How Anthropic’s Finance Skills Power Real‑World CFO Workflows

Blog workflow overview

The Anthropic blog describes a corporate‑finance analyst who prepares quarterly board decks, monthly financial reviews, ad‑hoc model work and cross‑team context extraction. For each workflow Claude performs a specific role:

Board deck (quarterly) – Claude Cowork aligns numbers across slides, checks narrative consistency and flags contradictions whenever a figure changes.

Monthly review (monthly) – Claude Cowork drafts the first‑version commentary in the company’s established voice, e.g. “Revenue is A vs B, down C% driven by D.” It can reference the prior month’s document to keep style consistent.

Financial model work (ad‑hoc) – Claude for Excel traces cross‑tab formula references, diagnoses model bugs and can summarise key drivers of a new model.

Cross‑team context (continuous) – Using Google Workspace / Slack connectors Claude extracts decisions and rationales from long discussions and documents.

These scenarios illustrate how Claude’s memory, system prompts and project isolation enable a single analyst to keep narrative and data in sync across many moving parts.

Claude Cowork Projects

A Cowork Project is a self‑contained workspace on the Claude desktop. Each project provides:

an independent folder with local files,

a persistent system‑prompt level instruction set,

project‑level memory that accumulates facts and decisions.

Projects can be created from scratch, imported from the Claude web UI, or linked to an existing local folder. The blog recommends separate projects for board decks and monthly reviews to isolate style, memory and context.

Financial‑services Skills repository

The anthropics/financial-services repository contains only Markdown SKILL.md files – no build step or runtime dependencies. The directory layout is:

plugins/
├── agent-plugins/        # 10 end‑to‑end agents, each self‑contained
├── vertical-plugins/    # 7 vertical industry packs; skill source files live here
│   ├── financial-analysis/   # core modelling (DCF, LBO, comps, 3‑statement)
│   ├── investment-banking/
│   ├── equity-research/
│   ├── private-equity/
│   ├── wealth-management/
│   ├── fund-admin/
│   └── operations/
└── partner-built/       # partner packs (LSEG, S&P Global)

Three reusable skill categories are identified:

Methodology / process skills – e.g. equity-research/skills/sector-overview generates a full‑length industry research report.

Excel modelling skills – DCF, LBO, three‑statement, comps and audit‑xls. They support a Python/openpyxl mode that runs in Claude Code or Cursor without Cowork.

QC / consistency‑check skills – e.g. financial-analysis/skills/ib-check-deck for PPT consistency, audit-xls for spreadsheet audit, deck-refresh for template‑driven deck updates.

Mounting Skills in Claude Code

Method A – Official plugin marketplace (recommended) :

# add marketplace
claude plugin marketplace add anthropics/financial-services

# install core modelling pack (required)
claude plugin install financial-analysis@claude-for-financial-services

# install vertical packs as needed
claude plugin install equity-research@claude-for-financial-services
claude plugin install investment-banking@claude-for-financial-services

After installation slash commands such as /dcf, /comps, /earnings, /sector and /ic-memo become available directly in Claude Code sessions.

Method B – Manual copy of SKILL.md :

# clone the repo
git clone https://github.com/anthropics/financial-services
cd financial-services

# copy a skill into your project’s .claude/skills directory
cp -r plugins/vertical-plugins/equity-research/skills/sector-overview \
    ~/your-project/.claude/skills/

Invoke the skill in Claude Code with @sector-overview.

Using Skills in Cursor

Cursor does not have native Skill support but can apply Rules. Copy the SKILL.md content into a .cursor/rules file:

# in the Cursor project root
mkdir -p .cursor/rules
cp /path/to/financial-services/plugins/vertical-plugins/equity-research/skills/sector-overview/SKILL.md \
    .cursor/rules/sector-overview.mdc

Cursor automatically applies the rule when the project is referenced.

Using Skills in GitHub Copilot

Copilot supports Custom Instructions. Place each Skill as a separate .github/copilot-instructions.md file and add an applyTo front‑matter to limit the scope:

# create instructions directory
mkdir -p .github/instructions

# copy the skill
cp /path/to/SKILL.md .github/instructions/sector-overview.instructions.md
---
applyTo: "**/*.md,**/*.xlsx"
---

Adapting the Skills for Chinese users

The default .mcp.json lists eleven overseas data sources (e.g., Daloopa, FactSet, Morningstar). These services are generally unavailable in China and must be replaced with domestic equivalents. The mapping is:

Listed‑company financials / valuation multiples – replace Daloopa / FactSet / S&P Kensho with Wind, iFinD or Choice (wrapped by a local MCP server).

Primary‑market / private‑equity transaction data – replace Pitchbook with CVSource, IT Juzi or Qingke.

Fund / asset‑management data – replace Morningstar with Tiantian Fund, Galaxy Securities fund research or Wind fund library.

Credit ratings – replace Moody's with China Chengxin, United Credit or Dongfang Jincheng.

Real‑time market data + Reuters news – replace LSEG (Refinitiv) with Financial Associated Press, Xinhua Finance or Wind real‑time data.

Financial news & earnings call transcripts – replace Mt Newswires / Aiera with Financial Associated Press, Giant‑Wave or Snowball call‑note archives.

PE fund operations – replace Chronograph with self‑built tools or Jiandaoyun.

Enterprise document management – replace Egnyte with Feishu, DingTalk or Enterprise WeChat document APIs.

US public‑company filings – replace SEC EDGAR with CNInfo (cninfo.com.cn) and Shanghai/Shenzhen exchange disclosure platforms.

Modification steps

Replace hard‑coded data sources in each SKILL.md. Example for comps-analysis/SKILL.md:

# Original
1. FIRST: Check for MCP data sources (S&P Kensho, FactSet, Daloopa)
2. ONLY if MCPs unavailable: Bloomberg Terminal, SEC EDGAR
3. NEVER use web search as primary source

# Chinese version
1. FIRST: Check Wind / iFinD / Choice availability
2. Fallback to 巨潮资讯网 (cninfo.com.cn) + exchange disclosure platforms
3. PROHIBIT using search engines as primary data source

Build a local MCP server that wraps domestic APIs (Wind, iFinD, etc.) so Claude can call them via the same endpoint URLs used in the original skills.

Localise terminology (e.g., comps → “可比公司分析”, LBO → “杠杆收购”).

Adjust accounting standards from US GAAP to Chinese Accounting Standards; IFRS to International Financial Reporting Standards where appropriate.

Quick‑start for individual users

For non‑production use, free data sources such as Tushare or AKShare can be wrapped in a minimal Python MCP server:

import akshare as ak
# expose ak.stock_financial_report_sina(stock="600519") as an MCP endpoint

The author notes a separate article will detail a full MCP server implementation.

Conclusion

The article extracts concrete, reusable Finance Skills from Anthropic’s open‑source repo, shows how to attach them to Claude Code, Cursor and GitHub Copilot without requiring Claude Cowork, and provides a detailed mapping of overseas data sources to Chinese alternatives. By following the step‑by‑step modification guide, Chinese finance teams can adopt Claude‑powered workflows with locally available data and terminology.