Early AI work focused on prompt wording. In production systems, the larger lever is what the model is shown: the right documents, tool descriptions, prior steps and rules, organized so the task is clear.
What the work involves now
- Selecting and ordering the information the model sees for each step.
- Keeping context lean: more text is not always better.
- Defining tools and outputs precisely.
- Testing changes against a fixed evaluation set.
Treat context as a designed component with owners and tests, not as a block of text someone edits when something breaks.