The skill that builds all my finance skills

I'm giving away the most-used file in my entire library. But... what does Margot Robbie have to do with it?

Hey Finance Engineers,

This issue starts something new. Until now I’ve been testing other people’s releases: Anthropic’s agents, Claude Fable. Today we start building your skill stack, level by level.

Level 1 is skills.

In the next 10 minutes you’ll learn what a skill actually is, how to build financial skills that survive contact with a real close, and you’ll walk away with the master file: the skill.md that builds all my other financial skills.

Let’s go.

First, what a skill actually is

You’ve used Claude in chat. Maybe Claude in Excel, Cowork, or Claude Code. And I can bet money on this: you’ve had the same prompt produce three different outputs on three different days.

That inconsistency is the tax you pay for working from prompts.

A skill is how you stop paying it.

In The Big Short, Margot Robbie explained subprime mortgages from a bubble bath in 30 seconds. I’m not Margot Robbie. And I don’t have a bath either. But I’ll try to be just as precise.

A skill is an SOP (Skope of Work) your model follows. It tells Claude exactly how to produce a specific output: the steps, the format, what to check, what to refuse. Every time. In your format. That’s it. That’s the whole concept.

Or even shorter: a prompt is you explaining the task again every morning. A skill is the task explained once, permanently, by someone who was paying attention.

This is the first real game-changer for your time and your mental health. And instead of handing you a library of someone else’s skills you’d still have to customize, I’m going to teach you to build your own.

Then give you the file that does it with you. How about that?

The problem nobody warns you about

When you start building financial skills, the first one is fun. The fifth is fine. By the fifteenth, you end up with a mess.

I know because I opened our skills folder last month and audited it. Every skill structured differently. Half of them with dollar thresholds hardcoded into the prompt. One file says “company profile,” another says “business context,” a third says “client info.” Same file. Three names. And when something breaks during close, you can’t tell if it’s the data or the skill, because there’s no standard to check against.

So here’s the rule I now live by, and I’ll say it the blunt way: a skill stuffed with hardcoded values is not a skill. It’s a script.

A script works for one project, one month, one threshold. A skill survives a new month, a new policy, a new client, without anyone touching the skill body. If your “skill” breaks when the threshold changes, it was never a skill.

The fix: a meta-skill that builds skills

We built a meta-skill. One md file, dropped into your Claude skills folder. From then on it runs in two modes.

BUILD mode. You say “build a skill,” and it interviews you before writing a single line. What’s the one responsibility? (Exactly one. Not two.) What feeds data into it, and what does it feed into next? Which context files does it read? What can’t be automated yet? Only after the interview does it produce the full SKILL.md, in a fixed 10-section template: role, scope, context, tools, input, process, escalation rules, output schema, what it does NOT do, and critical rules.

VALIDATE mode. You paste any existing SKILL.md and it audits the thing line by line. Hardcoded values, missing sections, banned terminology, weak escalation logic. You get a pass/fail report with exact quotes of every violation and the fix for each.

Build with it, then validate its own output. The skill checks its own homework.

3 rules worth stealing even if you never download my skill file (below)

1. The FM approval gate is a hard block. Every skill has a step where the agent stops and waits for explicit human confirmation. Not “human review recommended.” The agent literally cannot continue without it. In finance, this is the difference between an agent you can deploy and an agent you can demo.

2. Confidence scoring with deduction logic. The agent starts at 1.0 and deducts for specific issues: 0.1 per missing data field, 0.1 per unresolved source discrepancy, 0.1 per unclassified vendor. Below 0.7? It escalates instead of deciding. No vibes-based confidence. Remember: every downstream skill is betting on this one’s output being right, so the errors don’t add up, they multiply.

3. Never let the LLM “just calculate.” Every calculation follows the same pattern: find the formula, show it, build the logic, calculate, return formula plus inputs plus result. If the formula is missing, the agent escalates. It never estimates. This one rule kills 90% of the hallucinated-numbers problem.

If you take nothing else from this issue, take these three. Paste them into every skill you ever write. I have.

How to use it

1. Download the SKILL.md below

2. Add it to your Claude skills (Settings → Capabilities → Skills, or drop it into your Claude Code skills folder)

3. Say “build a skill”. Or paste your messiest existing skill and say “validate this”

Start with validate. Watching it tear apart a skill you thought was solid is the fastest way to internalize the standard.

Watch it work: building a revenue reconciliation skill in 8 steps

I ran the full loop so you can see exactly what the interview feels like. The task: a skill that reconciles Stripe/Chargebee revenue against QBO GL for the close month.

The skill asks you eight questions before it writes a single line. These eight questions ARE the method. Even if you never download the file, run your next skill through them by hand and it will come out better. Here’s each one, and why it’s there.

Step 1. What is the one responsibility of this skill? Exactly one. Not two. The moment a skill does two things, you can’t tell which one broke. Mine: reconcile Stripe/Chargebee revenue against QBO GL for the close month, flag mismatches.

Step 2. Which orchestrator does it belong to? Every skill lives inside a bigger workflow: Month-End Close, Controlling, Forecasting. Naming the orchestrator up front means you’re building a system, not a pile of disconnected tools. Mine: Month-End Close.

Step 3. What feeds data into it, and what does it feed into next? The data chain. Mine receives from data-ingestion-agent and passes to revenue-recognition-agent. If you can’t answer this question, you don’t have a skill yet, you have a wish.

Step 4. Which context files does it read? This is where hardcoded values go to die. The thresholds, the CoA mapping, the business specifics all live in named context files.

Step 5. Which connectors does it need? Only what’s actually required, with read/write spelled out. Mine: QBO (read), Chargebee (read). A skill that asks for write access it doesn’t need is a skill your future self will regret.

Step 6. Which steps can’t be automated yet? Anything still manual becomes an explicit [MANUAL STEP] block instead of a silent assumption. Mine: the FM confirms the Chargebee export covers the full month before the skill runs.

Step 7. Who owns it? Every skill has a named owner. A name. When it breaks during close, you know exactly whose Slack lights up.

Step 8. Has it been tested on a real close month? This sets the status field, and I love this part: Idea → Testing (1 month) → Validated (2+ months) → Standardized (team standard). A skill that’s never touched real data is an Idea, and the file makes you admit it.

Eight answers later, it produced the full SKILL.md: the role, the scope, the critical rule (”Never modify accounting records without a human approval. Never guess missing financial data. If confidence < 0.7 or risk is high → escalate.”).

Then I did the part I recommend you do first: I validated it. And to make it interesting, I planted two bugs: a hardcoded $500/1% threshold and a missing connector.

The validator caught both. 28 of 30 checks passed, and the two failures came back with exact quotes and exact fixes. The missing kpi_context surfaced as four separate violations, which is what happens with a real gap in real life: it doesn’t fail once, it leaks into everything downstream.

That’s the loop. Interview, build, validate, fix, deploy. No blank page anywhere in it. Your next skill takes twenty minutes, and it’s better than the ones you spent hours on.

My meta-skill file (built and tested by ~20 finance managers)

Here it is, the most-used file in our entire skill library. The skill that builds all the other skills.