Back to Details

generate-startup-ideas

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Startup Idea Machine

Startup Idea Machine

Turn a blank prompt into a multi-round idea hunt. Ask three focused questions, search current public signals, generate concise idea bursts, learn from the founder's reactions, and validate only the finalists.
Respond in English even when the user's request is written in another language, unless the user explicitly requests a different output language.
将空白提示转化为多轮创意挖掘流程。提出三个针对性问题，搜索当前公开信号，生成简洁的创意方案，根据创始人的反馈学习，仅对入围方案进行验证。
即使用户的请求使用其他语言，也请用英文回复，除非用户明确要求使用其他输出语言。

Load the right references

加载正确的参考资料

  • Read references/intake-and-rounds.md before starting or continuing an interactive hunt.
  • Read references/research-lenses.md before searching the web or selecting source types.
  • Read references/evidence-rules.md before citing signals, competitors, spending, or timing.
  • Read references/idea-session-schema.md before writing session, round, feedback, or validation JSON.
  • Read references/originality.md before generating a second round or checking diversity.
  • Read references/scoring.md before ranking finalists.
  • Read references/report-schema.md before producing the final report.
  • 在开始或继续交互式创意挖掘前，请阅读references/intake-and-rounds.md。
  • 在搜索网页或选择来源类型前，请阅读references/research-lenses.md。
  • 在引用信号、竞品、支出或时机相关内容前，请阅读references/evidence-rules.md。
  • 在编写会话、轮次、反馈或验证JSON前，请阅读references/idea-session-schema.md。
  • 在生成第二轮创意或检查多样性前，请阅读references/originality.md。
  • 在对入围方案排名前，请阅读references/scoring.md。
  • 在生成最终报告前，请阅读references/report-schema.md。

Core workflow

核心工作流

1. Start with three questions

1. 从三个问题开始

For an open-ended request, ask the following three questions together and wait:
  1. What are you unusually good at, curious about, or experienced in?
  2. What customers, markets, or business styles attract you, and what should be avoided?
  3. What constraints matter: time, budget, technical ability, team, geography, and ambition?
Keep the questions easy to answer. Let the user say "surprise me" or "no preference." If the user already supplied an answer, do not ask for it again. Ask at most one compact follow-up only when a missing constraint would materially change the hunt.
对于开放式请求，同时提出以下三个问题并等待回复：
  1. 你在哪些方面格外擅长、好奇或有丰富经验？
  2. 你感兴趣的客户、市场或商业模式类型是什么，哪些是需要避开的？
  3. 哪些约束条件很重要：时间、预算、技术能力、团队、地域及目标抱负？
问题要易于回答。允许用户回复“给我惊喜”或“无偏好”。如果用户已经提供了相关答案，请勿重复提问。仅当缺失的约束条件会实质性改变创意挖掘方向时，才提出最多一个简洁的跟进问题。

2. Create the idea session

2. 创建创意会话

Save the normalized answers as JSON in the workspace 
outputs/
directory. Run:
bash
node <skill-dir>/scripts/create_session.mjs <intake.json> <startup-idea-session.json>
Keep using the same session file throughout the hunt. Do not discard rejected ideas; they teach the machine what to avoid.
将标准化后的答案保存为JSON文件到工作区
outputs/
目录。运行：
bash
node <skill-dir>/scripts/create_session.mjs <intake.json> <startup-idea-session.json>
在整个创意挖掘过程中持续使用同一个会话文件。请勿丢弃被否决的创意；它们能帮助机器了解需要避开的方向。

3. Run a fast public-signal scan

3. 快速扫描公开信号

Browse the current public web after receiving the answers. Use at least four query angles and at least three source types from references/research-lenses.md.
Collect enough evidence to inspire and constrain the first burst without turning the opening into a long market report. Prefer original sources and capture URLs, dates when visible, actors, workflows, workarounds, and spending proxies.
收到答案后浏览当前公开网络内容。使用至少四个查询角度，以及至少三种来自references/research-lenses.md的来源类型。
收集足够的证据来启发和约束第一轮创意生成，避免将初始环节变成冗长的市场报告。优先选择原始来源，并记录URL、可见的日期、参与者、工作流程、替代方案及支出指标。

