Say there is something you do every week: sorting your to-do list on Monday, or writing the same kind of email. You want an AI skill for it, the kind of saved instruction set I described in Skills, then agents, then context. You could pick up someone else's skill from a marketplace, a shared library of skills other people have written, but it was built around somebody else's way of working, and you want yours to be as bespoke as possible. You could also ask the AI to write one for you, but you are never going to get it right in a single try. It is going to need a little bit of tweaking here and there.

Start with a stable source. By that I mean decades-old material: a framework or a workflow such as the Eisenhower matrix for sorting tasks (from a 1954 address) or George Orwell's rules for plain writing (from 1946). You need to already have an idea of what it is: something that exists and has social proof and business proof. You look at it and say, all right, the foundational parts of this do not change, and I want to create a skill from it that I can use in my work. Then find the Wikipedia page that describes it, give the AI the link, and tell it you need a skill for that.

Then talk it through. Unless you already have something really tightly defined and written down, just talk it through, in a chat inside a project that already has your history and context. If you use Claude, pick Opus for this.

Take it one step at a time. Have the AI give you questions, and answer them as they come. If you want it to compare other approaches as well, leave those as follow-up questions. Give it the link, see what it says, then ask what else could be explored. Otherwise it might come back with too much, or it might just make a decision on your behalf. You need to be part of the step-through to make sure you get what you are asking for.

Made-up example for illustration. The week and the AI's replies are invented. No AI is running on this page.

One question or all at once?

Ask an AI to build a skill that sorts a week of tasks: once in a single message, once with the AI asking you questions. Switch between the two to see which tasks end up in a different box.

You write one message.

Build me a skill called sort-my-week. It should sort my tasks into the four boxes of the Eisenhower method. Use the Wikipedia article below as the source. Do it all in one go and give me the finished skill. Then run it on my tasks for this week (list below).

Source: Time management, Eisenhower method (Wikipedia)

The week to sort

You run a small business. It is Thursday. These eight tasks are the same on both paths.

  1. File the VAT return. Due Friday.
  2. Reply to the supplier about the late delivery. They want an answer by Monday.
  3. Weekly team meeting. Friday, 9:00.
  4. Update the prices on the website.
  5. Plan next quarter.
  6. Send a client the invoice they asked for. They want it by Friday.
  7. Reply to a software seller. They want an answer by Friday.
  8. Tidy the shared folder.

Done. I built the sort-my-week skill from the article and ran it on your tasks.

Look inside the skill
  • Urgent means due within 48 hours.
  • Important means it has a deadline.
  • Emails and admin go to Delegate.
  • Anything without a date goes to Schedule.
  • Hand back four short lists.

Do now

Important and urgent. Do it now, yourself.

  • File the VAT return.
  • Weekly team meeting.
  • Send a client the invoice they asked for.

Schedule

Important, not urgent. Set an end date and do it yourself.

  • Update the prices on the website.
  • Plan next quarter.

Delegate

Urgent, not important. Hand it to someone else.

  • Reply to the supplier about the late delivery.
  • Reply to a software seller.
  • Tidy the shared folder.

Drop

Not urgent, not important. Take it off the list.

  • Nothing here.

The AI chose these rules. You were not asked about any of them.

Switch to "One question at a time" to sort the same tasks with your own answers.

This page sends nothing anywhere. Your answers are not saved or sent.

How to read this

There are two ways to ask for the same skill. In the first, you ask for the finished skill in one message. In the second, the AI asks you short questions and you answer each one.

Both end with the same eight tasks sorted into four boxes. Switch between the two paths to see which tasks end up in a different box.

Keep the first version broad. If you find you need a separate skill for every area of your work, the skill is too narrow. Cover everything first and add exceptions later, when you reach the weeds. Spell out your abbreviations as well, because the AI may not know what your shorthand stands for.

For a "sort my week" skill, start from Wikipedia's Time management article and its section on the Eisenhower method. It stems from a quote attributed to Dwight D. Eisenhower: "I have two kinds of problems, the urgent and the important. The urgent are not important, and the important are never urgent." The article sorts tasks into four quadrants. Important and urgent tasks are done immediately and personally, and important tasks that are not urgent get an end date and are done personally. Unimportant urgent tasks are delegated, and unimportant tasks that are not urgent are dropped.

Give the AI that page and ask for the skill. Before it sorts anything, it could ask what counts as urgent for you, then what counts as important, and only then whether you have anyone to delegate to.

For a "plain English email" skill, start from the six rules George Orwell set out in his 1946 essay Politics and the English Language, which the Orwell Foundation publishes in full. Two of them: "Never use a long word where a short one will do" and "If it is possible to cut a word out, always cut it out." Give the AI the essay and ask for the skill. Its first question could be who the email is for, and the second what you want them to do after reading it.

Pick a framework or workflow you already know that has not changed in decades, and paste a link to it into a chat. Ask the AI to ask you questions one at a time, and answer them. We can't assume the result is 100% correct, and hopefully it won't assume you are 100% correct either, so there are checks on both sides. Save the skill, test it on real work and tweak it.

Kit Library members can copy starter prompts for both skills from the members' shelf.

SkillsPromptingAnthropic