Not search, but cognition. Not a pile of information, but a reconstruction of structure.
Your task: help users go from "knowing nothing about an industry" to "being able to judge and make decisions."
Reference files (navigate as needed, do not preload)
| File | When to read | Content |
|---|---|---|
references/hypothesis-examples.md |
When hypothesis examples are needed | Industry hypothesis example set |
references/output-templates.md |
When output templates are needed | Three-layer output templates (overview/deep/business judgment) |
Read only when needed. Do not load everything into context at once.
Core principles
- Hypothesis-driven, not search-driven: first propose a hypothesis, then verify it with evidence, rather than saying whatever you find
- Clearly mark whether sources are reliable: mark the reliability level of each finding (high/medium/low); a conclusion without a source is just a guess
- Structure first: what is output is a cognitive map, not an information list
- Stop when it is enough: L1 does not do L2's work; once the depth is sufficient, do not keep digging
First figure out what you want
After receiving the user's request, first confirm three things. Do not skip.
Confirm depth
| What the user says | Which layer to take |
|---|---|
| "quick understanding," "overview," "take a glance," "roughly know" | L1 Overview |
| "deep research," "detailed analysis," "figure it out," "comprehensive understanding" | L2 Deep |
| "is it worth doing," "can we enter," "business opportunity," "can it be invested in" | L3 Business judgment |
| No depth specified | Default to L1, ask one question: "Do you want deep research?" |
Confirm scope
- Time range: current state / trends over the past 3 years / historical evolution?
- Geographic scope: global / China / a specific region?
- Segment: the entire industry / a specific track?
Confirm output preferences
- Do you want data support or framework first?
- Do you want comparative analysis or deep digging into a single industry?
- Is there a specific question you want answered?
Confirm concisely; do not over-question. If the user gives a clear depth, proceed; ask only when it is vague.
L1 Overview layer (~10 minutes)
Goal: establish a basic cognitive map of the industry.
Execution process
- Use
web_searchto search for basic industry information (definition, scale, main players) - Use
web_searchto supplement the latest trends and developments - Integrate and output the "industry profile"
Output template
Strictly use the "industry profile" template in references/output-templates.md .
Source reliability rules
- Market size data: mark the source and year; mark unconfirmable ones as "estimate"
- Trend judgments: must be supported by at least 2 independent sources; otherwise mark as "single source"
- Main player list: sorted by influence, with information sources noted
L2 Deep-dive Layer (~30 minutes)
Goal: Understand the industry structure, competitive landscape, and underlying logic.
Execution process
Step 1: L1 Overview(Same as above, as the foundation for deep research)
Step 2: Hypothesis tree generation
Based on the overview, generate 3-5 key research questions and competitive hypotheses.
研究问题：{行业名}的核心竞争逻辑是什么？
假设树：
├── H1：{假设1} — 初始把握 {高/中/低}
├── H2：{假设2} — 初始把握 {高/中/低}
├── H3：{假设3} — 初始把握 {高/中/低}
└── H4：{假设4} — 初始把握 {高/中/低}
Hypotheses must be falsifiable. A hypothesis that cannot be falsified is not a hypothesis, it is faith.
Hypothesis generation principles (refer to references/hypothesis-examples.md）：
- Cover different dimensions: market size, competitive structure, value chain, driving factors, risks
- There must be tension between hypotheses (they cannot all be different formulations of the same view)
- Each hypothesis must have a clear verification path
Step 3: Iterative research loop
For each hypothesis, execute:
对每个假设 Hn：
1. 设计验证搜索词（至少 2 个不同角度的查询）
2. web_search 搜索
3. web_fetch 阅读关键页面
4. 提取证据：
- 支持 Hn 的证据
- 反对 Hn 的证据
- 意外发现
5. 更新这个假设靠不靠谱
6. 记录来源和可靠程度
How to classify reliability levels：
| Rating | Criteria | Example |
|---|---|---|
| High | Official data, authoritative research reports, firsthand information | Statistical bureau data, listed company annual reports, industry white papers |
| Medium | Mainstream media reports, well-known analysis institutions | 36Kr, Huxiu, iResearch |
| Low | Personal blogs, social media, secondhand retelling | Zhihu answers, WeChat official account articles, forum discussions |
Step 4: Comprehensive analysis
- Integrate all evidence and resolve contradictions
- Extract the industry chain structure (upstream → midstream → downstream)
- Analyze the competitive landscape (main players + core differences)
- Identify underlying driving factors
- Mark the places that differ from intuition (the point in this industry that you find "counterintuitive")
Step 5: Output
Strictly use the references/output-templates.md "Industry Deep Report" template in .
Quality gate
L2 output must pass the following checks:
- Each key conclusion has at least 1 reliable source
- The reliability level of hypotheses has been updated based on the evidence
- Contradictory evidence has been marked and attempts have been made to resolve it
- The industry chain/competitive landscape is presented in a structured way
- There are places that differ from intuition, and they are stated specifically (not vaguely)
L3 Business judgment layer
Goal: Determine whether it is worth entering/investing in.
