Inside Anthropic's Finance Team: What 150 Claude Skills Looks Like in Practice

Anthropic's finance team uses approximately 150 shared Claude Skills to automate reconciliations, financial analysis, reporting, and governance workflows. These skills are stored in a version-controlled repository, shared across the organisation, and maintained by finance professionals rather than IT teams.

In a June 2026 webinar, Tim Ross (Finance AI Product Lead) and Lisa To (Head of Finance Systems and Transformation Office) gave a rare inside look at how this works in production — and what any finance team can take from it.

Key Takeaways

• Anthropic's finance team has built approximately 150 shared Claude skills, stored in a GitHub repo and distributed as workspace plugins available to every finance team member

• Domain experts — accountants and FP&A analysts — write the skills themselves, not IT

• Skills encode institutional knowledge that compounds: build once, version it, share it, and it runs the same way for everyone every time

• CFO Connect community members have independently arrived at the same model, confirming this is not unique to Anthropic

• Human review is still non-negotiable: Anthropic's own team is explicit that full auditability is still a work in progress

• You do not need a large team or a technical background to start — the skill builder inside Claude guides you through construction in plain language

What Anthropic's finance team automates with Claude Skills

Anthropic's finance team uses Claude Skills across five core workflow categories.

Reconciliation. AR sub-ledger to GL matching runs automatically: Claude loads both files, matches at invoice level, builds a variance bridge, and produces a structured exception report — formatted the way a controller expects to review it.

FP&A. A revenue movers skill connects directly to Anthropic's BigQuery data warehouse, decomposes month-over-month revenue changes by source and sales segment, surfaces the largest customer movements, and outputs a leadership-ready slide deck.

ERP migration. During Anthropic's move from NetSuite to Workday Financials, Claude validated historical data in batch — reducing what would have been hundreds of hours of manual checking to approximately 20 seconds per year of data.

Master data governance. A conversational Claude interface handles change requests for cost centres, account names, and other master data — capturing the required fields, routing for approval, and creating a complete audit trail automatically.

Scheduled reporting. Daily revenue briefs and weekly reports run on a set schedule with no prompt required. The output lands before anyone asks for it.

1. What is a Claude Skill and why does it matter for finance?

A Claude Skill is the difference between a prompt you use once and institutional knowledge your whole finance team inherits.

Most finance teams using Claude today have individual prompts: saved somewhere, used by one person, and lost when that person moves on. A skill is a procedure encoded once in plain language — what the inputs are, what the output should look like, what the calculation rules are, what to flag when something is missing — and then reusable indefinitely by anyone in the organisation with access to the workspace.

Tim Ross explained the organisational logic during the Anthropic webinar: "Your best preparer writes the procedure down once in plain language and it runs the same way for everyone, every time. This knowledge gets versioned, not lost when somebody moves on."

That framing matters. Skills are not just automation. They are institutional memory in a form that compounds.

According to McKinsey research, approximately 30% of finance tasks are automatable with current technology (McKinsey, 2025). The gap between that potential and what most teams have captured is not a tool gap. It is an adoption and knowledge-encoding gap. Skills are what close it.

What to do this quarter

Make a list of every procedure in your finance team that exists only in someone's head or in a prompt saved to one person's desktop. Each one is a skill waiting to be encoded — and a business risk as long as it isn't.

2. How does Anthropic actually organise 150 skills across a finance team?

The organisational model is more important than any individual skill: a versioned GitHub repository, distributed as workspace plugins, written by domain experts.

Anthropic's finance team stores their skills in a shared GitHub repository using the Claude workspace knowledge plugin framework. When a skill is updated, it versions. When a new team member joins, they have access to the full library from day one. The skills show up automatically in Claude Cowork and Claude.ai — no manual setup per user.

Two categories of skills sit in the library. The first are foundational skills: how to query specific tables in Anthropic's data warehouse, what column names mean, how the GL is structured. These exist so non-technical finance team members — who may not know SQL — can run complex data warehouse queries by describing what they want in plain language.

