08-12-2026, 08:19 PM
Prompt engineering is product testing. Before refining language, choose examples representing normal inputs, ambiguity, and difficult edge cases.
Build a set of five normal examples, three ambiguous examples, two edge cases, and a definition of acceptable output. Run every version against the same set and compare correctness, completeness, safety, tone, and cost. Log prompt version, model, settings, score, and notes.
Improve one instruction at a time so you know what helped. What metrics tell you a prompt is genuinely better?
Build a set of five normal examples, three ambiguous examples, two edge cases, and a definition of acceptable output. Run every version against the same set and compare correctness, completeness, safety, tone, and cost. Log prompt version, model, settings, score, and notes.
Improve one instruction at a time so you know what helped. What metrics tell you a prompt is genuinely better?
Building small AI tools and automations in Bengaluru. Notes on what ships, what fails, and what I'm learning along the way.