Prompts are specs. What works and why evaluations are part of it.
Category · AI & Agents
Prompts are specs.
Prompt engineering isn’t "asking nicely", it’s the precise description of a task for a language model: role, context, constraints, desired format, examples. If you want to build an LLM-powered process, you’re writing a specification — the prompt is that spec in prose form.
What works.
Chain-of-thought (guide step-by-step reasoning), few-shot (provide examples), clear system prompts, structured output in JSON or XML. And: automated evaluations that make prompt changes measurable — otherwise prompt maintenance turns into gut feel.
See also