4. Fire the first idea burst

4. 生成第一轮创意方案

Generate 12 distinct ideas. Keep each idea compact enough to scan quickly.
For every idea include:
  • Name
  • Customer
  • Pain
  • Product
  • Signal — one source-backed reason or a clearly labeled hypothesis
  • Why you — connection to the founder's answers
Cover multiple idea shapes and business models. At least eight ideas must connect to recorded public signals. Label the remaining ideas "exploratory" rather than fabricating evidence.
Do not create a large scoring table in the first burst. End by asking the user to react using:
text
Love: [numbers]
Maybe: [numbers]
No: [numbers]
Save the round JSON and run 
append_round.mjs
.
生成12个不同的创意。每个创意要足够简洁，便于快速浏览。
每个创意需包含：
  • 名称
  • 客户群体
  • 痛点
  • 产品内容
  • 信号 — 一个有来源支撑的理由或明确标记的假设
  • 适配性 — 与创始人答案的关联点
覆盖多种创意类型和商业模式。至少8个创意必须与已记录的公开信号相关联。剩余的创意标记为“探索性”，而非编造证据。
第一轮创意生成时不要创建大型评分表。结尾请用户通过以下格式反馈：
text
Love: [数字]
Maybe: [数字]
No: [数字]
保存轮次JSON文件并运行
append_round.mjs
。

5. Learn from reactions

5. 从反馈中学习

Record the user's reactions with 
record_feedback.mjs
. Extract positive and negative patterns from idea tags, customers, business models, and reasons.
Do not interpret rejection as proof that an opportunity is bad. It only indicates founder preference. Preserve an explicit distinction between market evidence and personal taste.
使用
record_feedback.mjs
记录用户的反馈。从创意标签、客户群体、商业模式及理由中提取正面和负面模式。
请勿将否决视为机会不佳的证明，它仅代表创始人的偏好。要明确区分市场证据和个人偏好。

6. Fire the refinement burst

6. 生成优化后的创意方案

Generate eight new ideas:
  • four direct descendants of loved ideas;
  • two adjacent ideas that preserve the attractive mechanism but change the customer or market;
  • two wildcards that challenge the emerging pattern without repeating rejected traits.
Do not rename old ideas and present them as new. Use the originality checks in references/originality.md. Keep the cards concise and ask the user to select one to three finalists.
Repeat the feedback and refinement loop when the user wants more ideas. A session can contain any number of rounds.
生成8个新创意：
  • 4个为用户喜欢的创意的直接衍生版本；
  • 2个为相邻创意，保留有吸引力的机制但更换客户群体或市场；
  • 2个为wildcard创意，挑战现有模式但不重复已被否决的特征。
请勿将旧创意重命名后当作新创意呈现。使用references/originality.md中的原创性检查规则。保持创意内容简洁，并请用户选择1到3个入围方案。
当用户需要更多创意时，重复反馈和优化循环。一个会话可以包含任意数量的轮次。

7. Validate the finalists

7. 验证入围方案

For each selected finalist, research:
  • repeated problem evidence;
  • current workaround and switching friction;
  • spending, hiring, or purchase proxies;
  • direct and adjacent competitors;
  • underserved wedge;
  • reachable first customers;
  • why now;
  • fatal assumption;
  • disconfirming evidence;
  • smallest useful MVP;
  • a seven-day validation test.
Use current sources and original pages. Do not claim market size, willingness to pay, customer interest, or competitive absence without evidence.
针对每个选中的入围方案，调研以下内容：
  • 重复出现的问题证据；
  • 当前的替代方案及转换成本；
  • 支出、招聘或采购指标；
  • 直接和间接竞品；
  • 未被充分服务的细分领域；
  • 可触达的首批客户；
  • 为何现在切入；
  • 致命假设；
  • 反证证据；
  • 最小可用MVP；
  • 七天验证测试。
使用当前的原始来源页面。请勿在没有证据的情况下断言市场规模、付费意愿、客户兴趣或无竞品。

8. Score and select

8. 评分与筛选

Score finalists using references/scoring.md. Show criterion-level scores and the reason behind each score.
The highest score is not automatically the winner. Choose the winner by balancing evidence, founder fit, reachability, buildability, and fatal risk. It is acceptable to conclude that none should be built yet.
Save the validation JSON and run 
complete_session.mjs
.
使用references/scoring.md对入围方案评分。展示各标准的得分及背后的理由。
最高分方案并非自动成为胜出者。需平衡证据、创始人适配性、可触达性、可构建性及致命风险来选择胜出方案。也可以得出暂不启动任何方案的结论。
保存验证JSON文件并运行
complete_session.mjs
。

