Fluent in every tool, proficient in none. The real edge isn't a bigger stack — it's mastery of a small one.

I have a confession. I've been AI promiscuous.

It's part of my job — I write about this stuff, so I try everything. Every new agent, every new writing tool, every new presentation generator that promises to replace the last one. And for a while, that felt like an edge. More tools, more output, more leverage.

Then I looked at what I was actually shipping. A lot more content. A lot of it was AI slop.

The reason was simple: many of these tools didn’t understand me. They lacked context. They didn't have my voice. They didn't know my style guide, my edge cases, my non-negotiables. Every new tool was a cold start. I was paying the context tax over and over, and the output showed it — generic, off-tone, vaguely on-topic. The stuff that gets unsubscribes.

So I took a step back. I still try a lot of tools — but only to stay educated. For the work that actually goes out the door, I curated a tight stack. Five categories, and inside each one, one or two tools I've taught my voice to. That's the whole move. Fewer tools, deeper context, better output.

Here's what's actually in the stack:

1. Email — Shortwave and the Gmail web client (you could use Superhuman here too — I tried it and didn't love it, but plenty of people swear by it). I don't use AI to draft important emails (the voice cost is too high). I use it to sort, summarize, and triage. The average knowledge worker spends 28% of their work week in email. Cutting that in half is the single biggest leverage move most people are sleeping on.

2. Productivity — Notion first, Google Slides second. Notion is my source of truth — everything writes back to it. Plenty of smart people I respect use Obsidian for the same job, and it's a great solution if local-first markdown is your thing. I land on Notion because it's a source of truth plus it has Claude embedded right in the workspace — I can use Opus 4.7 as a thought partner while I'm writing, in the same surface where my notes live. One less context switch, one less cold start. I use the Gemini that's embedded in Google Workspace because it's already there and it's good enough for in-doc work. I stopped chasing presentation tools. For decks that matter, I use Manus or Claude Code with Claude Designand get a result that actually looks like me.

3. Notetakers — Fireflies, period. It's automated, it plays nicely with Claude and ChatGPT, and the search-video-by-text feature means I can find the 90 seconds I missed when Woodford the dog needed an emergency trip to his favorite bush. (He has one. He's committed to it.) Automated capture plus searchable transcript is the whole game.

I also wear the Plaud Pin at conferences and meetups — when I want to remember who I met and what we talked about without taking notes mid-handshake. Is it a little creepy? Sure. So is my robot notetaker showing up in a meeting. We're all already on the record; I'd rather have the recall.

4. Agents — Claude Cowork (on my desktop) and Manus. You can't swing a dead robot without hitting an agentic company right now. I've tried a lot of them. The two that earned a daily slot: Claude for the work I do at my desk, and Manus for tasks I want to hand to my assistant and collaborate on asynchronously. Low barrier to entry. Good results. That's the bar.

5. Vibecoding — Claude Code and Vercel. I know "vibecoding" still sounds nebulous to a lot of people, so here's what it means in practice for me: I use Claude Code and Vercel to maintain a handful of websites. Each site has its own skill in Claude with my branding, structure, and style codified. When I need to add a new event for a meetup, drop in a sponsor logo, or push a small content update, I open the project and just ask Claude to do it. Zero coding. Once the system is set up, what used to take hours takes a few minutes.

Bonus: If you are interested in vibecoding, I created a system and instructions for maintaining a website using an agent on Github.

Notice what's not on the list: the seventeen other tools I tried this quarter. They were educational. They were not part of the work.

The principle underneath this: every new tool is a context debt. You have to teach it your voice, your rules, your edge cases — or it will give you the average output of everyone else who uses it. That's the definition of slop. The fix isn't more tools. It's fewer tools, more context, repeated reps. Stick with a few tools, and you get real compounding gains from built-up memory and fine-tuning. It's almost impossible to fine-tune two dozen point solutions.

Fluency is being able to use a thing. Proficiency is being able accomplish real work more quickly and with better quality. Fluency scales linearly — every new tool costs you another learning curve, another set of quirks, another cold start. Proficiency compounds — the more reps you put on the same tool with the same context, the better the output gets, and the gap between you and someone who just opened the app for the first time widens every week.

The market wants you to be fluent. Every launch, every demo, every "this changes everything" thread is selling you fluency. Resist it. Pick the two or three tools that map to the work you actually ship, and put the reps in until they know you.

And here's where I think this is all going. The tool drawer is a temporary problem. In the long run, we're not going to have twelve apps open — we're going to have a handful of agents that consolidate the functions we currently spread across them. Better yet, those agents will build the software we need on the fly, in the moment we need it. The interface stops being "which app do I open?" and starts being "what do I want done?"

If that's where this is headed, the skill that compounds isn't tool fluency. It's knowing your work well enough — your voice, your rules, your edge cases, your standards — to direct an agent that can do anything. The people who win the next decade aren't the ones with the most tools in the drawer. They're the ones who can clearly articulate what "good" looks like, and have the context to back it up.

Fewer tools. Deeper context. Better direction. That's the stack that survives whatever comes next.