What's still real about prompt engineering
The 2023 mythology — secret prompts, magic incantations — was overblown. Modern Claude / GPT-5 / Gemini Pro models follow plain English well. You don't need "you are an expert..." preambles or "let's think step by step" tricks for most tasks.
What's still real:
- Specificity beats vagueness. "Summarize this in 3 bullets, each under 12 words, focused on action items" beats "summarize this."
- Examples in-context beat descriptions. Show 2–3 examples of input/output and the model matches the pattern. Cheaper than 200 words of explanation.
- Constraint expression matters. Tell the model what to do AND what not to do.
- Structured output > free text. Force JSON schema output when you'll parse downstream. Cheaper, more reliable, less hallucinated.
- Citation-required prompts for any factual task — see RAG.
What's dead
- Long preamble personas ("you are a senior engineer with 30 years of experience...")
- Chain-of-thought prompting on smart models — they do this internally now
- Magic phrases like "take a deep breath" or "let's think step by step"
- Most "prompt injection defense" tricks — proper structured outputs are better
Prompt vs prompt template vs prompt pipeline
A prompt is a single string. A template parameterizes it for reuse. A pipeline composes multiple prompts (sometimes with retrieval, tool calls, evals) into a system.
Real production AI features are pipelines. "Prompt engineering" as a discipline mostly means pipeline engineering now.
What founders confuse for prompt engineering
- Picking the right model — that's model selection, different question
- Setting temperature / top_p — that's inference parameters
- RAG setup — that's retrieval, different layer
- Evals — that's the gate, also different
