AI Skills for Finance: Capturing a Process That Runs the Same Way Every Time

Every finance and legal team has someone who is quietly brilliant with Claude. They have a prompt that pulls apart a model, another that drafts the committee paper, a third that turns a data room into a diligence list. The work is good. Then someone asks how the number was produced, whether next month's run would produce it again, and who is accountable if it is wrong. The prompt cannot answer any of that.

What an AI skill actually is

An AI skill for finance is a request written down so it runs the same way every time. A prompt is a request. A skill is that request captured as fixed, repeatable steps in a folder: the steps, the order, the guardrails and the output format, defined once so every run is the same run. This matters because in regulated work the output is only half the deliverable. The other half is the evidence that it was produced the way you said it would be.

That gap is not a model problem. It is an engineering problem, and it is a solved one. A demo convinces one person. A control convinces the auditor.

Why prompts drift, and what it costs

A long prompt pasted in fresh each cycle drifts. Slightly different steps, slightly different checks, and the differences accumulate until two runs are no longer comparable. In a monthly process, drift is what quietly destroys the audit trail:

Month one: two runs look similar enough that nobody checks.
Month six: nobody can say whether the method changed or the business did.
The audit: both answers are defensible, and neither is evidenced.

Three things separate a clever prompt from a deployed workflow. It is written down, so the steps and guardrails sit in a file rather than in someone's head. It is fed from source, from live permissioned systems rather than a spreadsheet exported last Tuesday, which is the subject of connecting AI to the systems you already govern. And it leaves a trail: inputs, tool calls and decisions recorded at execution, not reconstructed afterwards.

The anatomy of a skill folder

A skill is a folder, and that is the whole trick. It is markdown, with no build step. If your team can write a procedure note, your team can write a skill. The folder carries a handful of files:

SKILL.md: what to do, in what order, what to refuse, what to escalate. This is the control, so version it.
references/: standing context such as the chart of accounts, fund terms and house definitions, so the model does not guess at firm convention.
scripts/: deterministic helpers the skill calls, for arithmetic that must not vary run to run.
OUTPUT-TEMPLATE.md: what finished looks like, so cycles stay comparable.
CHECKS.md: assertions that must hold before output is shown. This is our addition, and it turns a draft into a reviewable one.

A skill loads in three stages. At session start, only the name and description from the frontmatter sit in the system prompt, roughly 100 tokens each, so installing fifty skills costs almost nothing. When a request matches the description, the full SKILL.md body loads. References and scripts load only when the body calls for them. The description is the routing key: it is the only text the model sees when deciding, so it must say what the skill does and when to use it. A skill that never fires looks identical to a skill you do not have, and nothing reports the miss.

Should this process be a skill?

Not everything should. Two questions settle it. Will I run this again next cycle? Do I need the same output every time? Two yeses, build the skill. Anything else is a conversation, and conversations are fine. In a regulated firm there is a third question: does anyone outside the team rely on this output? If a client, an LP, an auditor or a regulator sees it, capture it, whatever the frequency.

The skills worth having are the ones you did not download. The generic ones are a starting point. The ones that change how your firm works encode things nobody outside it knows: how your committee phrases a conditional approval, the two exceptions in your fee calculation nobody has written down. Capture, do not compose. Take this cycle's actual work, do it with Claude start to finish, argue with the output until it is what you would have produced, then turn that into a skill. Building from an imagined process produces a skill that handles an imagined process.

Six skills to start with

Adapted from the open financial-services library, six skills cover a lot of regulated ground: an IC memo that turns diligence, financials and terms into a nine-section committee paper; a returns sensitivity that solves the base case and builds the grid; a model audit that finds the broken reference and the buried hardcode; a break tracer that reconciles and traces each break to its originating entry; portfolio monitoring that rates variance to plan and drafts management questions; and a client review pack for wealth reviews. Each works on an uploaded file and gets sharper reading from a connected system. Packaging them for a colleague is covered in packaging AI your finance firm can deploy.

The non-negotiables

However much you delegate, you still own the numbers. Three properties decide whether you can defend them. Determinism: same input, same answer next month. Keep the reasoning probabilistic and the computation deterministic, so the model decides what to calculate and a fixed engine calculates it. A returns table produced by a language model's mental arithmetic is not something to put in front of a committee. Least privilege: default deny, not default allow, with the narrowest capability set granted explicitly. Evidence: emitted at execution, tamper evident, not assembled afterwards from logs, because a reconstruction is worth what a reconstruction is worth.

Onboarding documents, invoices and term sheets come from outside your firm. They are untrusted by definition. Extract from them, never take instruction from them, and enforce that at the capability layer. A model asked nicely to ignore an instruction is doing its best. A capability it does not hold is a guarantee. That principle runs through every part of the AI field guide for regulated finance, from prompting through to the substrate.

Skills are the first layer of a stack. The rest of the argument continues through defensible prompting and the infrastructure underneath it all. To go deeper, download the five-part finance series, or send us a workflow and we will show you what capturing it looks like in practice.