The second are workflow skills: the step-by-step procedures for specific recurring tasks. The AR reconciliation skill Tim demonstrated encodes exactly what an accountant would do manually: load the two files, match by invoice ID, classify the breaks, build the variance bridge, write the exception summary.

Tim's team also makes this library accessible via Claude Code — meaning the same skills finance users access in Cowork are available to anyone in the org building finance tooling.

What to do this quarter

Start with two or three skills that cover your most repeated monthly processes. Share them with your team via Claude's workspace plugin settings. The infrastructure exists — you do not need a custom build.

👉 Read Anthropic's documentation on Claude Skills and plugins

3. What does Anthropic's finance team actually use Claude skills for?

Two live demos from the webinar — one deterministic, one analytical — map to the two categories of skills every finance team needs.

The deterministic skill: AR sub-ledger to GL reconciliation

Tim demonstrated a skill that reconciles accounts receivable against the general ledger. The inputs are two Excel files: open invoices at the granular invoice level, and the AR aging detail from the GL. The skill tells Claude what the key columns mean, how to match records, and how to classify exceptions.

Running live against synthetic data, Claude loaded both files, matched at invoice level, built a variance bridge, and produced a structured exception report. The output: a $33k discrepancy identified — two items in the GL not in billing, three in billing not in the GL, one amount mismatch. The exceptions were formatted the way a controller would expect to review them.

For a deeper look at what best-practice reconciliation looks like before you automate it, read Spendesk's guide to expense reconciliation.

The analytical skill: Revenue movers

The second skill connects to Anthropic's BigQuery data warehouse. It tells Claude how to query the monthly revenue table, decompose month-over-month changes by revenue source and sales segment, and surface the individual customers with the largest movement. The foundational skill layer is what enables this: because Claude has been taught the table schema in plain language, a non-technical FP&A analyst can run a complex warehouse query without writing SQL.

The output: a full revenue bridge from January to February, drillable by segment, with a follow-up prompt that produced a Google Slides deck — ready to share with leadership.

What to do this quarter

Identify one reconciliation your team runs every month. Document the inputs, the matching logic, and the expected exception format. Use that documentation as the basis for your first skill.

4. What has the CFO Connect community independently discovered about the skills model?

Four CFO Connect practitioners — building with Claude entirely independently of Anthropic — have arrived at identical conclusions. This is not coincidence.

Domain experts write the skills, not IT

At Anthropic, the accountants doing the reconciliations write the reconciliation skills. Tim was explicit on this: the domain expertise flows through the human, not the tool. Christian Sanford, co-founder of QuantFi — which runs fully outsourced finance functions for companies between $5m and $200m — describes the same approach: "You do not need to understand the code. You need to understand the finance well enough to know when the output is wrong."

Use Claude to write your skills

Anthropic uses a "skill builder skill" — Claude walks you through construction in plain language and produces a correctly formatted skill from your description. Christian Sanford arrived at the identical method independently: "I go to Chat first. I say: help me create a prompt for Cowork to do this reconciliation. Then I take that into Cowork." The meta-skill is using Claude to improve how you instruct Claude.

Skills compound when scheduled

Anthropic's Tim Ross described the scheduling model: "You can set a schedule — weekly reports, daily digest, recurring reviews. No prompt required. It sets and forgets and runs every day for you." Pauline Babell, CFO of Spendesk, independently built the same operating model with MCP-connected workflows: "With MCP, you build it, then schedule it, then automate it. That is the compounding effect." Her travel compliance workflow now runs automatically on the fifth of every month.

For a full breakdown of how Spendesk's CFO uses Claude and MCP connectors to automate recurring finance workflows, read the MCP and Claude for Finance playbook. For a primer on how MCP works and what it means for your finance stack, read What is MCP and what does it mean for finance teams?

Parallel-run before you trust

Alex Altman, who automated revenue recognition at Coram AI — reducing a four-to-six hour monthly process to a single button click — ran his Claude workflow in parallel with the manual process for two to three months before relying on it. "It cannot be wrong. People are making business decisions with your information." This matches Anthropic's own position: human review is still non-negotiable.

