AI Skills for Business Procedures: 7 Steps to 4 Minutes

One of our clients onboarded every new customer onto their platform by hand: seven steps, seven different owners. We turned it into a skill an AI agent runs in four minutes. Here is how to do the same with one of your own repetitive procedures, starting from a checklist you already have.

What It Is

A skill is a procedure written in a file that a person can read and an agent can run. It is not a long prompt saved in a document. It has a name, a clear input, steps and a result you can check. In Claude Code, for example, a skill is a folder with an instructions file (SKILL.md) the agent loads when it needs it.

For the team it is a procedure. For the agent it is code. For the company it is the same thing seen from two sides.

Why It Matters

The client showed us how they set up a new tenant, meaning a new customer's space inside their software. Someone created the environment. Someone else configured DNS (the web address). Then the CRM, billing, the message to the team, the first deploy and the welcome email. Each step had a different owner, and each owner a different day. If one stopped, the chain waited.

AI was not what was missing. A fully written process was. It lived in chat, in checklists and in two people's heads.

As a skill, it runs in four minutes, with the same chain and no skipped steps. Across eighty real onboardings: zero errors, 94% less operational time and a full trail in git (the tool that records every change to a file). Tickets for wrong DNS, wrong billing plans or welcome emails sent to an internal address disappeared. They were rare, but expensive.

The team was not replaced. They stopped doing the work they hated and got wrong out of boredom. The agent is not smarter than the team: it is more consistent.

How To Do It

1. Pick a process that repeats every week and today lives in a checklist. The test: if you can explain it to a new colleague in ten minutes and they can run it, it is a skill. If every time "it depends", decide first and automate second.
2. Answer four questions on one page. Who runs it today, by name, and who covers when they are away. Which system is the source of truth for each step (one, not two). Which result a person can tick off in ten seconds. How wrong it can go before it gets expensive. If the answers do not fit on one page, they will not fit in a skill.
3. Write the steps, each with its own stop. Every step says what to do and when to stop. Example below.
4. Remove fake exceptions. Either they are rules, and become a step, or they are noise.
5. Run the first ten by hand following the new procedure, with no agent. If those ten do not come out the same, the agent will not save you: it will multiply the mess.
6. Then let the agent run it. Review everything at first. Once results are stable, review only the exceptions.
7. Store the skill in git, not in a shared spreadsheet. Every change to the procedure becomes a revision with an author and a date. If March onboarding differs from April's, you can see why. When someone asks "who created this tenant?", the answer is a log, not a meeting.

Here are the seven steps from our case, each with its stop:

• Environment: create it with the tenant name and permissions. If a field is missing, stop. Do not invent.
• DNS: create the record, wait, verify. If it does not resolve, do not continue, and the error stays in the log.
• CRM: record, owner, status. No duplicates "just in case".
• Billing: plan, start date, internal reference. If the plan does not exist, stop.
• Team notice: who needs to know, which channel, which link. Once.
• First deploy: the agreed version, not whatever was left on yesterday's machine.
• Welcome email: text stored in the skill, not pasted from a six-month-old thread.

Ready to copy, to have your agent write the first version of the skill:

Small Test

Ten minutes. Take the checklist of a procedure you run at least once a week and paste it into the prompt. Read the skill that comes out.

It worked if every step has a clear stop and the agent asked you at least one question you could not answer on the spot. That question is the weak point of your procedure, with or without AI.

Watch Out For

• It does not work for judgement calls. A negotiation, a credit decision, a conflict between two teams. There the agent can prepare the file, but it should not close it.
• It works where the outcome is yes or no and the systems have an API (a way to talk to other software) or at least one step that can be automated. Onboarding, offboarding, opening a store, a staging environment, a welcome kit, renewing a certificate.
• Exceptions need a channel, not a skill that swallows them. A ticket, an owner. The same exception three times: it becomes a step. Once: a person handles it.
• If exceptions keep growing, AI is not the problem. Usually someone is selling too many variants. Fix it there.
• The numbers come from one case. Four minutes and 94% less time apply to that onboarding. Measure yours: time, errors and exceptions per week.

Next step: if you are not sure which procedure to pick, start with the 5-step AI audit. To make the agent remember exceptions it has already solved, read how to give it a memory that sticks. If you have a procedure that lives in seven chats today, we can build it with you: the code stays yours.

In short: A procedure you repeat every week can become a skill: readable by a person, run by an agent, stored in git, with a stop point at every step.