You Don't Need More AI Tools. You Need One You've Finished.

Look in your bookmarks bar, and there's a folder of AI tools you tried once. There's the meeting-notes one, and near it the one that was going to organise your research.

Each tool got an afternoon and worked about as well as promised, and then each lost its place to the next one.

None of them changed your week. Why not?

You paid with the afternoon, but the benefit never arrived. A tool you've tried isn't a tool you use, and a tool you used for a week isn't one you've finished.

Mia Kiraki writes a newsletter called Robots Ate My Homework, and she has the line for this. You can't scale a thought, but you can scale a well-documented process. In other words, an idea in your head can't grow, but a job written down step by step can be repeated and passed on. Her advice is to stop collecting tools and build one deep way of working for one real bottleneck, which is the slow job that holds everything else up. Give it thirty days, then write it up so a new hire could follow it.

The thirty-day rule is hers, and I think she's right.

Compare the folder with a bookshelf. A bookshelf has a fixed number of shelves, so at some point you decide what stays. The deciding is what makes the books a collection rather than a pile.

Folders have no shelves, so nothing in yours ever gets decided.

Left alone, the folder grows until it's useless. Nothing else will set a limit, so the limit has to come from you. Kiraki calls the person who sets that limit a curator, and the person who doesn't a hoarder.

What comes out of thirty days on one bottleneck? A process, which is the written-down way you now do that one job, with the AI inside it. You can hand the process to someone or reuse it next year.

Thirty days on one slow job

Which job is the bottleneck? It's the one that's slow every week, and that you'd be relieved to have off your plate.

The bottleneck is rarely glamorous. Think of the weekly report, the quote follow-ups or the inbox on Monday morning. If you can't tell which job it is, put the question to the AI.

Here's a rough picture of my working week. [describe it, or paste your calendar]. I want to pick one recurring job and build a proper way of doing it with AI, over thirty days. Help me find the bottleneck. That's the thing that's slow every week, that I dread slightly, and that AI could plausibly help with. Rank three options by how much time they'd free. Tell me which one you'd pick and why. Don't suggest tools yet.

Use the AI tool you already have. Claude or ChatGPT, whichever's open, will do. A new tool means a new afternoon lost to settings, and the point is the process.

Use the process every time the job comes up, for thirty days. Write one line after each go, saying what worked, what you had to fix and what you'd change. By day ten you'll see a shape.

By day twenty, a saved set of instructions, such as a Skill or a Project, will be doing most of the job. By day thirty you'll know whether the process is worth keeping.

Some bottlenecks turn out not to be a job for AI. Maybe the report is slow because the figures are in six places, and no AI fixes that. If by day ten the AI isn't the slow part, stop, and count that as the lesson, because you've found the real problem.

At the end, write the process up on one page, as if for a new hire. Cover what the job is, what to give the AI, what to check, and what never to let the AI do.

If a colleague could follow the page, the process is finished, but if they'd need you in the room, it isn't finished yet.

Every line on that page came from a go you checked, so the page is yours to hand on.

Isn't trying lots of tools how you find the right one? People who tried twenty tools know things the people who committed early don't, and the tool your bottleneck needs might be one you haven't met.

That's fair. So give it a week of looking and then a month of doing, rather than a year of looking.

When the month's done, open the old folder. Most of what's in it answers a question you no longer have.

This week, leave the folder shut. Paste the bottleneck prompt into whatever's already open.

With thanks to Mia Kiraki. The line "you can't scale a thought, you can scale a well-documented process", the curator and hoarder idea and the thirty-day rule are all hers. They come from "To go fast with AI, first go slow" on Robots Ate My Homework (September 2025).