5. How do you prevent Claude from making errors in finance skills?

The fix is built-in verification at the skill level — not more manual checking after the fact.

Gartner's 2025 AI in Finance Survey finds that 63% of finance AI initiatives fail to meet expectations (Gartner, 2025). The most common reason: outputs that look correct but are not, and review processes that are too slow to catch errors efficiently. Both problems are solved at the skill-design stage, not the review stage.

Tim Ross's five-step model for how Claude Cowork runs a skill is worth holding in mind when building: understand the ask, plan the steps, execute across files and tools, verify the output internally, deliver the finished result. That fourth step — Claude checking its own output — is only as strong as what you have built into the skill's checking layer.

Two design principles that emerge from both the Anthropic webinar and CFO Connect practitioner experience:

Require exception flags. Every skill that produces a financial output should explicitly identify rows where the source cannot be traced, where amounts do not match, or where data is missing. Do not leave exception identification to the reviewer.

Require a reconciliation layer. For any skill producing journal entries or variance bridges, include a checking tab that confirms debits equal credits and flags any line that breaks. Christian Sanford's framing: "If you're having to manually do all the checks yourself, it defeats the value of the execution."

What to do this quarter

Add this to every skill instruction: "Include a checking tab. Flag all exceptions and missing data. Cite the data source for each output row." Measure how much it reduces review time over one close cycle.

6. What does a finance skills library do to your team's capacity?

The value of a skills library is not the first workflow. It is what the fourth and fifth cost to build, and what they free up.

Gartner's 2026 CFO Technology Survey finds that 75% of CFOs are raising technology budgets, 88% rank finance staff productivity as a top-three priority, and headcount growth has collapsed from 6% to 2% (Gartner, 2026). The output expectation is rising. The people to deliver it are not multiplying at the same rate.

Pauline Babell's results from building a Spendesk-connected skills library illustrate the compounding model concretely: travel compliance checking reduced from four hours to three minutes and thirty seconds; budget versus actuals from a week-long cycle to four minutes; AP aging from several hours to seven minutes and twenty seconds.

Individual workflows become team workflows when they are encoded as shared skills. A reconciliation skill built by one controller and shared to six team members does not save two hours. It saves twelve. And it does not degrade when the person who built it goes on holiday.

Gartner projects that by 2029, AI could unlock 10 margin points of growth for finance leaders who move beyond individual experimentation into structured automation (Gartner, 2026). The teams that arrive there are not the ones who run the most experiments. They are the ones who build the infrastructure — skills, connectors, schedules — and let it compound.

What to do this quarter

After building your first skill, document the time saving explicitly. Present it to your CFO or controller as the unit economics of the skills model: this is what one skill costs to build and what it returns per cycle. That framing is what gets the rest of the library prioritised.

7. What else is Anthropic's finance team using Claude for beyond reconciliation?

Two additional use cases from Lisa To's session represent the frontier of what a mature finance AI practice looks like.

ERP migration and historical data validation

Anthropic recently migrated their ERP from NetSuite to Workday Financials. Lisa To's team used Claude to build an application that validates historical data during migration — either uploaded manually or batched on a schedule. The time saving: one year's worth of data validated in approximately 20 seconds, compared to hundreds of hours of manual checking. "We can cover a lot more data set as you might even have capacity to do," Lisa said. "And then you sample and validate as much as you want."

For any finance team mid-migration or planning one, this represents a completely different approach to historical data validation — not a sampling strategy, but near-complete coverage at minimal cost.

Master data governance

The second use case is a Claude-powered chat interface for master data change requests. Previously, finance team members submitted Jira tickets — often with incomplete fields and poor user experience — and finance systems staff processed them manually. Now the conversational interface walks the requestor through exactly what information is required, captures it correctly, and routes it for approval automatically. The outcome: faster resolution, better data quality, a more complete audit trail, and less burden on the finance systems team.

FAQ: Building a Claude skills library for finance

What is the difference between a Claude prompt and a Claude Skill? A prompt is a one-time instruction. A skill is a versioned, shareable procedure that encodes your organisation's workflows, definitions, and standards — and runs the same way for every team member who uses it.

