My AI stack right now. What's yours?

Four tools, four jobs: my default, my second opinion, my build tool, and where each one bites.

Paul Cobb
August 03, 2026

Most operators are still hunting for the one best AI. I stopped, and started giving four of them different jobs. Here's the stack I actually run right now, and where each one still bites.

1. Claude, by Anthropic. My default, and now my whole way in.

Most of my work starts here. I run Opus 5 as the default for the thinking work: underwriting conversations, pressure-testing whether a deal actually pencils, pulling insight out of a pile of numbers. When the job is heavier data crunching, I drop to Sonnet 5 to get it done faster and cheaper. But the real shift this quarter isn't the model, it's the interface. Cowork is now my default way of working, not a chat box I visit. It has memory and it builds real artifacts (live dashboards, working documents, actual files), so I run it like standing staff instead of a Q&A tool. It covers the departments a solo shop can't hire: media, finance, rental ops. When I'm building something technical, I drop into Claude Code. The honest knock is still the interface. It's deep, but you have to spend a little time learning where things live before it pays off.

2. ChatGPT, by OpenAI. Back in the rotation as my second opinion.

I'd cooled on it. I've warmed back up. Its Work function (on model 5.6) does the same agentic gather-and-produce work: pull across your connected apps, hand back finished sheets and docs. But it earns its keep two ways for me. First, it's my second opinion. When I'm weighing an opinion or a high-value question, I run it past ChatGPT to pressure-test what Claude told me, and the gap between two strong engines that don't always agree is exactly where I slow down and look closer. Second, it's my design room. I lean on it for rehab visualizations, rendering a kitchen layout onto a photo of the real room before I commit a dollar to cabinets.

3. Cursor. Where I actually build the software.

New on the list. Cursor is the backbone for building my forVEX ecosystem, the core apps that actually run the business, and it complements Claude Code rather than replacing it. If you've crossed from prompting AI to building real tools with it, a coding-native environment earns its own slot. That's a different job than the chat-and-artifact work above.

On the bench: Perplexity. I don't live in it, but I keep it close for the research-first jobs: go find out, then do something with it. Its Comet browser and local-file features put it in the same agentic shape as the rest, with a research spine underneath.

The moral, this quarter. Notice the list isn't really about which chatbot is smartest. It's default, second opinion, build tool, research bench. Four jobs, four tools. That's the tell that these stopped being answer boxes and became infrastructure. The operators pulling ahead aren't the ones who found the one magic model. They're the ones who gave each tool a job and built a system. If your mental model of AI is still one tab you paste questions into, you're running a staff of one and calling it a search bar.

The layer most people skip: make the tools yours. Using these tools is table stakes. Customizing them is where the leverage is, and it's gotten easy. Two moves I lean on hard now:

Skills. Instead of re-explaining myself every session, I build skills that carry my rules and voice so the AI works like a trained team member. There's even a skill whose whole job is helping you build skills. I've got a small library running my content and deal workflows. I added two this week: one that publishes a blog post end to end, one that preps a newsletter issue for send.
Connectors. I wire the AI into the apps my work already lives in, like Gmail and Google Drive, so it reads and acts on my real data instead of a copy I pasted in. That's the jump from a smart chat to something that operates inside your business.

Where it still bites. The new trap isn't picking the wrong tool. It's running all of them shallow. Four half-used tools lose to one you've gone deep on. So pick a default (mine's Claude and Cowork), and use the others deliberately, not out of FOMO. And the old warning still holds: an agent that can act can act confidently wrong. Read the artifact before it goes near a seller, a lender, or your books, and never let one set a number unsupervised. Natural language on top, your logic and your review underneath. The tool does the task. You still own the call.

The play: give one business function a Project, not a chat. Here's the move that separates people who use AI from people who run on it. Pick one core function (your bookkeeping review, your rental portfolio, your listing pipeline) and stand it up as a Project. Claude or ChatGPT both do this. Not a fresh chat every time. A Project with three things a chat doesn't have:

Instructions: its job and your rules, written once.
Memory: so it remembers last month and builds on it.
Artifacts: the running sheet or dashboard it maintains for you.

A cold-start you can paste today:

"You are my [role, e.g. bookkeeping reviewer] for [my business]. Each month I'll upload [the statements/exports]. Your standing job: [flag miscategorizations, compare to prior months, list what to ask my bookkeeper]. Keep a running log across months, and at month-end hand me a one-page summary."

Fill in your own function. The difference between that and a one-off chat is the difference between a temp you re-explain everything to and a team member who already knows the account.

Know an operator still treating AI like a search box? Forward it. That's how this list grows. And hit reply: what's your stack right now, and what's the one business function you'd hand to a Project? I read every one.

Written by Paul Cobb, real estate operator (Louisville / Kentuckiana), forVEX builder. I run the plays I write about. More depth: forvex.app/journal