SKILL: Idea Generator (想法发生器) v1

Position in Decision Flow

没方向 → [Idea Generator 想法发生器] → 候选清单
                                        ↓
                                   [Idea Evaluator 想法质检] → 评分≥6 → 六模式审评

This skill answers: "What directions should I even be looking at?"
Idea Evaluator answers: "Is this specific idea any good?"

---

Pre-Generation: Context Inventory (上下文盘点)

Before running the 7 recipes, inventory the user's available inputs. Each input is fuel for idea generation — the more specific, the better the output.

inventory_checklist:
  skills:
    - Core technical skills (languages, frameworks, domains)
    - Non-technical skills (sales, writing, domain knowledge, languages spoken)
    - Rare skill combinations (e.g., "Go + quantitative finance + AI agents")

  industry_exposure:
    - Current job industry and daily workflow observations
    - Previous job industries
    - Outsourcing/freelance client industries
    - Industries of close friends/family

  pain_points_observed:
    - Problems you personally experience repeatedly
    - Problems you've seen colleagues/clients struggle with
    - Complaints you've heard from specific people (not abstract "market needs")
    - Broken workflows you've witnessed firsthand

  assets:
    - Existing code/products (even frozen ones)
    - Existing user base (even small)
    - Domain knowledge accumulated
    - Relationships/access to specific user groups

  recent_changes:
    - New technologies that just became viable
    - Regulatory changes affecting industries you know
    - Behavioral shifts you've noticed (post-COVID, AI adoption, etc.)
    - Industry disruptions in your orbit

Output: A filled context card. This card is referenced by every recipe below.

---

The 7 Recipes

Recipe 1: Start From Your Own Problems (从自身痛点出发)

method: >
  List every friction/annoyance/inefficiency you personally encounter
  in a typical week — at work, at home, managing finances, learning,
  communicating, building software, managing projects.

prompt_questions:
  - "What did you do this week that felt unnecessarily painful or slow?"
  - "What task do you keep postponing because it's annoying?"
  - "What manual process do you repeat that should be automated?"
  - "What tool do you use daily that frustrates you?"

quality_filter: >
  Best results come from problems you experience as a PROFESSIONAL,
  not as a consumer. Consumer problems attract massive competition.
  Professional/B2B problems are often invisible to outsiders.

output: List of [problem + who else has it + current workaround]

Recipe 2: Start From What You're Uniquely Good At (从独特能力出发)

method: >
  Identify your rare skill combinations. A single skill (Go, AI, frontend)
  is common. Two skills combined (Go + AI) is less common. Three skills
  combined (Go + AI + quantitative trading infrastructure) is rare.
  Look for problems that sit at the intersection of your rare combination.

prompt_questions:
  - "What can you build in a weekend that would take most teams a month?"
  - "What do people ask you for help with that they can't easily get elsewhere?"
  - "What technical problem have you solved that most engineers haven't faced?"

⚠️ SISP guard: >
  This recipe is the highest-risk for SISP (Bug 5). The output must be
  [capability → problem it solves → who has that problem], NOT
  [capability → cool thing I could build]. If you can't name a specific
  person with the problem, the idea is SISP. Flag it and move on.

output: List of [rare capability intersection + specific problem it solves + for whom]

Recipe 3: Start From Changes You've Noticed (从你观察到的变化出发)

method: >
  What has recently changed in technology, regulation, behavior, or
  industry structure that creates a new gap? The best startup timing
  comes from riding a wave that just started — not one that's peaking.

prompt_questions:
  - "What became possible in the last 12 months that wasn't before?"
  - "What industry is being forced to change by new regulation or technology?"
  - "What behavior shift have you observed in people around you?"
  - "What used to be expensive/hard that is now cheap/easy?"

timing_check: >
  Cross-ref with Model 06 (Contrarian Timing). If "everyone" already
  knows about this change, you're late. Look for changes that are
  obvious to insiders but invisible to outsiders.

output: List of [change + new gap created + who is affected + timing assessment]

Recipe 4: Start From Industries You're Inside (从你接触的行业出发)

method: >
  Your current job, freelance clients, and personal network give you
  insider access to specific industries. Most outsiders can't see the
  real problems inside these industries. This is your information
  advantage — use it.

prompt_questions:
  - "What do people in your industry complain about at lunch?"
  - "What process at your company is shockingly manual or outdated?"
  - "What data exists in your industry that no one is using well?"
  - "What do your outsourcing clients keep asking for that isn't a product yet?"
  - "What would make your current job 10x easier?"

