Skip to content
Help center

Skills

A skill is a step-by-step finance workflow. Copy it from Gatlas and paste it into your agent, which runs it against your datasets the same way each time.

Run a skill

You need Gatlas added to your agent first, with an agent key. If it is not, follow Connect your agent.

1

Open Skills

Open the Skills page. It is not in the sidebar, so open it from this link. Each card shows what the skill does, the datasets it needs under Required datasets, and whether it can run. Click a card to read when to use it and which of your datasets it will read.

2

Copy the prompt

Choose Ask your agent. The whole procedure is now on your clipboard.

3

Paste it with a company and period

Paste it into a new conversation and add what you want, for example:

Run this for Northwind for August.

Your agent asks for anything it still needs, such as the company or the month, then reads your datasets and writes the result.

Tip

Within one conversation you rarely need to paste the skill twice: asking "now do the same for last quarter" usually follows the same method.

Required datasets

A skill works on your datasets, not on raw data. Gatlas matches each name under Required datasets against your dataset names: a dataset whose name contains it, ignoring case, counts. The badge on the card tells you where you stand:

  • Ready: every required dataset exists, is approved, and has AI context.
  • Needs context: the datasets exist, but at least one is not approved or has no AI context. Approve any dataset that is not approved first (switch on Approved), because agents read only approved datasets. A dataset without AI context still works, and answers improve once you fill it in.
  • Missing datasets: at least one required dataset has no match. Ask your agent stays off until you create it.

So when you build the datasets a skill needs, give them names it recognizes, such as "Northwind Monthly P&L". Ask your agent to build them with you; see Datasets for the other ways.

The skills

Cards show each skill by its name. This is what each one does, and the dataset names it looks for.

finance-monthly-pnl

A monthly profit and loss statement for one company, with the change on the month before and on the same month last year, notes on the biggest movers, and a summary you can put in front of a board.

Looks for datasets named: Monthly P&L

finance-portfolio-rollup

Revenue, gross margin, operating result and net result across every company in the portfolio, with each company ranked by what it contributed.

Looks for datasets named: Monthly P&L

finance-variance-vs-prior

The P&L lines that moved most against the month before and the same month last year, ranked by impact, with a short note on what drove each.

Looks for datasets named: Monthly P&L, Monthly Trial Balance

finance-cashflow

A company's cash position over a window you choose (the last 90 days, week by week, unless you say otherwise), with month-end balances, net movement and a projection from the recent burn or build.

Looks for datasets named: Cash Position

finance-ar-aging

What customers owe, grouped by how long it has been outstanding (current, 1 to 30 days, 31 to 60, 61 to 90, over 90), with the customers who pay latest called out and the total owed.

Looks for datasets named: Invoices

finance-ap-aging

What the company owes suppliers, with the largest amounts called out and the total owed, grouped the same way where its data has due dates.

Looks for datasets named: Suppliers

finance-customer-concentration

How much of a company's revenue its largest customers bring in (the top 10 over the last 12 months, unless you say otherwise), where it depends on one customer, and whose share has shifted.

Looks for datasets named: Invoices

finance-vat-prep

Output VAT and input VAT for a period, by VAT code and rate, with the net amount due or refundable. It is a summary to check, not a filed return: you still verify it and file through your tax authority.

Looks for datasets named: VAT

dataset-builder

Builds a new dataset with you. Claude asks what you want in plain language, checks the idea against your real data, proposes the dataset, and creates it once you say yes, approved at once. It needs no datasets to start, and it needs a Read/write agent key because it creates something.

Limits

  • The finance skills only read. They never send reminders, pay a bill or change anything in your accounting.
  • A skill sees what your agent key sees. If the key is limited to some companies or some datasets, so is the skill.
  • A skill is only as good as the datasets under it. When a number looks wrong, check the dataset first, then its AI context.

Note

Anyone who can log in to Gatlas can copy a skill. Running one needs an agent key, which an admin creates on Agent access; see Agent access.