Luopan routes a research question to an industry or a company workflow, then makes you say why

One router file decides whether you meant an industry or a company

Luopan is not two skills that happen to live in the same folder. It is one router plus two judgment models. The root SKILL.md does nothing but decide what kind of object you are asking about and load the matching sub-skill, and the actual methods live in modes/industry/SKILL.md and modes/company/SKILL.md. The split the router enforces is short enough to read as a tree:

罗盘 SKILL.md
├── 行业、赛道、产业链 → 行业研究 Skill
└── 具体公司、股票、岗位 → 公司研究 Skill
    ├── 投资为主
    ├── 求职为主
    └── 投资与求职并重

The industry side answers how the sector runs, who holds the discourse, where the profit sits, and whether entering is worth it. The company side answers how the company makes money, how sound its development is, and whether it is worth investing in or joining. Both sides share the same source grading, the same adversarial verification pass, and the same honesty rules, but they run different judgment models on top of that shared base. A user installs the whole repository and invokes one name, rather than hunting for two skills separately, and the repository is mostly its own output: review-output/ holds the published company reports, test-output/ the industry reports, screenshots/ the figures, and agents/ the helper definitions. The default branch is main, the license is MIT, and there are no GitHub releases, so the published pages rather than a release history are the version record.

The router pauses and refuses to search until you name the purpose

The most interesting line in the flow diagram is the one that stops. When you ask about a company without saying why, the router does three things in order: it asks you to choose, it waits, and it explicitly does not search, fetch, or generate a report before you answer. That branch is the difference between this and a generic research prompt.

The choice has three options, and the third is gated. Investment first means financial quality, business model, moat, management, valuation, and margin of safety. Job first means business prospects, organizational stability, role value, what the job builds for your career, and interview reverse questions. Both lines at once runs only after you say so explicitly. The stated reason for the question is that a job seeker should not have to read long valuation sections, and an investor should not be pulled into role information, and the design refuses to infer your purpose from the type of company. It asks once, and the reasoning is about rework: a report finished in the wrong direction has to be thrown away and redone.

A request that names both objects runs both routes in order, building the industry map first and then placing the company inside it. The example inputs the documentation uses are:

研究行业：帮我研究一下 AI Agent 行业
研究投资：从投资角度研究一下 NVIDIA
研究求职：我准备面试字节跳动，帮我判断是否值得去
用途不明：帮我调研一下腾讯 → 罗盘会请你选择投资、求职或两者并重
组合研究：先研究游戏行业，再看腾讯在行业中的位置

Every fact carries an A, B, or C grade and the data floor is one vendor

The provenance design is the part worth examining, because it is stricter than most research prompts. Structured financial data, covering quotes, financial reports, valuation, consensus expectations, research reports, and capital flow, comes from the Tencent Securities official interface, the same origin as its WeChat mini program and its mobile app, with the deepest coverage claimed for A-shares and Hong Kong stocks. Industry and macro data are cross-checked against authoritative primary materials, and each report tags every item with a source grade of A, B, or C, so a reader can see which claims rest on filings and which rest on secondary writing.

Listed and unlisted companies are handled with different evidence boundaries. A listed company is worked from regulatory disclosure, financial reports, and announcements. An unlisted company without reliable data is explicitly marked as having limited public information, and the stated rule is that media rumors will not be assembled into a false precision of revenue, profit, and valuation. The ByteDance case demonstrates it: because the company is not publicly listed, the repository states that no apparently precise financial model will be built from media estimates, and the job line carries that case instead.

The trade-off is a single-vendor dependency. Concentration in one interface makes the structured layer consistent and cheap to query, and it makes A-share and Hong Kong coverage deep. It also means the grade on a claim is only as good as that one source, and any number you take out of a report has to be re-checked against a filing before money moves.

One fact base, two verdicts, and no combined score

The dual-line case is the clearest statement of the project's thesis. For a company where you ask about both investing and working, the fact base is built once: company identity, financials, business model, and ecosystem position. Then the two judgments run separately and are reported separately, and the design explicitly refuses to blend them into a single score.

