Andrew Dunn

I sell dump trucks of rocks on the internet ( Host AI Friday in NOLA. Build weird tools with AI.

[Andrew Dunn @DunnWithAI]

Finally, I have my ChatGPT Dot managing my email and pings me about what needs attention. It's the most proactive but I've found.

Also played with Claude Duet. The channels idea seems promising, but I tried to get to replace a process I have running in a GrokBot and it's super limited.

[Andrew Dunn @DunnWithAI]

Vibe Check: Here's a thread of the AI tools I'm using right now.

I was a heavy Claude user to start of the year then GPT5.5 and Sol took my heart. Claude recently pulled me back after I was spinning my wheels on some project sin Sol. Gave it to Opus 5.5 and got unstuck.

[Andrew Dunn @DunnWithAI]

Lauren (@poteto) over at @spacexai cannot be beat. She's so good at sharing her processes and so maybe my issues with the Grok model is a skill issue and I need to go deeper on some of the stuff she's posted.

https://t.co/csPEZhpJ2R

[lauren @poteto]

my favorite use case for Grok Bot is for the first and the last mile of getting work done.

the first mile is figuring out what work you should even be doing at all. that involves connecting Bot to the different apps and services you use, so that it can grab whatever context it needs - whether it's your calendar, your CRM, slack, email, or the many other connectors available. then, the first mile begins when you combine these connections, allowing you to provide context to your bot ("look at this slack channel for new features we've built and then look through our CRM to see which customers we should nudge") instead of micromanage it.

you can automate further by setting up routines - they let your bot do something on a schedule ("check my email every morning") or as a reaction to something ("do this whenever this slack channel has a new message"). so a new lead can kick off a whole workflow that you setup for your bot, allowing you to delegate the work you would otherwise have to do to your bot!

the last mile involves actually following through and closing the loop. one example of this is building a new feature, or fixing a bug that your customers have reported. this is where Bot can hand off tasks to Cursor - our engineering focused agents that can write code in the cloud, on their own computers.

when you've gotten the hang of this, the next thing you can try is building a loop. here's a simple example that i use daily:

1. watch this slack channel for feedback about our app
2. create a ticket for this in Linear
3. create a Cursor cloud agent to triage and reproduce the issue on its own computer, using pstack
4. if the issue clearly reproduces, put up a fix and fuzz the pull request with a small swarm of agents
5. if the fuzz has no issues, ping me on Slack, and rebase and auto-merge the PR after 1 hour unless i request changes

you can also install my bot Dr Eggbot and ask it to teach you how to setup a loop like this for yourself!

https://t.co/U1shj39y3J

[Andrew Dunn @DunnWithAI]

Thanks to @TheIdeaVillage's Jon Atkinson for shouting out our "boring company that sells rocks online."

From the @NOLAnews Louisiana Innovates panel: https://t.co/K4kHnII2Ex

Couldn't have described us better if I tried.

[Andrew Dunn @DunnWithAI]

Helpful prompt to use: "If our traffic doubled overnight due to bots, what would it cost us?"

And be thoughtful about the cat and mouse game of blocking them (it never ends). Can you stay performant? And can you make sure your downstream systems (particularly analytics) won't blow up your bills?

[Andrew Dunn @DunnWithAI]

Bots tripled traffic to https://t.co/jYstNZCu5O for three weeks last month. Leads were flat to down (expected as we get closer to winter).

They ran real Chrome, executed our JavaScript, and fire analytics events. PostHog labeled 99.9% of them human.

Our @Cloudflare usage went from $342 to $1044. @posthog bill went from $614 to $1121. So +$1K for bot traffic.

[Andrew Dunn @DunnWithAI]

You can't just block them all. On our peak day, more than half the traffic was verified "good" bots and one AI search crawler made 1.27M requests my itself. We're cool with that, we want AI search reading our product pages.

[Andrew Dunn @DunnWithAI]

I don’t know if this helps but our son was a bad sleeper. We also tried everything. And yeah in hindsight I wish I wouldn’t have tried so hard. Seems easier to just embrace it than be frustrated all the time.

Now he sleeps great, and when he doesn’t he’ll just wake up and quietly read a book and everything is fine.

I also now understand why sleep deprivation torture is a war crime and I’m still thinking about calling the authorities on my little war criminal.

[Andrew Dunn @DunnWithAI]

I sort of bet on this with my taxes! In April Claude downloaded the IRS forms, mapped the fields, and checked the math, but I filed an extension to finish and verify with better models.

By October it was much more capable. Checked my email for every "your tax documents are ready" message as a checklist, reviewed the tax code, and used browser use to enter everything into tax software to cross-check its own work.

[Andrew Dunn @DunnWithAI]

In case you're wondering what happens if you try to put your Dot and Instinct in the same chat.

[ChatGPT @ChatGPT]

By popular request, you can now text your dot.

Give it a task to take off your plate or something to keep an eye on, using text, photos, or even voice notes. It will text you proactively with relevant updates.

Safe, secure texting via iMessage or RCS. You’re the only one who can text your dot, and your dot can only text you. Rolling out now as a preview for Pro users in the US.

[Andrew Dunn @DunnWithAI]

Here's the prompt I used. It was a quick throwaway prompt just to see what I'd get.

so, I love this: https://t.co/U0paafuuzH

https://t.co/yKtSTKw6se

But:

- It's not the best/easiest to use on mobile.
Just downtown is pretty limiting.
- I think we can think more expansively ambitiously with this: https://t.co/YqFwiKypqy

Thinking out loud is we do some deep research, come up with a good name for this project, host it on https://t.co/oP01tc4Zk8 and cloudflare (I can give you my stack.md later).

[Andrew Dunn @DunnWithAI]

The NOLA Downtown Development District put out a public art map. It's neat, but it only covers downtown and it's hard to use on a phone.

So I gave Claude the link, said "think more ambitiously," answered a couple of questions, and forgot about it. When I checked that night it had built a map of the whole city: 372 murals, sculptures, and monuments across 42 neighborhoods, with walking routes.

https://t.co/om7AmYjTOc

[Andrew Dunn @DunnWithAI]

I had my agents count every prompt I've typed this year into ChatGPT and Claude. I was hot and heavy with Claude for a while, then I became a Sol man. By September, 92% of my coding prompts went to Codex.

Then Sol started feeling dumb and grinding on problems it used to just solve. Everyone around me loved Opus 5.5, so I switched back a couple of weeks ago and got unstuck right away. Plus the Claude Projects feature is amazing.

ChatGPT is still better at using the computer and at going and doing things on its own when it needs something done. Claude is the better model and currently my daily driver.