Four Thinking Models Top Prompt Engineers Use

In AI creation, prompts act like a compass that tells the model the desired content type, style, and details, directly shaping the quality and direction of the output.

1. Role‑play – A generic prompt such as "You are a nutritionist, design a weight‑loss plan" yields vague results, while a detailed role description (e.g., "You are a clinical nutritionist with 10 years experience, served 300+ clients, prefers non‑starvation methods, hates extreme fasting. Now a 32‑year‑old mother, 160 cm, 65 kg, breastfeeding, limited budget, needs a plan") produces output that closely matches the imagined expert.

2. Chain‑of‑thought – Break a complex task into granular steps. For planning a 5‑day Sanya trip, the author suggests:

Ask for three key pieces of information (budget, companions, preferences).

Decide whether the itinerary should be relaxed or activity‑heavy.

Allocate one main attraction plus one backup per day.

Provide a pitfall checklist (e.g., avoid popular shops, certain time slots).

The chain should stay within 7–8 steps; longer chains cause the model to lose focus.

3. Few‑shot (少样本) – Provide a few labeled examples that include explicit boundary explanations. Using an intent‑classification task with categories [consultation / complaint / refund / chit‑chat], the author shows three user inputs and the model's outputs, noting that bracketed comments in the second example explicitly mark the decision boundary, which is the most valuable aspect of few‑shot prompting.

4. Self‑critique – First let the model generate content, then ask it to review from a critical perspective. Example: ask the model to write a birthday wish for a mother, then request a “picky friend” to point out clichés, lack of personal details, or awkward phrasing, and finally produce an improved version. The author observes that the second‑round output is typically a level higher.

Overall, effective prompting is about directing the model’s attention: instead of asking "What should AI do?", ask "How will AI interpret this sentence and where might it fail?" By designing prompts from the AI’s viewpoint, each sentence becomes an instruction rather than a vague wish, achieving true communication power in the AI era.