Prompt engineering that works

Most bad AI answers are bad prompts: missing context, no examples, no definition of done. A working prompt has four parts — role and context, the task in one sentence, two or three input-output examples, and hard constraints (format, length, what is forbidden).

Iterate like code: change one thing, rerun the same three test prompts, keep what scores better. Vague prompts («write about testing») lose to specific ones («write a 5-step checklist for API testing beginners, each step one sentence») every single time.

Shapes that work

Three shapes cover most needs. Instruction style («you are a senior QA engineer, review this checklist») sets tone and strictness. Few-shot style shows two solved examples and asks for the third in the same shape — the most reliable way to fix output format. Decomposition style («first list the risks, then write cases for each») forces multi-step work instead of a one-shot guess.

Pick the shape by failure mode: wrong tone means weak instruction, wrong format means missing examples, shallow answers mean no decomposition.

Know the limits

Prompting cannot inject knowledge the model lacks: version numbers, internal APIs and yesterday news will be hallucinated no matter how elegant the wording. For facts, attach sources or use RAG; for repeatable team behavior, write a skill. Prompts steer reasoning, they do not create knowledge.