchpad.
The intern's loop
A junior pastes one function and asks for a fix. The model sees no consumer, no test, no surrounding convention. It returns code that's locally correct and globally wrong: the API drifts, the type doesn't match, the naming clashes with the rest of the file. A 'simple change' breaks four things downstream because the model never knew they existed.
An engineered context window
Every meaningful task gets its context engineered: the file under change, the consumers, the type definitions, the existing tests, the convention doc, the architectural decisions that constrain the answer. Context-loading tooling (CLAUDE.md files, indexed code, scoped agents) keeps the model's view of the codebase richer than a single function.
A repo the model can actually see
A CLAUDE.md, a project conventions file, an explicit context discipline. Subagents scoped to specific concerns, not the entire repo. Every future change (by anyone, with any model) runs faster because the context is engineered, not improvised.
- 01Does the model see more than the file under your cursor when you ask for a change?
- 02Is there a CLAUDE.md (or equivalent convention doc) at the root of every active repo?
- 03When the model returns code, does it match the conventions of the rest of the codebase?
- 04Do you know what's in the model's context window before you hit enter?