9. Create the native Codex report

9. 创建原生Codex报告

Run:
bash
node <skill-dir>/scripts/generate_report.mjs <startup-idea-session.json> <startup-idea-machine-report.md>
Save the report in the workspace 
outputs/
directory. Verify that source links, reaction history, scores, risks, unknowns, and the seven-day plan render correctly.
Return clickable links to both the Markdown report and session JSON.
运行：
bash
node <skill-dir>/scripts/generate_report.mjs <startup-idea-session.json> <startup-idea-machine-report.md>
将报告保存到工作区
outputs/
目录。验证来源链接、反馈历史、评分、风险、未知项及七天计划是否正确显示。
返回Markdown报告和会话JSON的可点击链接。

Hunt modes

创意挖掘模式

  • open-web: Search across markets and idea shapes. Default when no niche is supplied.
  • niche: Stay within one industry, audience, or workflow.
  • boring-business: Prioritize recurring operational work, legacy tools, and clear budgets.
  • manual-work: Look for spreadsheets, copy-paste work, assistants, agencies, and repeated services.
  • unbundle: Look for valuable pieces of expensive or bloated software.
  • new-tech: Combine newly available technology with an old, expensive workflow.
  • github-demand: Prioritize feature requests, issues, scripts, and open-source commercialization gaps.
  • local-opportunity: Prioritize location-bound businesses and workflows.
Use 
open-web
by default. The user can mix modes.
  • open-web：跨市场和创意类型进行搜索。未指定细分领域时默认使用此模式。
  • niche：聚焦单一行业、受众或工作流程。
  • boring-business：优先考虑重复性运营工作、传统工具及明确预算的业务。
  • manual-work：寻找电子表格、复制粘贴工作、助理服务、代理机构及重复性服务相关的机会。
  • unbundle：寻找昂贵或臃肿软件中的高价值功能模块。
  • new-tech：将新可用技术与老旧、高成本的工作流程相结合。
  • github-demand：优先考虑功能请求、问题反馈、脚本及开源商业化缺口。
  • local-opportunity：优先考虑受地域限制的业务和工作流程。
默认使用
open-web
模式。用户可以混合使用多种模式。

Bundled scripts

内置脚本

  • create_session.mjs
    — normalize founder answers and initialize session state.
  • append_round.mjs
    — validate and append a source-backed idea round.
  • record_feedback.mjs
    — persist love/maybe/no reactions and derive taste signals.
  • complete_session.mjs
    — attach finalist research, scoring, winner, and validation plan.
  • generate_report.mjs
    — render the session as native Markdown.
  • create_session.mjs
    — 标准化创始人的回答并初始化会话状态。
  • append_round.mjs
    — 验证并追加有来源支撑的创意轮次。
  • record_feedback.mjs
    — 保存喜欢/待定/不喜欢的反馈并推导偏好信号。
  • complete_session.mjs
    — 添加入围方案调研、评分、胜出者及验证计划。
  • generate_report.mjs
    — 将会话内容渲染为原生Markdown格式。

Quality rules

质量规则

  • Start ideating after the questions; do not bury the user in methodology.
  • Prefer concrete customers and workflows over broad categories.
  • Keep the first two rounds varied and fast.
  • Link material validation claims to sources.
  • Separate observed evidence, inference, and speculation.
  • Search for counterevidence before recommending a winner.
  • Never present generated people, quotes, revenue, market size, or customer interest as factual.
  • Do not bypass logins, paywalls, access controls, rate limits, or private communities.
  • Do not use leaked, private, or sensitive personal data.
  • Never contact prospects, submit forms, or publish anything without separate user authorization.
  • 先提出问题再开始创意生成；不要让用户淹没在方法论中。
  • 优先选择具体的客户群体和工作流程，而非宽泛的类别。
  • 前两轮创意要多样化且快速推进。
  • 重要的验证声明需链接到来源。
  • 区分观测证据、推论和猜测。
  • 在推荐胜出方案前先搜索反证证据。
  • 切勿将生成的人物、引用、收入、市场规模或客户兴趣当作事实呈现。
  • 不要绕过登录、付费墙、访问控制、速率限制或私有社区。
  • 不要使用泄露的、私人的或敏感的个人数据。
  • 未经用户单独授权，切勿联系潜在客户、提交表单或发布任何内容。

Default output

默认输出流程

Three Questions → 12-Idea Burst → Founder Reactions → 8-Idea Refinement → 1–3 Validated Finalists → Winner or No-Go → Native Markdown Report
三个问题 → 12个创意方案 → 创始人反馈 → 8个优化创意 → 1–3个验证后的入围方案 → 胜出方案或暂缓启动 → 原生Markdown报告