Don't Build AI Skills, Build Systems
First published on LinkedIn, 22 October 2026.
A skill is a routine written in plain text so that an AI tool such as Claude or ChatGPT can follow. If you repeat a workflow, like a tax pack review, you can probably turn it into a skill. But a skill that works is much more than a just written-up process. In this article I use a small skill I built to book calls to show what that means. The lessons apply to any tax process you want to hand to AI.
Since I started my own tax technology consulting business, a lot of new conversations begin with “let’s have a call”. My first idea was simple: a skill in Claude that reads my calendars and suggests a call time. I quickly realised that getting the AI to find a random gap in my calendar wasn’t enough.
1. Embed priorities
When there’s more than one gap in my calendar, I set the following preferences:
- stack calls 5 to 10 minutes apart
- afternoons over mornings, because 09:00 to 13:30 is protected for writing and coding
- where possible, no calls on the afternoon before travel or a big event
Left unstated, the skill would spread calls randomly and before I know it my week would break into multiple half-hours, which ruins my deep work time.
2. Build tolerance for edge cases
An exception came up quickly. A contact was in Singapore, and the skill proposed a call at midnight her time. So I added a rule: first establish where the invitees are, then where our working hours overlap, and only then apply my preferences. It taught me that the order of the checks matters as much as the checks themselves.
3. Enrich context before acting
An introduction call and a product walkthrough need different lengths, pre-reading and research. The skill now checks what the call is for before applying my preferences. The preparation itself runs as a separate routine, on purpose, so that each part stays small enough to check.
4. Learn from patterns and iterate
A dozen uses in, two more patterns emerged. Bookings land faster when I propose a few slots first, instead of asking for availability outright. And my booking link (shan.tax/book) is almost never used when options come first. The link still goes in alongside two proposed slots, in case neither works for the other person.
5. Leave room for flexibility
More chats with enterprise teams made me realise that people who live in Teams aren’t keen on a Google Meet link, and vice versa, so the skill now checks which they might prefer. And because the skill was built in layers, it is easy to add extensions or bring in connectors.
Beyond the skill
What I have now is a small system and the basis of my own CRM: a booking skill, a separate preparation routine, and a contact record with preferences. The skill is the building block, and good prompt writing is what the block is made of.
The same habit can sit under a tax process: a filing cycle, a provision process or a cross-border reporting review. Take a tax pack review. What would your priorities be? Materiality, or which entities and balances get attention first? Order matters: confirm each entity’s status, jurisdiction and filing deadline before reviewing a single number. Context sets the depth: a dormant company needs a lighter review than a trading subsidiary with intercompany balances, or one where a significant change took place in the last year. And the points reviewers raise each year should feed back into the next iteration of the system.
AI tools can do more and more of the low-level work. That does not make the work unimportant, and it does not allow us to be complacent about it. What becomes more important is setting priorities, finding the edge cases with solutions, and reviewing what actually runs. I think getting these right is undervalued, and it may be a harder job than the one many of us are used to in tax.
Check out the free Prompting Techniques for People in Tax, the layer under any skill.