Prompting techniques that hold up across models and over time — not the latest TikTok hack. Role-context-task-format scaffolds, few-shot examples, chain-of-thought, and the eval discipline that lets you iterate.
Yes. Models have improved, but the gap between a beginner prompt and an expert prompt is still 2–5x in output quality on subjective tasks. The skill has shifted from clever tricks to structured prompting plus eval discipline.
For multi-step reasoning, yes — 15–40% accuracy gains depending on task. For one-shot factual lookups, no — it just adds latency. Add "think step by step before answering" only when the task genuinely has steps.
Specify the format with an example, not a description. Show one full input → output pair in the prompt, then ask the model to follow the same shape. Add a JSON schema or function-call definition when reliability really matters.
Long enough to set role, output format, and 2–5 few-shot examples. Past ~1,000 tokens, you start diluting the signal — models average over everything in the prompt, and important rules drown in the noise.
Build a small eval set: 20–50 real inputs with the desired outputs. Run the prompt against the set, score by accuracy and format compliance. Any prompt change that does not improve the eval score is not really an improvement.
