Set up AI to work your way

You ask AI to prepare a weekly staff update. You explain the format, what belongs in it and which decisions still need your approval. Next week, you explain it again.

The instructions you repeat are a useful starting point for a reusable setup. Begin with a task you understand well enough to judge.

Consider a fictional local shop. Its weekly update covers confirmed changes, staff assignments and questions awaiting a decision. The owner wants it brief, with customer details left out.

Some expectations stay the same each week. Dates, assignments and unresolved questions change. Save the stable guidance, then supply the current facts for each run. A well-written procedure cannot make last week's information current.

Anthropic's engineering article on Agent Skills describes packaging instructions and supporting resources so Claude can load them when relevant. It also recommends starting with representative tasks, observing where the agent struggles and improving the guidance from those observations. This is technical guidance, rather than a study proving time savings for business owners.

You can work out your instructions in an ordinary document before choosing a feature. A Claude Skill is one way to package them; the current Skills documentation explains the mechanism. Setup differs by product.

For the fictional shop, a starting instruction could be:

Prepare a short staff update from these notes. Separate confirmed changes from proposals awaiting approval. Include an action owner and date only when supplied. Flag missing details, and leave customer information out.

Now give that instruction something worth checking. Try one fictional set of notes with complete details, then another where a delivery change is only proposed and has no start date.

Compare the drafts with the notes. Does the second draft keep the proposal unconfirmed? Does it flag the missing date? Can the reviewer see which decision is still theirs? A polished announcement that invents Monday as the start date has failed this task, even if it reads beautifully.

Keep the first result, revise the guidance and try a different example. That gives you a small record of what improved and what still needs attention. One successful test supports that example; it does not establish reliability across future situations.

Human review remains part of the design. In its February 2026 study of agent autonomy, Anthropic observed that more experienced Claude Code users both allowed more automatic actions and interrupted more often. The study uses data from Anthropic's own products and warns that important findings may not transfer beyond software work. It does not tell us how much autonomy a local shop should grant.

Our practical takeaway is to decide where review matters in your task. Who checks the facts? Who approves a schedule change? Producing a draft, running it every Friday and sending it to staff are separate steps that need their own setup and checks.

Choose one recurring task this week. Save its stable instructions, identify the facts that change and try an incomplete example. Keep the correction you would want the system to remember.

A question to consider: Which correction do you keep making when AI helps with your work?

Subscribe to Tongdao Note for practical reflections and a place to reply with your questions. An unfinished attempt is welcome.

Practise with your own work

If you would like help building and checking instructions for a recurring task, private AI coaching and tailored team training can use the tools you already work with.