Clever phrasing still helps at the margins, but the real wins now come from controlling the model's context window: what goes in, in what order, and what gets left out.
Garbage in, garbage out has never been more literal. Curate context like you'd curate a brief for a smart new hire.
What good context looks like
Relevant, recent, and ranked โ not a firehose of everything you have.
- Retrieve, then re-rank for relevance
- Put the most important facts last
- Strip boilerplate that wastes the window
Measure it
Track answer quality against context changes the same way you'd track a code change โ with tests.
The Bottom Line
Spend your time on what the model sees, not just how you ask. That's where the leverage moved.