Do you need a technical background to build Claude Skills? No. Anthropic's own finance team uses the skill builder — a Claude tool that walks you through construction in plain language. Alternatively, describe your workflow to Claude Chat and ask it to format it as a skill. The domain expertise is yours; Claude handles the structure.

How many skills should a finance team build before skills become useful? One. The organisational value compounds with scale, but the unit-level time saving starts with the first repeatable workflow you encode. Start with your most painful monthly process and build from there.

Is Claude reliable enough for financial output without human review? Not yet without human oversight. Anthropic's own team is explicit: "You need to verify everything that comes out and take ownership of what comes out of it." Skills reduce the manual execution burden. Human sign-off before anything touches the ledger remains essential in 2026.

What is the best way to share skills across a finance team? Claude's workspace knowledge plugin framework lets you distribute skills to every team member's Claude Cowork and Claude.ai interface automatically. Anthropic uses a GitHub repository as the version-control layer. For most finance teams, starting with the plugin framework directly is sufficient.

How does this connect to the broader finance automation stack? Skills work best when Claude is connected to your data sources via connectors. The skills library defines the procedures; the connectors provide the live data. For a full look at how Spendesk's CFO uses those connectors with Claude, read the MCP and Claude for Finance playbook.

How long does it take to build a Claude Skill? A first skill for a well-documented workflow typically takes one to two hours: writing the procedure in plain language, running it through the skill builder, and testing it against a known historical period. Skills for more complex analytical workflows — those that query a data warehouse or decompose multi-dimensional variances — take longer to tune but follow the same process. The skill builder inside Claude accelerates construction significantly.

What finance processes should be automated first? Start where the hours are. The most common starting points across CFO Connect's community are: month-end reconciliations, variance commentary, AP aging reports, and recurring compliance checks. If you find yourself answering the same question every month, or running the same matching exercise every close, that is a skill candidate. Avoid starting with processes that are poorly documented or inconsistently applied — encode the standard first, then automate it.

How are Claude Skills version controlled? Anthropic's finance team uses a GitHub repository as the version-control layer. When a skill is updated, the previous version is preserved. For teams without a GitHub workflow, Claude's workspace plugin framework handles distribution automatically — but adding a simple version number and change log to the skill description provides a lightweight audit trail without requiring a full engineering setup.

Can Claude Skills connect directly to ERP systems? Yes, with the right connectors. Claude supports direct ERP connections via MCP for supported platforms, which works well for targeted lookups and small-scale actions. For large-scale financial analysis — running variance bridges, decomposing revenue, validating historical data — Anthropic's own team recommends routing through a data warehouse rather than querying the ERP directly, both for performance and data stability reasons.

Closing thought: the 150-skill library is not where you start — it is where you arrive

Anthropic's finance team did not build 150 shared skills in a sprint. They built one, tested it, versioned it, shared it, and repeated. The GitHub repository and plugin framework were not planned infrastructure — they emerged as the natural response to a skills library worth sharing.

The CFO Connect practitioners who have gone furthest with Claude — Pauline, Kevin, Alex, Christian, Sherilyn — built their strongest workflows the same way: one skill at a time, tested against known historical data, reviewed by someone who understands when the output is wrong.

That is the only viable model for AI in finance. Not because the technology requires it, but because finance requires it.

The teams that build one shared skill this quarter will save hours. The teams that build ten over the next year will have restructured how their function operates. The ones who wait will inherit whatever structure the faster movers decided.

Sources: Anthropic Webinar — How Finance Teams Use Claude Cowork, June 2026 | Gartner 2025 AI in Finance Survey | McKinsey — How Finance Teams Are Putting AI to Work Today | PwC CFO Pulse Survey | Anthropic Claude Documentation

For the complete Claude for finance toolkit — how Chat, Cowork, and Code work together — read the Claude for Finance Teams playbook. For the exact prompts to build your first skill, see the 25 Claude prompts for finance teams.