The blog

For you who prefer listening to reading, this post is also available as an episode of the ""Done!"" podcast:

Less small talk

Instead of chat­ting with AI, I build agents and so-called skills (and, beyond that, more or less advanced apps) that make things eas­i­er for me. The point of cre­at­ing an agent instead of con­tin­u­ing to write in long chat threads is:

• that you save ​“tokens” (which, sim­ply put, are what get con­sumed when you chat and may have a cost attached to them) because the AI reads the con­cise agent instruc­tion instead of the entire chat his­to­ry in the thread,
• that the results become more con­sis­tent between runs, with­out the AI grad­u­al­ly drift­ing in some direc­tion over time,
• that the agent is eas­i­er to dis­trib­ute to oth­ers who want to use it than if it were only a thread and an approx­i­mate prompt.

How do you get started?

How­ev­er, from con­ver­sa­tions with clients, I’ve real­ized that get­ting start­ed with build­ing agents can be a hur­dle. How do you begin? After all, you’ve only just got­ten good momen­tum with chat prompting!

That’s why I’d like to share a method that I’ve come to use with good results, and that has become my most com­mon way of build­ing agents, skills, and some­times even sim­ple apps.

It’s based on the AI describ­ing itself!

Do this

1. It all starts when you get an idea of some­thing the AI could help you with.
2. To see whether your assump­tion is cor­rect, you test it in a chat, just as usu­al. You upload a doc­u­ment and ask the AI to do some­thing with it. Or you let it find some­thing online, process it, and present it to you in a for­mat you’re hap­py not to cre­ate your­self. Or you prompt it about some­thing else you need help with.
3. When you see that the AI suc­ceeds in doing exact­ly what you want­ed, not just once but sev­er­al times, ask it to describe what you’ve just done togeth­er. The pur­pose is for you to be able to ​“tell” the agent so that it does the same thing.
   
   I usu­al­ly write a prompt such as: ​“I’d like to turn what we’ve done here into an agent. Can you describe what we’ve done as a prompt I can give the agent when I set it up?” or ​“Could you for­mu­late a skill based on what we’ve done togeth­er here?”
4. When you’ve received the descrip­tion, read it and look for whether the AI has mis­un­der­stood some­thing or failed to men­tion some­thing impor­tant. If you’re doing this in Microsoft Copi­lot, you’ll, in true Copi­lot fash­ion, be asked whether you want the AI to describe it in a com­plete­ly dif­fer­ent way or in a more advanced way than it just did. It’s up to you whether you accept.
5. Then, when you cre­ate the agent, it’s this descrip­tion that you paste into the box where you’re asked to describe the agent you want to cre­ate. If you’re cre­at­ing a skill in Claude instead, you’ll get a skill file that you dou­ble-click to install and put into operation.
6. Test the agent or your new skill care­ful­ly to ensure that it does what you want. Adjust the instruc­tions if you need to refine it.

Let me add that this approach is best suit­ed to sim­pler agents and skills. If you want to build some­thing more advanced that has mul­ti­ple data sources, where it’s crit­i­cal that the results are reli­able every time, or that will be used broad­ly across the orga­ni­za­tion, it may be bet­ter to build the agent with more direct hands-on involvement.

A more effi­cient and sta­ble solution

If you let the AI describe itself in order to quick­ly build an agent, you’ll get start­ed with agent cre­ation rapid­ly if you haven’t already. You for­mal­ize a good AI expe­ri­ence so that you can use it more often and with more con­sis­tent results. It’s like doc­u­ment­ing a well-func­tion­ing process you can show oth­ers, rather than doing the work auto­mat­i­cal­ly based on your exten­sive experience.

Your part of the busi­ness becomes a lit­tle more stable.

How do you do it?

What agent or skill have you built that you’re espe­cial­ly pleased with? Write to me and tell me, because I’m always look­ing for new ideas about what I can cre­ate myself so that I can get more work done more eas­i­ly dur­ing my workdays.

(Have you con­sid­ered that you can also use AI to help train oth­ers, such as a new colleague!)