The investment line covers how WeChat, games, advertising, payments, and enterprise services monetize together, the quality of revenue, profit, and free cash flow, AI spending against regulatory risk, and a three-scenario valuation with a margin-of-safety test. The job line covers company and business stability, where a specific business group sits, whether the role owns resources, decision space, and results you can attribute, and how to read the real state of a team through naturally asked questions. The published Tencent case lands differently on each line. The investment side calls the company quality acceptable but finds a base-date price of about 590.5 Hong Kong dollars closer to fair value without enough margin of safety. The job side says the company is worth further look while noting that differences between internal business lines may matter more than the difference between this company's brand and its competitors' brands. Same company, two answers, no average.

Valuation runs three scenarios and then reverses the question

The investment path for the NVIDIA case shows the shape of the method. It judges how the company moved from being a chip supplier to being an AI compute platform, checks whether CUDA, networking interconnect, and the full stack form a durable moat, and looks at revenue, profit, free cash flow, and the quality of growth. Then it values the company under pessimistic, baseline, and optimistic scenarios, works backward from the current price to the growth rate that price implies, and computes a good price that still leaves a margin of safety at the target return rate.

The conclusion it reaches is a rejection on price rather than on quality. Company quality passes, but a base-date price of about 210.96 dollars requires normalized cash flow to keep growing at roughly 31 to 32 percent a year for five years, and that fails the margin-of-safety test. The headline line is that a good company is not the same as a good price today. The follow-up questions the report invites show what the method is for: what would the fair price be at target annualized returns of 8, 10, and 12 percent, and how would cloud vendors cutting capital expenditure change the answer. That is a reusable calculation rather than a verdict, and it depends on inputs you supply rather than on a conclusion you are handed.

The job route ends in ten interview questions, not an offer verdict

The ByteDance case runs the other route, and it skips the purpose question entirely because the request already said it was about an interview. The steps are: judge whether the company and the target business are stable, separate core growth from cash cows, strategic experiments, and marginal shrinking lines, check whether the role accumulates results that are provable and transferable, identify the direct supervisor, team stability, and organizational change risk, and finish with ten natural, low-defense questions to ask in the interview itself.

The published conclusion is deliberately narrow. Worth interviewing does not mean worth accepting the offer. The argument is that platform scale, global operations, and AI investment can supply dense career capital, while whether one specific role is worth joining depends on where the business sits, who you report to, whether you own your results, and what you want yourself. The word that gets rejected is the company name as a career credential.

There is also a stated upgrade path. Sending the specific job description, level, and compensation, or sending the answers different interviewers gave, moves the judgment from company level to business level or offer level. That is the mechanism for the weakness in any company-level job report: it cannot know the team before you have been in the room, so the design routes the uncertainty into a question list instead of hiding it in a score.

Reports rank power by discourse, not by position in the chain

The report template is fixed and unusual in one specific way. Every report opens with the core judgment, meaning the answer comes first and is followed by the data, the source, and a so-what line. It then maps upstream, midstream, and downstream power, and it does that ranking by discourse power rather than by position in the value chain. After that come per-company analyses, then the industry outlook and action advice. The distinction matters: a component supplier sitting upstream is not automatically the party with bargaining power, and the project's own worked examples argue that the opposite is often true.

Each formal company report is generated in three formats from the same structured facts: HTML for reading, Markdown for local keeping, and JSON as a reviewable data draft. The industry reports carry numbers that make the power argument concrete, including a knowledge-payment tool at 75.7 percent gross margin against a content platform with no comparable figure, an AI chain where the compute seller is profitable, the model builder is losing billions, and an application layer sits at 91 percent gross margin, and a cross-border seller whose revenue share on the largest platform has fallen to 52 percent. A fourth industry, on emotional consumption, keeps its screenshots in the repository while the full HTML report is absent, which the documentation states rather than hides. The closing table of counterintuitive findings is where the project stakes its own claims, and that table is cut off partway through its last row in the file as published.

Editorial conclusion

Luopan is built for someone who already knows the company or the sector and wants a structured verdict rather than a summary, and its refusal to guess your purpose is the feature that makes the output usable for either an investment or a job decision. It is the wrong tool for a quick factual lookup, and it is a poor fit if you need data sources beyond one Chinese brokerage's interface or if you will not read the A, B, and C grades attached to each claim. Before trusting a number, check three things yourself: which source grade carried it, whether the base date still holds, and whether a private company case was marked as having limited public information rather than quietly estimated.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. zhangxiaoqiang1991/luopan on GitHub