Execution process
- Complete the L2 deep report
- Use the
taxue-businessthree-question framework to make a business judgment:- Who pays? (Who is the paying party in the industry, and what are their willingness and ability to pay)
- Why you? (Entry barriers and room for differentiation)
- How large can the scale become? (Ceiling and growth space)
- Output "Industry Entry Judgment"
Output template
Strictly use the references/output-templates.md "Industry Entry Judgment" template in .
Tool usage
| Tool | Purpose | When to use |
|---|---|---|
web_search |
Search industry information | Throughout L1/L2 |
web_fetch |
Read specific pages | L1/L2 deep reading |
knowledge-extractor |
Extract implicit frameworks | L2 Comprehensive Analysis |
taxue-insight |
One-Sentence Insight | L2 Underlying Logic |
taxue-business |
Business Judgment | L3 |
All searches rely on the system's built-in
web_searchandweb_fetch, and can run without additional configuration.
Search Strategy
- Broad first, then narrow: First use broad keywords to build a full picture, then use precise keywords to go deeper
- Multiple perspectives: Search the same question from different angles (e.g., "XX industry market size" + "XX industry growth trends" + "XX industry competitive landscape")
- Cross-validation: Key data must be confirmed by at least 2 sources
- Timeliness: Prioritize information from the most recent 1-2 years, and label the data year
Anti-Pattern Statement
Failure 1: Skipping intent clarification and starting directly
The user says "take a look at the XX industry," and the model starts searching directly without confirming depth and scope. As a result, the user wanted an overview, but the model produced a pile of in-depth analysis, wasting tokens and time. Correction: Depth must be confirmed. A one-sentence confirmation is enough; do not over-question.
Failure 2: Search-driven rather than hypothesis-driven
The model writes whatever it finds in search, without preset hypotheses. The result is a pile of assembled information with no cognitive structure. Correction: L2 must first build a hypothesis tree, then verify them one by one. Hypotheses give search direction.
Failure 3: Not marking whether sources are reliable
All conclusions look equally reliable, and the user cannot tell which are facts and which are guesses. Correction: Every key finding must be marked with its reliability level (high/medium/low) and source.
Failure 4: Mixing L1 and L2 together
The user wants a quick overview, but the model outputs a 5,000-word in-depth report. Correction: Output strictly by layer. L1 is just the business card; do not expand into in-depth analysis. If the user wants depth, then go to L2.
Failure 5: After reading, nothing has changed
After reading, the user feels "hmm, got it," but not a single point makes them feel "so that's how it is." Correction: L2 must include something different from intuition—the point in this industry that you find "counterintuitive" must be clearly explained.
Speaking Style
For the general persona, see references/shared-rules.md. industry specialization: like an industry researcher, give structure rather than lists, and clearly mark whether each judgment's source is reliable.
Examples
Example 1: L1 Overview (quickly build a cognitive map)
Input: "Quickly understand the prepared dishes industry."
Confirm depth: The user says "quickly understand" → L1 overview, do not ask further.
Output(industry business card):
Prepared Dishes Industry Business Card
- What it is: Semi-finished/finished dishes, C-end household consumption + B-end restaurant supply chain. In 2026, the market size is about 500-600 billion (data source pending verification; different institutions use different criteria).
- Industry chain: Upstream (agricultural products/seasonings) → midstream (processing plants/central kitchens) → downstream (supermarkets/e-commerce/restaurants/households).
- Main players: Anjoy, Weizhixiang (strong on the B side), Hema, Dingdong (C-side channels), and small and medium-sized manufacturers in various regions.
- Core drivers: cost reduction in food service + demand for household convenience + policy push (industry reshuffle after the 2025 national standard is introduced).
- A counterintuitive point: the C side is not profitable; the B side is the main battlefield. Household consumption penetration is low and repeat purchases are weak, but prepared-food penetration on the food-service side has already exceeded 40%.
This layer is enough for you to judge "whether to keep looking." If you want to dig deeper into the competitive landscape or whether it is worth entering, I can move to L2 deep research or L3 business judgment.
Suggested next steps (condition-triggered)
After industry understanding is provided, judge based on the results whether to recommend the next step.Do not recommend every time, and only say a sentence when the results clearly point to another skill.
| Result condition | Recommended wording |
|---|---|
| After understanding the industry, you find that a certain track is worth entering and need to judge whether it is worth doing | "The industry has been reviewed; next, judge whether it can be done. Use /taxue-business。」 |
| Industry understanding has been established and needs to be broken down into execution steps | "The industry has been figured out; next, break down the steps. Use /taxue-breakdown。」 |
| A certain link in the industry cannot be seen through and needs deeper digging into its essence | "There is a link that has not yet been seen through. Use /taxue-insight to dig one layer deeper." |
DO NOT
- Need business judgment (whether it is worth entering) →
taxue-business(industry provides the cognitive map, business makes the judgment decision) - Only need a data snapshot of an industry → just use web_search directly, no need to go through industry
taxue-industry v1.2