I keep running into the same problem when I talk to finance teams about AI. They’ve got access to Claude, ChatGPT, or Copilot. They’re using it. But every time someone opens a new chat, they’re starting from scratch — re-explaining the same context, the same standards, the same way the firm does things.
That’s not a model problem. That’s a context problem. And there’s a surprisingly simple fix for it.
The AI Amnesia Problem
AI tools are intelligent. But they’re amnesiac. Every session starts at zero. The model has no idea your team uses a specific materiality threshold, that your revenue recognition process follows a particular internal memo, or that escalations above $500K go straight to the partner.
So what do people do? They paste the same context into every chat. They build a library of saved prompts. They re-explain things constantly. It works, but it doesn’t scale — and it definitely doesn’t survive someone leaving the firm.
What a Skill File Actually Is
In October 2025, Anthropic formalized something called Agent Skills — a standard for encoding that context into a portable, readable markdown file. A SKILL.md file is just structured text. No code. No deployment. No IT department required.
Here’s a simple example of what one looks like for a revenue recognition review:
- Identify the five-step model under ASC 606
- Flag variable consideration and constraint risks
- Note any series guidance applicability
- Reference the firm’s internal policy memo from Q3 2024
- Escalate to partner if contract value exceeds $500K
The AI agent reads that file before starting work. Every time. It doesn’t matter which session it is or who opened the chat — the context is there from the start.
This Is a Knowledge Management Problem, Not a Tech Problem
When I was at PwC, one of the workarounds people used was saving their best prompts so they could paste them in at the start of a new session. That was the state of the art — a copy-paste library of context. Skill files are the natural evolution of that idea, just made portable, versionable, and shareable across a team.
The people who should be writing skills aren’t engineers. They’re your senior reviewers, your technical accounting leads, your tax managers — the people who have done the work a thousand times and know exactly how it should be done. They’ve always had this knowledge. Now there’s a format that makes it operational at machine scale.
Two Things This Means for Accounting and Finance
1. Your expertise is encodable at machine scale
A senior accountant who has run 500 audit procedures or lease reviews can write that methodology into a skill file once. From that point forward, every AI session starts with that expertise already loaded. No re-explanation. No drift. Consistent output every time.
2. Institutional knowledge risk is real — and now addressable
Firms lose knowledge when people leave. That’s always been true. But AI makes it worse if your best people have been quietly building expertise inside personal Custom GPTs and saved prompts that live in their individual accounts. When they go, that goes with them.
Skill files flip that dynamic. They’re readable, auditable, and institutionally owned — the same governance mindset you apply to workpapers and procedure documentation. You can review them, update them, and hand them off. That’s a meaningful shift.
Where to Start
The practical starting point is simple. Find the workflow where your team keeps re-explaining the same context to an AI tool at the start of every session. That friction is your first skill waiting to be written.
Write it down in plain text. Give it structure — steps, context, firm-specific standards. Put it somewhere the team can access it. Then iterate. The skill gets better through use, and each improvement compounds across every future session.
AI models are getting smarter every quarter. Your institutional knowledge isn’t getting any more accessible on its own. Closing that gap doesn’t require a software project — it requires a markdown file.
Key Takeaways
- AI tools start every session from zero — skill files fix that by encoding your firm’s context once and applying it every time
- A SKILL.md is plain markdown text — no code or engineering required, domain experts can write them
- The most durable skills encode organization-specific methodology — your thresholds, procedures, and standards — not generic tasks the model will eventually learn on its own
- Treat skills like institutional documents: version them, own them, and don’t let them walk out when someone leaves
- Start with one workflow where your team keeps re-explaining the same context — that’s your first skill
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply