Prompt Engineering — the practical framework

A prompt is the instruction you give to an LLM. Two kinds of prompt: the system prompt (the fixed instructions: 'you are an assistant specialized in...'), and the user prompt (what the user types). Both matter — the system prompt shapes behavior, the user prompt carries the specific request.

A reliable prompt has 6 ingredients: role (who is the AI?), task (what is it supposed to do?), context (what info does it need?), constraints (what must / must not it do?), format (what does the output look like?), examples (1-3 illustrative examples).

Concrete example. Bad prompt: 'summarize this article'. Result: arbitrary length, unclear style. Good prompt: 'You are an analyst preparing a daily briefing. Summarize the article below in 3 bullet points, each of max 20 words, focused on the business impact. If a number is mentioned, keep it. Article: [...]'. Much more predictable result.

Iterate like an engineer: test on 5-10 varied inputs, note failures, refine the prompt, retest. Treat it like TDD — you're building a small regression suite. When something works, save the prompt with the version in git.

Common pitfalls: (i) contradictory instructions ('be concise but exhaustive') — the model picks one or averages; (ii) no examples on complex tasks — hard to infer the expected format; (iii) prompts inflated beyond 2000 tokens — you pay for useless context; (iv) ignored edge cases — the prompt works on happy-path but fails when input is empty, too long, or in an unexpected language.