Now part of Bank CSV to QuickBooks & Xero Pro. From Bank CSV to QuickBooks & Xero Pro v1.1.0, the same four skills ship inside it (in the finance-analyst/ folder) at no extra cost, alongside its main skill. If you already bought this pack, your download keeps working and stays available.
Four Agent Skills that turn Claude into a careful junior finance analyst. Each skill comes with Python scripts that do the arithmetic, so your numbers come from code you can rerun, not from the model's head.
| Skill | What it does | You get |
|---|---|---|
messy-sheet-cleaner |
Fixes ugly bank, accounting and CRM exports: title rows above the header, (1,234.00) and 1.234,56 amounts, Cr/Dr columns, mixed date formats, total rows, duplicates. It checks the cleaned total against the file's own total row. |
_clean.csv, _clean.xlsx, a change log |
ledger-reconciler |
Bank vs books (or Stripe/Razorpay payouts vs invoices, GL vs subledger). Five matching passes, including date windows, fuzzy descriptions and split payments. Every open item gets a likely reason, and a proof line shows the difference is fully explained. | reconciliation.xlsx (Summary, Matched, Open in each side), summary |
budget-variance-analyst |
Budget vs actual by line and month, with favourable/unfavourable logic for income vs costs, materiality thresholds, YTD and plain-English commentary for management. | variance.xlsx, chart, commentary |
month-end-finance-pack |
From one transaction list and a chart of accounts: monthly P&L, direct-method cash flow, KPIs (margins, burn, runway) and tie-out checks. | finance_pack.xlsx with native Excel charts, 3 PNG charts, summary |
Built for founders, freelancers, small-business owners, bookkeepers and accountants who already use Claude and are tired of re-explaining the same spreadsheet chores.
Try it
Upload your files in a chat with the skills installed, then ask things like: - "Clean up this bank export and give me a proper CSV." - "Reconcile my September bank statement against the ledger and tell me what's unmatched." - "Compare Q2 actuals to budget and write the variance commentary for the board." - "Build the month-end pack from these transactions: P&L, cash flow and KPIs."
Or open samples/ and ask Claude to run any skill on the sample files first, so you see exactly what you get.
AI disclosure and limits
- The scripts compute; Claude explains. If a file has a layout the scripts can't read, Claude tells you instead of guessing. Every report shows the checks it ran.
- The month-end pack is cash-basis from your transaction list. It's not an audited statement and doesn't replace an accountant's accruals.
- Currency-agnostic. The samples use INR, and the US/EU number and date formats are covered in the tests.
- Your data stays wherever Claude runs: your computer for Claude Code, or Claude's sandbox. The scripts make no network calls.
- Not accounting, tax or audit advice.
- Claude writes the explanations and commentary; that text can be wrong. Check it before it goes to a client, a bank or a board.
- Not tested: every bank or accounting export layout in the world, or Claude's commentary quality on your files.
What you need
- A Claude plan with Skills and code execution turned on, or Claude Code, or the Claude API.
- For Claude Code on your computer: Python 3.9+ with pandas and openpyxl (matplotlib optional, for charts).
- Your exports as CSV or XLSX.
5-minute setup
- Unzip the pack.
- In claude.ai, open Settings > Capabilities. Turn on code execution.
- Open Customize > Skills. Click +, then Upload a skill.
- Upload each zip in
claude-ai-uploads/. One zip is one skill. - Check the skills are on. Start a chat, attach your files and ask in plain words.
- Not sure it works? Try a sample first: ask Claude to run
budget-variance-analyston the files insamples/.
Using Claude Code or the API instead? See Install below.
Install
You need Python 3.9+ with pandas and openpyxl wherever the scripts run. Claude's code-execution sandbox already has them. matplotlib is optional: without it you still get every table and report, just no PNG charts.
Claude Code (CLI or desktop)
Option A: copy the skills (simplest)
unzip oarsman-finance-analyst-pack.zip
cp -r oarsman-finance-analyst-pack/skills/* ~/.claude/skills/ # all your projects
# or, for one project only: cp -r oarsman-finance-analyst-pack/skills/* .claude/skills/
Start a new session and ask for the task in plain words (see "Try it" below). Claude picks the right skill by itself.
Option B: install as a plugin (one command to update later)
unzip oarsman-finance-analyst-pack.zip
claude plugin marketplace add ./oarsman-finance-analyst-pack
claude plugin install oarsman-finance-analyst-pack@oarsman-finance-analyst
In plugin mode the skills are namespaced, e.g. /oarsman-finance-analyst-pack:<skill-name>, and they still trigger automatically.
Claude.ai (web, desktop app)
- Turn on code execution: Settings > Capabilities > Code execution and file creation (in the newer layout it's called "Create and edit files").
- Open Customize > Skills, click +, choose Create skill, then Upload a skill.
- Upload each zip in
claude-ai-uploads/. Each zip holds one skill, so repeat for every skill in the pack. - Make sure the skills are toggled on. Then upload your files in a chat and ask in plain words.
On Team or Enterprise plans, an owner has to allow Skills and code execution under Organization settings > Plugins & skills first.
Claude API (Agent Skills)
Upload each skill folder with the Skills API (POST /v1/skills, beta header skills-2025-10-02) and attach it in the container of a Messages request that has the code-execution tool enabled. The API sandbox has no internet access and no pip install. Every script in this pack uses only the standard library plus pandas and openpyxl, which are preinstalled.
Other agents that read SKILL.md
The folders follow the open Agent Skills format (agentskills.io), so tools that support it can load them too: copy skills/* into that tool's skills folder. Only Claude has been tested.
Check it works on your machine
cd oarsman-finance-analyst-pack
python tests/run_tests.py
This reruns every script on the bundled sample data and checks the numbers. You should see 29/29 checks passed. The full record of our own run is in TESTS.md, and the files it produced are in expected/.
What's in the folder
skills/budget-variance-analyst/ SKILL.md, scripts/variance.py
skills/ledger-reconciler/ SKILL.md, scripts/reconcile.py
skills/messy-sheet-cleaner/ SKILL.md, scripts/clean_table.py
skills/month-end-finance-pack/ SKILL.md, assets/chart_of_accounts_template.csv, scripts/build_pack.py
samples/ sample data used by the tests and for trying the skills
tests/ run_tests.py, validate_skills.py, fixtures/
expected/ the exact files our test run produced
TESTS.md test record: every check, its result, and excerpts of real outputs
.claude-plugin/ plugin.json + marketplace.json for Claude Code
Licence and support
One buyer, one business (yours or the client you're working for). You can change the skills and deliver the outputs to clients, but you can't resell or share the pack. Full terms are in LICENSE.txt.
Support is self-serve: this README and TROUBLESHOOTING.md (fixes and FAQ). There is no call, chat or email support. Fixes ship as new versions at your download link; CHANGELOG.md lists every change.
Built by Oarsman AI, with AI-assisted development. Every number in the examples comes from a real test run (see TESTS.md). Not affiliated with Anthropic. Claude is a trademark of Anthropic, PBC.