Ashish Chandra7 minAI EngineeringNewsletter: Prompting Patterns That Actually Scale
This month's issue covers structured prompting, evaluation habits, and why your prompt library needs version control like any other codebase.
This month's issue covers structured prompting, evaluation habits, and why your prompt library needs version control like any other codebase.
Prompting As Engineering
Prompting stops being a trick and starts being engineering the moment you version it, test it, and measure it. Until then you are tuning by vibes and hoping the improvement was real.
- Store prompts in the repository, not in a document
- Version them alongside the code that calls them
- Keep a fixed evaluation set you re-run on every change
- Record why a prompt changed, not just what changed
Key Takeaway:
If you cannot tell whether last week's prompt change helped, you do not have a prompting problem — you have a measurement problem.
Patterns That Survive
Explicit role and task framing, structured output contracts, and few-shot examples drawn from real failures have all outlasted several model generations. Clever phrasing tricks generally have not.
“Treat prompts like code, because that is what they have become.”
