LLM agents make it possible to automate so much more than before. The reason for this is that they add a level of flexibility possible that wasn't possible before. The way I think about this is as a pyramid. When you want to automate a digital task, think about it in these levels:
- Prompt
- Skill
- Product
First level: Prompt
When you first want to automate something, just ask the agent. Fire up Codex or Claude Code and ask it to do the thing you want. This is really useful for one-offs. For example, "I have this folder of documents, give them all titles based on the contents". I use that prompt after I scan paper documents I received. iOS names these "Scanned documents" and prompting is faster than thinking of a name. And the name isn't that important other than that it should be good enough to recognize what is in it. But this kind of document I probably only read once.
Second level: Skill
You reach this level when you notice yourself writing the same prompt over and over again. A great example of this is Matt Pocock's grilling skill. Instead of asking "Ask me questions about this", you invoke the skill with /grill-me and the result will be a little bit more rigorous and consistently formatted. You don't have to use his skill, you can make your own with your own preferred format. The main idea here is that you don't have to specify all your preferences every time.
Third level: Product
The moment you have a consistent process that needs to be handled deterministically, this is the place to go. Build your own private-first product. This is the level where you get nice UIs, good visualizations and consistency. The product will always work the same way, assuming you spend some time hammering out the bugs. Some of my own examples are: Checklister, an app I made to really quickly create repeatable checklist templates. 4Daagse tijden, an app to see at what time to spectate the 4Daagse. And OmniConverter, an app to quickly convert quantities. They are extremely personal products, but creating them is way cheaper than before LLMs.
Bringing it together
I've seen a lot of people intuitively navigate these levels already. But I've found the mental model helpful to categorise and structure my own thinking of how to automate digital processes. May it help you too.