Summary
A practitioner’s guide to 6 prompt techniques (心法, “mental methods”) for more effective AI conversations, written by a 2-3 year AI power user. The core insight is that prompt engineering is less about syntax and more about conversation design — expressing context clearly and structuring the interaction so the AI surfaces what it needs before answering. The 6 techniques: AI-selected expert persona, Socratic pre-questioning, adversarial debate, pre-mortem failure simulation, reverse prompt engineering, and dual-layer explanation.
一位使用 AI 兩三年的重度使用者分享的 6 個 Prompt 心法，核心洞見是 Prompt 工程不是語法技巧而是對話設計——讓 AI 在回答前先獲取它需要的上下文。六個技巧：AI 選定專家角色、蘇格拉底式追問、對抗辯論、預演失敗、反向提示工程、雙層解釋法。
Key Points
- Tech 1 — AI-selected persona: instead of guessing which expert role to assign, ask AI to pick the most appropriate one first, then describe the problem
- Tech 2 — Socratic pre-questioning: “Ask me questions one at a time until you have 95% confidence in understanding my real need, then answer” — the 95% threshold prevents both shallow answers and infinite loops
- Tech 3 — Adversarial debate: explicitly tell AI to act as an opponent trying to disprove your idea — counters AI sycophancy; author spent 3 hours in debate and was genuinely changed
- Tech 4 — Pre-mortem simulation: “Assume this project failed — what were the decay signals, the fatal decisions, the ignored risks?” based on real similar failures; surfaces cognitive blind spots
- Tech 5 — Reverse prompt engineering: give AI a desired output, ask it to reverse-engineer the prompt that generated it, with explanation of each element’s function — useful for learning style/structure
- Tech 6 — Dual-layer explanation: request two versions simultaneously: beginner analogy (accessible, no jargon) + expert technical version; cross-reference for deeper learning
- Core philosophy: “Make dialogue a collaboration, make questions a design” — you are the director, not the audience
Insights
- The 95% confidence threshold in Tech 2 is the most precise formulation I’ve seen for this pattern — it gives the AI a stopping condition that’s neither too low (shallow) nor impossibly high (infinite loop)
- Tech 3 (adversarial debate) is the antidote to a structural problem in RLHF-trained models: they’re trained to be agreeable, which actively harms their usefulness for stress-testing ideas
- Tech 4 (pre-mortem) is the AI version of a classical decision-making technique from Gary Klein — the insight is that asking “how could this fail?” unlocks different reasoning than “what’s your plan?”
- The “洗脚城大爷” (foot bath parlor elderly man) persona for beginner explanations is funnier and more effective than “explain like I’m 6” because it targets adult life experience without triggering childish vocabulary
- Tech 5 (reverse engineering) reframes prompt learning: instead of memorizing prompt formats, you analyze outputs you admire and extract the underlying structure — this is learning by reverse engineering rather than by imitation
Connections
- Prompt Engineering
- AI Agents
- Context Engineering
- Learning
- Everyone using AI has about 12 months to develop these 3 moats
Raw Excerpt
让对话变成你跟AI的协作，让提问变成你对AI的设计。你不需要做Prompt大佬。你只需要做自己问题的导演。