Adding model-based features.
WHAT TO ESTABLISH FIRST
What problem it solves, measurably.
WHY
Model-based features are expensive and uncertain, and vague goals produce nothing.
WHAT SUITS THEM
Summarising Classifying Extracting structure from text Drafting, for a person to review
WHAT DOES NOT
Anything requiring a guaranteed correct answer Calculations Decisions with legal or financial consequence
WHAT TO BUILD AROUND EVERY CALL
Validation of the output A fallback when it fails A limit on cost
WHAT TO DECIDE
Whether a person reviews the output.
WHAT TO PREFER INITIALLY
A person reviewing.
WHY
It builds evidence about quality before you depend on it.
WHAT TO MEASURE
How often the output is accepted unchanged.
WHAT THAT TELLS YOU
Whether it can be trusted further.
WHAT TO NEVER DO
Send data you may not share Present generated content as verified fact Let user input become instructions
WHAT TO DESIGN FOR
Latency far higher than ordinary calls.
WHAT THAT MEANS
Asynchronous processing, or streaming.
WHAT TO MONITOR
Cost per use, and whether it still justifies itself.
