claude-for-financial-services-cn: 63 Claude Skills for A-share Financial Work

The A-share gap in Anthropic's financial services skills

Anthropic's claude-for-financial-services was built around Wall Street workflows: Bloomberg and FactSet for data, JPM and GS report formats, US GAAP for accounting, US Treasury yields as the risk-free rate, and GICS for industry classification. None of that maps cleanly onto A-share practice. A DCF built on a US Treasury rate is not the model a Chinese analyst would defend in a review meeting, and a comparable-company table classified by GICS will not line up with how a domestic research desk organizes coverage.

This project is a direct adaptation of that upstream work for the domestic market. The README's comparison table sets out the substitutions: Wind, iFind and AkShare replace Bloomberg, FactSet and PitchBook; 中金, 中信 and 华泰 replace the English-language bank formats; 中国会计准则 replaces US GAAP; 中债 replaces US Treasuries; and 申万 or 中信 classification replaces GICS. The intended user is a practitioner who already knows these conventions and wants Claude to follow them without being told each time.

The repository is Apache-2.0 licensed and the last push was on 2026-09-14. The primary language is Python, and the top-level layout includes .claude-plugin/, agent-plugins/, managed-agent-cookbooks/, mcp-servers/, scripts/ and vertical-plugins/.

How the 63 skills and 4 agents are organized

The skills are grouped into six vertical plugins rather than shipped as one undifferentiated bundle. china-finance holds 31 skills and is the core A-share research set: china-comps and china-comps-analysis for comparable valuation, china-dcf and china-dcf-model for discounted cash flow work, china-lbo-model, china-3-statement-model, china-earnings-analysis, china-earnings-preview, china-model-update, china-sector-overview, china-catalyst-calendar, china-idea-generation, china-thesis-tracker, china-morning-note, china-initiating-coverage, and a set of file and data utilities including china-deck-refresh, china-ib-check-deck, china-ppt-template-creator, china-pptx-author, china-xlsx-author, china-audit-xls, china-clean-data-xls, china-market-data, china-variance-commentary, china-accrual-schedule, china-break-trace, china-gl-recon, china-roll-forward, china-deal-screening and china-skill-creator.

The other five plugins are narrower. investment-banking has 10 skills covering pitch decks, merger models, CIMs, teasers, buyer lists, datapacks, process letters, strip profiles, deal tracking and competitive analysis. private-equity has 9, from DD checklists and IC memos through portfolio monitoring, returns analysis, deal sourcing, unit economics, value creation planning, AI readiness and DD meeting prep. wealth-management has 5, fund-admin has 6, and operations has 2 for KYC document parsing and a KYC rules engine.

Above the individual skills sit four end-to-end agents: china-pitch-agent chains valuation modeling into a roadshow deck, china-market-researcher moves from sector overview to competitive landscape to a target pool, china-earnings-reviewer runs from earnings interpretation through model update to a research report, and china-model-builder produces DCF, LBO or three-statement Excel output. The README presents skills and agents as usable independently, which matters if you only want the modeling piece and not the full pipeline.

The tiered data layer and where it can break

The data layer uses a four-tier priority scheme. Tier-0 is wind-mcp, described as paid Wind data spanning A-shares, Hong Kong and US equities, funds, indices, bonds, macro, research reports and analytics. Tier-1 is ifind-mcp, a paid 同花顺 iFind service. Tier-2 is akshare-mcp, the free open-source AkShare source covering quotes, financial statements, industries and indices, and the README states it is used as an automatic fallback when iFind is unavailable. Tier-3 is china-news-mcp, free, drawing on 财联社, 东方财富 and exchange announcements.

The design intent is that commercial sources take priority and degrade automatically to free ones so that some data is always available. That is a reasonable default, but the README does not document what the degradation looks like in practice. It does not say which fields AkShare can and cannot supply relative to Wind, whether a skill fails loudly or silently produces a thinner model, or how a user would detect that a run happened on Tier-2 rather than Tier-0. If you are producing a valuation that will be reviewed by someone else, that distinction matters more than the fallback's existence.

Only the paid tiers require credentials. Wind uses WIND_API_KEY, obtained from the Wind AI financial terminal, with keys beginning ak_. iFind uses IFIND_AUTH_TOKEN, a JWE token from the iFind MCP platform's key management page. AkShare and the news tier need no registration. The README also notes iFind service tiers with different concurrency limits: the free version covers stock, bond and index queries at 2 requests per second, the personal version unlocks fund, news and edb services at 5 per second, and the enterprise version raises concurrency to 10 per second.

Installing the plugin marketplace and running a first task

The README's one-minute setup assumes Claude Code is already present. The first command installs it globally through npm, the second registers this repository as a plugin marketplace, and the third installs the full china-finance bundle.

npm install -g @anthropic-ai/claude-code
claude plugin marketplace add jwangkun/claude-for-financial-services-cn
claude plugin install china-finance@claude-for-financial-services-cn

If you do not want all 31 core skills, the README lists per-vertical installs instead. These pull in the investment banking, private equity, wealth management or fund administration sets individually.

claude plugin install investment-banking@claude-for-financial-services-cn
claude plugin install private-equity@claude-for-financial-services-cn
claude plugin install wealth-management@claude-for-financial-services-cn
claude plugin install fund-admin@claude-for-financial-services-cn

After installation the README shows plain conversational prompts rather than a CLI invocation. The examples given are writing a 贵州茅台 annual report review in 中金 format, building a 比亚迪 DCF using the 中债 10-year rate for WACC, producing a semiconductor sector A-share strategy, and generating a roadshow deck. What you should see is Claude responding with output shaped by the installed skill rather than a generic answer.

