The work after the work
Most practitioners finish a project and move on. The deliverable goes to the client or the filing authority. The process that produced it — not the specific reasoning, but the way you worked through the problem — lives in one person's head. Next time someone in the firm faces a similar question, they don't just lack the answer. They lack the process. They have to figure out how to think through the problem before they can even start thinking through it.
Even if you sat down and explained your process to a colleague, they'd still have to work out how to execute it themselves. That's the gap. And it's a waste of everything you just learned.
The reflective question — what worked, what didn't, what do we do differently — is the mechanism that converts a one-off project into institutional knowledge. But here's what changes everything: when you ask those questions to the AI, it doesn't just answer them. It builds the instruction. It writes the process down in a form that guides the next person through the same methodology — step by step, with the same rigour, the same phase gates, the same adversarial stance. The AI doesn't just capture knowledge. It operationalises it.
Instructions vs. skills — two ways to encode what you learn
In Claude Cowork, there are two concrete ways to capture this knowledge. The first is an instruction file — a document that lives in a specific folder and shapes how AI operates in that context. When I open my tax research folder, Claude already knows my jurisdiction preferences, my documentation standards, and the analytical frameworks I use. I didn't configure that in a settings menu. I wrote it down in a file, in plain language, and the AI reads it every time.
The second is a skill — a portable methodology that's available across every session, regardless of which folder I'm working in. The research dialogue process we built from the cross-border project is a skill. It encodes the three-phase approach (research and challenge first, architecture second, drafting last), the adversarial stance, the loose ends tracking, the research brief that builds as we go. Anyone with that skill gets the same structured process I developed through trial and error — without the trial and error.
The distinction matters. Instructions are context — they tell AI where it is and how to behave here. Skills are process — they tell AI how to do a specific type of work, anywhere. Both are just text files. Both compound over time. And both flip the relationship between practitioner and tool. The AI isn't waiting for you to tell it what to do. It's guiding you through a proven process — your process, encoded once, executed consistently every time.
This is context engineering — and it's your competitive moat
Every instruction file you write, every skill you build, every process you document in a way that AI can follow — you're building structured knowledge that makes your practice more capable. Not just you personally. Your team. The person you hire next year who needs to do the work you're doing today.
This is context engineering in practice. A firm that has encoded its research methodology, its client onboarding process, and its advisory frameworks into reusable instructions and skills doesn't just use AI faster — it uses AI better. Every project makes the next one sharper. Every instruction file reduces the gap between what your best person knows and what everyone else can do. That's the rebuild.
The question that changes everything
After your next significant piece of work — the next tax research question, the next advisory recommendation, the next regulatory interpretation — stop before you move on. Ask the three questions. Write down what you learn. Turn it into an instruction or a skill that someone else can use.
You'll be surprised how quickly the knowledge adds up. And you'll be more surprised by what happens when you hand that skill to a colleague and they produce work at a level that used to require your direct involvement.
That's not a productivity gain. That's a practice asset. And it started with getting the process out of your head.
The research dialogue skill we built is included in Your Practice, Automated — the guide and implementation video course for using Claude Cowork in CAS workflows. It walks you through how to capture methodology as portable instructions, build skills your team can reuse, and encode institutional knowledge in a form that AI can execute consistently. Get the Kindle edition on Amazon, or visit theaiaccountant.com to get both the book and the free 5-day implementation video course. The skills you build today compound into a practice asset that works for your firm tomorrow.
Tomorrow I'm going to push this further. Every time you tell AI it's wrong — every correction, every pushback, every "that's not how this works" — you're exercising a skill that the profession has never developed deliberately. And it might be the most valuable thing you do with AI all week.
If you want a copy of the research dialogue skill we built from this process, DM me.