⚠️ schlep_check: >
  Industry problems often involve schlep — messy integrations, legacy
  systems, regulatory compliance, relationship-dependent sales.
  If you feel yourself dismissing an idea because "it's too industry-specific"
  or "not techy enough", that's Bug 6 (Schlep Blindness). Flag it
  and force yourself to score it anyway.

output: List of [industry + specific problem + who owns budget for this + how they cope now]

Recipe 5: Start From Recent Tech Breakthroughs (从技术突破出发)

method: >
  A new technology capability (LLMs, AI agents, multimodal models,
  cheap inference, voice AI, etc.) has just made something possible
  that wasn't before. What specific workflow can now be 10x better?

prompt_questions:
  - "What task required a human expert last year that AI can now handle?"
  - "What was too expensive to automate before but is now cheap?"
  - "What data was unstructured/unusable before but LLMs can now parse?"
  - "What manual QA/review/analysis process can now be automated?"

⚠️ SISP guard (strict): >
  This recipe has the HIGHEST SISP risk. The question is NOT
  "what can AI do?" but "what painful workflow can AI fix for
  a specific person who will pay?" If your answer starts with
  "AI can..." instead of "[Person X] struggles with...", stop.
  Reverse the direction.

⚠️ tar_pit guard: >
  "AI + [broad category]" is almost always a tar pit.
  "AI writing assistant", "AI study tool", "AI personal assistant"
  — these are graveyards. Be specific: AI for [specific role]
  doing [specific task] in [specific industry].

output: List of [new capability + specific workflow improved + who benefits + why now]

Recipe 6: Talk to People (跟人聊)

method: >
  Structured conversations with potential users — not friends, not
  other developers. People who do real work in real industries.
  The goal is NOT to pitch an idea. The goal is to discover problems.

conversation_template:
  1. "What's the most tedious part of your job?"
  2. "What tools do you use daily? What do you hate about them?"
  3. "If you could wave a magic wand and fix one thing, what would it be?"
  4. "How much time/money does [that problem] cost you per month?"
  5. "Have you tried to fix it? What happened?"

who_to_talk_to:
  - Your outsourcing clients (you already have access)
  - Colleagues at your current company
  - People in WeChat/Telegram industry groups
  - Friends who work in non-tech industries (they have unsolved problems tech people never see)

minimum_bar: >
  Talk to 5 people outside your immediate circle before concluding
  "there are no good ideas". If you haven't talked to 5, you haven't
  looked — you've just thought.

output: List of [person's role + their problem + severity + current solution + willingness to pay]

Recipe 7: Find Incumbent Weaknesses (找巨头的弱点)

method: >
  Large companies and established products always have blind spots.
  They can't serve small niches profitably. They can't move fast.
  They can't customize. They can't care about individual users.
  Find a specific segment they're underserving.

prompt_questions:
  - "What large software product do people in [industry] use reluctantly?"
  - "What's the most common complaint about [dominant tool]?"
  - "Who is being overcharged by [incumbent] relative to the value they get?"
  - "What user segment is too small for [big company] to care about?"

quality_signal: >
  The best answers come from actual user complaints — app store reviews,
  Reddit threads, V2EX posts, industry forums. Not from your imagination.

output: List of [incumbent + their weak point + underserved segment + what "good enough" alternative looks like]

Recipe 8: 10-Lens Product Audit (十维产品审视)

method: >
  Pick a concrete product, tool, or workflow area you know deeply (your own
  product, a tool you use daily, a client's system). Force yourself through
  each of the 10 lenses below and write down what you find. Each lens answers
  "where is the value leak?" — the output is a list of specific problems,
  each of which can become an idea candidate via other recipes.

when_to_use: >
  When you have deep familiarity with a specific product/area but no clear
  direction. Also the best recipe for re-mining an existing product for
  expansion opportunities.

lenses:
  - speed:        "What takes too long?" (e.g., instant search, predictive loading)
  - automation:   "What's repetitive?" (e.g., auto-scheduling, smart defaults)
  - intelligence: "What could be smarter?" (e.g., recommendations, anomaly detection)
  - integration:  "What else do users use alongside this?" (e.g., calendar sync, export)
  - collaboration:"How do users work together here?" (e.g., sharing, comments, real-time)
  - personalization: "How is every user different?" (e.g., custom views, preferences)
  - visibility:   "What's hidden that shouldn't be?" (e.g., dashboards, progress tracking)
  - confidence:   "What creates anxiety?" (e.g., confirmations, undo, previews)
  - delight:      "What could spark joy?" (e.g., polish, celebrations that don't feel gamified)
  - access:       "Who can't use this yet?" (e.g., mobile, offline, accessibility)

rules:
  - BE SPECIFIC — "better UX" is not a finding. "One-click rescheduling
    from notification" is.
  - SMALL CAN BE HUGE — don't dismiss simple findings. Sometimes one button
    changes everything. A tiny lens hit can anchor a whole product.
  - CITE EVIDENCE — reference what you actually saw in the product or heard
    from users, not imagination.

output: List of [lens + specific problem found + who has it + current workaround]

---

Unstick Playbook (卡壳急救)

当任何 recipe 跑不动、产出为空或感觉"想不出来了"时，按顺序试：

1. 换 lens / 换 recipe — 卡住通常不是没想法，而是当前视角已耗尽。跳到另一个 lens 或另一个 recipe。
2. Ask yourself the unstick questions:
   • "什么会让用户主动跟朋友说这个产品？"
   • "用户每天做的那件有点烦的事是什么？"
   • "如果我们团队扩大 10 倍会做什么？缩小到 1/10 呢？"
   • "竞品要做出什么才能打败我们？"
   • "高级用户现在手动在做什么、我们可以做成原生功能的事？"
   • "我们手里有但用户看不到的数据/洞察是什么？"
   • "哪个功能听起来很疯但也许能成？"
3. 回到 Recipe 6 — 跟 5 个真实用户聊。想不出来时，答案几乎总在外部而不在脑内。
4. 接受间隔 — 如果今天真的挖不出来，停手。每轮生成建议间隔 2 周，中间需要新的信息输入。

---

Post-Generation: Candidate Assembly

After running relevant recipes (not all 8 are needed every time — pick 3-4 most relevant based on the context inventory), assemble candidates.

Deduplication & Clustering

Group similar ideas together. Often multiple recipes point at the same underlying opportunity from different angles — that's a strong signal.

Quick Viability Filter

Before sending to idea-evaluator, apply a 3-second gut check on each candidate:

kill_if:
  - You can't describe who pays and why in one sentence
  - The only user you can imagine is "everyone"
  - You feel excited about the TECHNOLOGY but bored by the USER PROBLEM
  - You've seen this exact idea fail 3+ times (tar pit)

keep_if:
  - You can name a specific person who has this problem
  - The idea makes you slightly uncomfortable (schlep signal)
  - You can see how to test it in <48 hours
  - Multiple recipes converged on this direction

---

Output Format

## 想法生成报告 (Idea Generation Report)

### 上下文盘点
- 技能栈: [summary]
- 行业接触: [summary]
- 已观察痛点: [summary]
- 已有资产: [summary]
- 近期变化: [summary]

### 使用的Recipes
[List which recipes were run and why]

### 候选清单

| # | 想法 | 来源Recipe | 目标用户 | 核心痛点 | Schlep指数 | 初筛 |
|---|------|-----------|---------|---------|-----------|------|
| 1 | [idea] | R[X] | [who] | [pain] | 高/中/低 | ✅/❌ |
| 2 | ... | ... | ... | ... | ... | ... |

### 收敛信号
[哪些方向被多个recipe指向？这些是最值得优先评估的]

### 下一步
对初筛通过的候选，逐个运行 idea-evaluator 想法质检。
建议优先评估: #[X] 和 #[Y]（理由）

---

Usage Notes

• 不需要每次跑全部8个recipe。根据上下文盘点结果，选3-4个最有燃料的recipe跑。
• 已有具体产品/工具时优先跑 Recipe 8（十维审视）；完全没方向时从 Recipe 1/3/6 起步。
• Recipe 4（行业内部）和 Recipe 6（跟人聊）产出质量通常最高，因为它们基于真实信息而非推测。
• Recipe 2 和 Recipe 5 的SISP风险最高，每个产出都要过SISP检测。
• Schlep指数高的想法不是坏想法——恰恰相反，它们往往竞争更少。标注它是为了让你正视它，而不是过滤它。
• 如果跑完所有recipe产出为零，问题不是"没有好想法"，而是你的观察输入不够。回到Recipe 6，跟5个真实用户聊。
• 每轮生成建议间隔2周。连续生成想法不会提高质量——中间需要新的信息输入（工作观察、用户对话、行业动态）才能产出不同的候选。

Final step

After completing the analysis, call the tool SaveIdeaReport with:

• title: a concise, filesystem-safe title for the idea
• content: the full Markdown report content (include # headings, problem, alternatives, recommendation, next steps)
• type: "generator" (required — must be exactly "generator")

Do not write the file directly; always use SaveIdeaReport so the output is validated and stored under .ody-code/ideas/.