Just Type What You Want

Every week someone posts a new framework for working with AI. There are lots of them: spec-driven development, role-based multi-agent orchestration, RAG, prompt engineering, context engineering, MCP server stacks, sophisticated memory systems.

Meanwhile, the most useful thing I did with Claude this week was type a sentence describing my problem, read the answer, type another sentence, and stop when I had what I needed.

We’ve already seen this pattern recently.

We already did this with Zettelkasten

Niklas Luhmann was a German sociologist who produced an enormous volume of work throughout his career (over 70 books and 400 papers). He used a wooden box and slip cards. When he had a thought, he wrote it on a card, gave it a number, and put it in the box. He called it Zettelkasten.

Around 2020, the productivity community discovered Luhmann and his Zettelkasten and decided that it was his slip-box that was the secret to his productivity. Immediately, an ecosystem arose: Roam Research, Obsidian, Logseq, Tana, Capacities. Each one promised new features, better UI, and better productivity. People built elaborate vaults with them — daily notes, atomic notes, literature notes, permanent notes, evergreen notes, backlinks, maps of content, graph views, plugins for spaced repetition, custom CSS, templates for templates. Books were written on taking notes.

The thing is, the people with the most elaborate vaults had written the least. The majority of “digital gardens” I’ve seen are now note graveyards.

Now we’re doing it again with AI

The AI discourse is on the same arc. A genuinely useful tool arrives. Almost immediately, the energy moves to building things around the tool instead of using it. And the same realization is going to happen, probably faster: the people with the most sophisticated AI workflows are getting less done than the people who just type what they want.

The pattern under both is the same. If you give someone a powerful tool, they will reliably build systems around the tool to avoid the vulnerability of using it. Using a tool means producing an output, which means being judged. Building systems around the tool feels like work, looks like work, generates content about the work, and protects you from being judged. It’s sophisticated procrastination.

I’ve fallen for both versions of the pattern. I used Obsidian for CBT for half a year. It definitely helped me fight depression through behavioral activation. Not to mention, I was able to expose some core dysfunctional beliefs and patterns. After that, it started to become productivity theater: vault tweaking, meta-notes on taking notes and psychology. All instead of doing the thing: changing the behavior. It was the same with AI. I started simply, just by chatting with an LLM. Surprisingly, sometime later, I started to get decent results. Then I decided that I wasn’t doing it right because everyone and their dog was arguing about prompt engineering, orchestration, subagents, etc. I went down this rabbit hole and, at some point, found myself doing a lot of meta-work instead of doing the work.

To be fair: these AI workflows do provide value for large engineering teams shipping production systems, where reproducibility and orchestration aren’t optional. What I’m saying is that the majority of solo users can simply get by typing what they want into a text box.

What I do now

The actual method, for almost everyone, almost always: ask for what you want. If the answer isn’t right, say what’s wrong with it. Repeat. Paste a screenshot. Then repeat again. Stop when you have what you need. That’s the whole thing. Notes get written by writing notes. Software gets built by asking the AI to build it.

Luhmann had a wooden box.

You just type what you want.