MVP 失败大多是因为团队跳过验证、直接造了自己想造的东西。下面 15 个 Prompt 设计成本以天计的预先验证实验。覆盖：落地页测试、假门按钮、Concierge MVP、Wizard-of-Oz 原型、烟雾投放，以及总被忽略的 kill 标准——让失败的实验真正杀掉想法，而不是被解释过去。
适合哪些场景
决定下一步做什么的创始人、做想法验证的成长期 PM、在 5 个想法里选一个的独立开发者、大公司里跑结构化测试的创新团队。
什么时候不建议这样写 Prompt
已经验证过的 roadmap 功能不要用——直接写 PRD。合规或基础设施工作也不要用——那是可靠性问题不是需求问题。
Prompt 结构公式
MVP 验证 Prompt 一定带这六个要素：
- 角色：让 AI 扮演谁（资深 PM / 独立创始人 / 产品设计师 / 独立开发者 / 增长负责人）。
- 上下文：阶段（想法 / MVP / 增长 / 规模化）、团队规模、流量或 ARR、平台（web / iOS / Android）、受众、限制。
- 目标：一个具体交付物——一段 PRD、一组用户故事、一个实验设计、一篇上线公告。
- 限制：时间线（本 sprint / 本季度）、要砍的范围、不能动的东西（现有流程、计费、合规）。
- 输出格式：表格、清单、可贴 ticket 的 JSON、或带标签的段落，能直接粘到 Linear / Notion / Jira。
- 示例 / 信号：1-2 份你欣赏的参考或竞品、加 1 个想避开的反例。
这套 Prompt 适合用在哪
- 预先想法验证
- 在 3-5 个想法里选一个
- 融资阶段需求证明
- Pivot 决策实验
- Hackathon / sprint 想法测试
15 个可直接复制的 Prompt 模板
1. 落地页需求测试
最便宜、杠杆最高。几天造好、一周可证伪。
You are a startup validation coach. Design a landing-page demand test for {idea}: (1) headline + subhead, (2) 3 outcome bullets, (3) email-capture CTA, (4) success metric (conversion rate from cold traffic), (5) sample size needed, (6) traffic source ({reddit / X / paid ads / cold email}), (7) kill criterion (what conversion below which we kill the idea). Total budget: less than $300 and 5 days.
Idea: {paste}可替换变量： idea, audience, budget
优化建议： 成功标准模糊时追加：“Pick a numeric kill criterion — for example, less than 5% email capture from 500 visits — and commit before launching.”
2. 假门按钮测试
Design a fake-door button test for {feature inside existing product}: button placement, copy, what users see when they click ("coming soon" page or interview signup), click-through threshold for greenlight, threshold for kill, ethical disclosure approach. Include the 1-line message shown after click.3. Concierge MVP
Design a concierge MVP for {idea} where humans manually deliver the value before any product exists. Define: target customer to recruit (first 5), what value we deliver manually, how we measure satisfaction, how long we run it ({4-8 weeks}), what we learn that automation would not show. End with success criteria for moving to a real build.4. Wizard-of-Oz 原型
Design a Wizard-of-Oz prototype for {idea}: user-facing UI that looks automated, human backend that fakes the response, success criteria (does the user behave as if it were real?), ethical caveat (when to tell them), and how to transition to real backend.5. 烟雾投放
Design a smoke-test ad campaign for {idea}: 3 ad creative variants targeting different angles, exact target audience, budget allocation ({$50-150 per variant}), landing page measurement, success threshold per variant. Define what would prove "no one wants this".6. 预售验证
Design a pre-sale validation for {idea}: pricing offered, refund policy, how many pre-sales needed to confirm demand, delivery timeline, what data we collect from buyers. End with the boundary between "validated" and "should refund".7. 5 客户承诺测试
Design a "5 customers commit before we build" test for {idea}: how we recruit the 5, what commitment we ask for (money / time / signed letter of intent), timeline, what counts as confirmation. End with the disqualifier — what would invalidate this evidence.8. 手工 workflow 演示
Before building, design a manual workflow that solves {problem} using existing tools ({Notion + Zapier + email}). Demo to 3 target users. Measure: would they pay, would they switch tools, would they wait for us to automate it. Success criteria: at least 2 of 3 say yes to all three.9. 可证伪假设写法
Convert {idea} into a falsifiable hypothesis: "{Target users} have {problem} so badly that they will {specific action} when shown {minimal offer}." For each clause, explain how it would be measured and what would falsify it. End with the experiment that tests it most cheaply.10. 5 个想法之间的 triage
I have 5 product ideas. For each: 1-line problem, 1-line solution, target user, the cheapest validation experiment, expected time and cost, what specifically would kill it. Output as a 5-row comparison table. Recommend which to validate first.
Ideas: {paste}11. 用户访谈设计
Design a 30-minute customer interview to validate the problem behind {idea}: 8 open questions (no leading), what behavior to observe, what to NOT mention until they bring it up, success signal (they describe the pain unprompted). End with red flags (interviewer-leading questions) to avoid.12. 自建 vs 购买 vs 验证 决策
For {feature/idea}, decide: build now, buy a third-party, or validate further. For each path: cost, time, reversibility, learning. Recommend one with reasoning. Add a tripwire: "if X happens we reconsider".13. 验证 post-mortem
Below is the data from a validation experiment we ran. Write the post-mortem: (1) hypothesis tested, (2) what we measured, (3) result, (4) honest read (succeeded / failed / inconclusive), (5) what we will do now, (6) what we would do differently if rerunning.
Data: {paste}14. 避免问卷偏差
Below is a customer survey designed to validate {idea}. Audit for leading questions, social-desirability bias, "would you pay" trap, vague metrics. Rewrite the 3 most biased questions and explain why each rewrite is more reliable.
Survey: {paste}15. 三个月验证 roadmap
For {idea cluster}, design a 12-week validation roadmap with 4 experiments. Each experiment: hypothesis, method (landing / fake-door / concierge / pre-sale), cost, success criterion, kill criterion, what we move to next. Output as a Gantt-style table.容易踩的坑
- 先造再验证——每个团队都觉得自己是例外，其实都不是。
- 没有 kill 标准——失败也会被解释过去。
- “你愿意付吗”问卷而没有真实付款——回答会撒谎。
- 验证流量只挑朋友 / 网络内——他们总说 yes。
- 把”激动”当验证——激动会衰减，行为不会。
- 因为”不可扩展”就跳过 Concierge——验证阶段 scale 无关紧要。
- 一次成功就拍板——复现一次再 commit。
优化技巧
- 实验开跑前先写 kill 标准。
- 先验证问题，再验证方案，再验证价格——顺序不能乱。
- 验证用户从网络之外招募；朋友会让结果偏 3 倍。
- 看行为（点击、付款、时长）胜过看表态。
- 先承诺一个便宜实验；pre-mortem 防 scope creep。
- 每个想法在 /validation 日志里记录尝试——多数团队会忘。
- 验证阶段 Concierge / Wizard-of-Oz 几乎总打败完整 MVP。
FAQ
- 最便宜的验证是什么？: $100-300 冷流量 + 一周窗口的落地页测试。用模板 1。
- 需要多少流量？: 落地页至少 300-500 个独立访客才能看转化。再少都是噪音。
- 只做访谈够吗？: 不够。访谈验证问题，不验证方案和付费意愿。要配模板 7（5 客户承诺）。
- 什么时候停止验证、开始造？: 至少 5-10 个目标用户有付款承诺或行为信号、且 kill 标准未被触发时。
- AI 能告诉我想法好不好吗？: 不能。AI 帮你设计便宜实验。“好”由真实客户判断，不是模型。