My AI Toolkit After a Year of Experiments: What Stuck and What I Dropped

I have a problem: I sign up for every AI tool that crosses my feed. Over the past year, that's meant dozens of free trials, a frightening number of browser extensions, and at least three "AI-powered" to-do apps.

Most of them lasted about a week. A few became part of how I work every day. Looking back, the difference between the two groups is surprisingly consistent.

What stuck

A general-purpose chat assistant

This is the one I use most, and it's also the least exciting to describe. I use it to think out loud: to untangle a confusing error message, to draft a tricky email, to explain a concept from a field I know nothing about, or to argue with me when I've made up my mind too quickly.

The key shift was treating it as a thinking partner, not an oracle. I don't ask "what's the answer?" as often as "what am I missing?"

An AI coding agent in my editor

This changed how I build side projects more than anything else. Not because it writes perfect code — it doesn't — but because it lowers the activation energy of starting. "I'll just set up the boilerplate" used to cost an evening. Now it costs ten minutes and a review.

I wrote about rebuilding this very site with one a few weeks ago.

Transcription

I record voice notes while walking — ideas, outlines, rants about bad UX. Accurate transcription turned a pile of audio I never listened to into searchable text I actually use. Several of the posts on this blog started as a ten-minute walk.

Image generation (occasionally)

I use it for rough concepts and placeholder art, not final work. It's great at "show me three moods for this landing page" and much less great at anything requiring consistency or precise text.

What I dropped

• AI note-taking apps that summarized everything automatically. I stopped reading the summaries, which meant I stopped reading the notes.
• Writing "enhancers" that rewrote my sentences to sound more professional. They made everything sound like the same person, and that person was a LinkedIn post.
• Autonomous "do everything" agents that promised to run my errands across the web. Fascinating demos. Terrifying in practice when they confidently book the wrong thing.
• Most browser extensions. Each one wanted to read every page I visited. That's a lot of trust for a tool I used twice.

The pattern behind what lasts

Looking at the list, the tools that survived all share three traits:

1. They do one thing I already did, but faster. I was already thinking, coding, recording notes. The AI accelerated an existing habit rather than inventing a new one.
2. I stay in control of the output. I review the code. I edit the draft. I pick the image. Tools that acted on my behalf without review didn't last.
3. Being wrong is cheap. A bad suggestion in a chat costs me five seconds. A bad action by an autonomous agent costs me an afternoon and an apology.

If you're evaluating a new AI tool, try asking: What habit does this speed up? Who checks the output? What happens when it's wrong? If you can't answer all three, it's probably a demo, not a tool.

I'll keep experimenting, obviously. But my toolkit is smaller and calmer than it was a year ago, and I'm getting more done.