If you want the iFind data source, it has its own setup separate from the plugin install. You install its Python dependencies, then supply the token either as an environment variable or in a config file that the repository's .gitignore excludes from commits.

cd mcp-servers/ifind-mcp
pip install -r requirements.txt
export IFIND_AUTH_TOKEN="your-jwe-token-here"

The alternative configuration route is a JSON file at mcp-servers/ifind-mcp/mcp_config.json with an auth_token key. The README also shows the same pattern for Wind, where the key can be exported as WIND_API_KEY or written to mcp-servers/wind-mcp/mcp_config.json.

What the README does not tell you

The most significant gap is verification. A skill that produces a DCF is only as good as the inputs it pulled, and the README gives no procedure for confirming that a given run used Wind rather than AkShare. There is no documented output marker, log line or metadata field that records which tier served a request. For exploratory work that is fine. For a model that feeds into a published report, the absence of that trace is a real operational problem.

Second, the paid data sources are not optional in the way the tier table implies. If you have neither WIND_API_KEY nor IFIND_AUTH_TOKEN, you are running entirely on AkShare and the news feed. The README describes AkShare's coverage as quotes, financial statements, industries and indices, which is narrower than Wind's stated scope of A-shares, Hong Kong and US equities, funds, indices, bonds, macro, research reports and analytics. Skills that depend on consensus estimates, institutional holdings or ESG data have no documented free-tier equivalent.

Third, the repository has no releases. The README's install commands reference a plugin marketplace rather than a versioned artifact, so there is no changelog to consult when a skill's output changes between runs. The last push was on 2026-09-14, three days before this writing, which means the codebase is moving, but without releases you have no pinned version to fall back to.

Finally, the README is in Chinese and the skill names are romanized. That is consistent with the target audience, but it means an English-speaking reviewer cannot audit what a skill does from the name alone and would need to read the skill definitions in vertical-plugins/.

How it differs from using the upstream Anthropic skills directly

The obvious alternative is Anthropic's own claude-for-financial-services, which this project is built on. The difference is not feature count but market fit. Upstream targets US and European workflows: Bloomberg and FactSet as data sources, US GAAP as the accounting basis, US Treasury yields as the risk-free rate, GICS for classification, and English-language bank report formats. If your coverage is US equities, upstream is the more direct choice and you avoid a translation layer entirely.

If your coverage is A-shares, upstream will produce output that needs correction at every convention boundary. You would be manually replacing the risk-free rate, reclassifying the comparable set, and reformatting the report to a domestic bank template. This project encodes those substitutions as skills, which is the whole point. The cost is that you inherit a dependency on domestic data providers and their key management, plus a Chinese-language documentation set.

A second alternative is to skip skills entirely and use AkShare directly in your own scripts. That gives you full control over the data path and no plugin layer, but you lose the report formats, the model templates and the agent chains. The choice comes down to whether your bottleneck is data access or output formatting. If it is data access, AkShare alone is simpler. If it is producing a 中金-format review from a 贵州茅台 filing, the skills are doing work you would otherwise write yourself.

Maintenance, licensing and upgrade cost

The repository is Apache-2.0, which permits commercial use, modification and redistribution provided you retain the licence and attribution notices. That is permissive enough for internal deployment at a financial institution, though the licence file itself is the authority and nothing here should be read as legal advice. Note that the licence covers this repository's code, not the data sources it connects to: Wind and iFind access is governed by your own agreements with those providers, and AkShare carries its own terms.

On maintenance, the last push was on 2026-09-14. There are no retrieved releases, so there is no version history to reason about. Upgrades arrive through the plugin marketplace rather than a package registry, which means `claude plugin marketplace add` and the install commands are the update path. Because skill behavior is defined in files under vertical-plugins/, an update can change output formatting without a version number to signal it. Teams that depend on a specific report format should keep a local copy of the skill definitions they rely on.

The data layer adds its own maintenance burden. iFind's tiered service levels have documented concurrency limits, so a workflow that runs many queries in parallel may hit the 2/s ceiling on the free tier and need a paid upgrade. Wind keys are obtained from a separate portal and expire or rotate on that provider's schedule, not this project's.

Editorial conclusion

Adopt it if you already work in A-share research, investment banking, private equity or fund administration and want Claude to produce output in 中金/中信/华泰 report formats with 申万 or 中信 industry classification and 中国会计准则. Do not adopt it if you need a general-purpose finance assistant for US or European markets, since the entire skill set is oriented around A-share conventions and domestic data sources. Before relying on it, verify which data tier you actually have access to: check whether WIND_API_KEY or IFIND_AUTH_TOKEN is set in your environment, and confirm that the AkShare fallback returns the fields your workflow needs, because the README describes the fallback as automatic but does not document which fields degrade when a paid tier is missing.

Frequently asked questions

Official sources

1. Issues
2. jwangkun/claude-for-financial-services-cn on GitHub
3. License: Apache-2.0
4. README