For most of the AI hype cycle, I didn’t really feel it.

I’m not a developer or an engineer. I work on business development, operations, product, and strategy. Of course, I was using ChatGPT and other LLMs every day. They were useful, but it didn’t feel like a breakthrough. I was moving a bit faster, but my workflows didn’t really change.

That shifted toward the end of 2025.

Not because AI suddenly got smarter, but because I started discovering the right tools and intentionally building them into my daily workflows. Small improvements. Quiet tools. Things that saved minutes instead of promising revolutions.

That’s when it started to compound.

Below are the tools and habits that actually stuck for me. This isn’t a universal recommendation list. It’s just what reduced friction in my day as a founder.

Granola AI

I’ve tried many meeting note-takers. Most of them feel intrusive. They announce themselves on calls, interrupt the flow, or feel like surveillance software with a UI.

Granola feels different.

It runs quietly, doesn’t disrupt the meeting, and produces notes that are actually usable. It handles multiple speakers well, switches between languages smoothly, and includes a clean text area where you can take your own notes during the call.

After the meeting, you can chat with the transcript like it’s a dataset: pull decisions, clarify context, or reconstruct what happened days later when your memory is already gone.

One important rule: always ask for consent and clearly tell people you’re recording. That’s non-negotiable.

Also worth mentioning: the free plan is generous enough to be genuinely useful.

MacWhisper

I use this alongside Granola, but for different reasons.

MacWhisper runs Whisper locally on your machine. It transcribes audio and video files, but the biggest value for me is dictation.

Typing is slow. Slower than most people want to admit.

Speaking is roughly 150–180 words per minute. Average typing speed is closer to 38–40 WPM. That’s a 3–4× difference, especially noticeable when writing long prompts, drafts, or messages.

I dictate rough thoughts, prompts, and half-formed ideas, then clean them up later.

The premium version is a one-time purchase (around $70). You choose the model, everything runs locally, and there’s no subscription.

Claude (Cowork / Agent Mode)

This is the first AI tool that felt like it was actually doing work instead of helping me do work.

I mostly use it for early-stage research and monotonous operational tasks: screening competitors, parsing messy inputs, and turning unstructured data into something I can think about.

It’s not perfect. It’s slower than Comet by Perplexity, and it can lag.

But for repetitive tasks I can do but shouldn’t be spending time on, it’s genuinely helpful. It reduced a lot of operational drag.

Claude Code is also interesting. I used it to prototype a basic financial tracking tool and a simple Telegram bot for group chats. Nothing production-grade, but enough to unblock ideas quickly.

That’s a separate topic.

Comet Browser

For years, Arc was my default browser, and most of my workflows were built around it.

The problem: Claude browser extensions don’t work in Arc. That’s how I ended up trying Comet.

So far, it’s better than I expected. Perplexity’s search feels faster than Claude’s desktop tools. You can choose which model to use, and while it’s less opinionated than Arc, it’s more practical for research and synthesis.

I’m still early here. Consider this a tentative yes.

Prompting Guides as .md Files

This sounds obvious, but few people actually do it.

At some point, we needed consistently high-quality images and videos for a startup. Real outputs, not prompt spam. We were using Flux and Kling AI, both of which have solid official prompting guides.

Instead of “learning prompting,” I copied their official guides into clean .md files and gave them directly to my LLM (or assigned the task to Claude Cowork).

That’s it.

Now I had a custom agent that writes prompts optimized for the exact model we use. No guessing. No vibes.

Bonus trick: you can do the same for writing. Take structured guidelines on detecting AI-generated text, turn them into constraints, and feed them to your LLM as instructions. It works better than most people expect.

Shortcuts

I ignored keyboard shortcuts for years. I thought I was fast enough.

I wasn’t.

A designer friend of mine works in Figma like he’s playing an instrument. He barely touches the mouse. He isn’t rushing. He just avoids unnecessary movement.

That was the lesson.

Whatever tools you use daily, learn the shortcuts you repeat every day. Not all of them. Just the ones that matter.

For me, that’s:

• macOS

• Chrome

• Notion

• Figma

Where I Landed

None of this made me 10× faster overnight.

What it did was remove friction in places where I was already spending time: meetings, notes, research, writing, and repetitive operational work.

I still make the decisions. I still think. I still judge.

I just get there with less drag and fewer context switches.

In 2026, the real advantage isn’t “using AI.”

It’s identifying which parts of your day never required judgment in the first place—and refusing to keep doing them manually.