Building AI sales agents 101:
1. Start simple
A cart-recovery agent is the easiest first agent. Landing page + form → checkout.
Non-buyers get the conversation.
2. Always plan the exit
Define clear escalation criteria: • wants a human • missing data to answer • unusual situation
3. Examples beat instructions
Paste the exact ideal conversation in the prompt. The agent becomes way more assertive.
Your first agent shouldn’t be impressive. It should work.
This week I was setting Black November offers and uncertainty hit hard.
“Will everything work?”
I deleted fear from my vocabulary.
Replace it with: risk.
We take the risk to collect the reward.
That’s how you stay rational when the market and entrepreneurship are full of unknowns.
As a founder I’m done with third-party software I can’t change in 10 minutes.
CRMs. ERPs. Marketing tools. Checkout. Messaging.
I can’t just open Claude Code and make them work the way my operation actually works.
That’s the new standard:
Every piece of software should have a chat interface so I can talk to it, modify the interface, and fork it to my needs.
If vendors don’t build this, we’re going to replace them with vibe-coded tools.
We’re already hunting. 🔥
I see my Claude Code sessions in my IDE superset as a “locomotive builder.”
The core idea is to run multiple agents at the same time, with workers spawned in parallel, each one building its own wagon (SPEC).
In the image below, I have 12 workers simultaneously building 12 different specs for the same project.
At the same time, I use the LOCAL repository as the locomotive engineer, the leader of the pack.
And here, the workflow is different.
The engineer ships PRs to staging and production in series, one by one, pulling them from the line of wagons being built in parallel.
I’ve set predefined rules in the system prompt so that no worker branch can ship directly to staging or production. Only the local branch has permission to ship.
So the architecture is basically a locomotive:
Workers build in parallel.
The locomotive engineer ships in series.
Run multiple agents at the same time. Parallelize the building. Serialize the shipping.
Claude coding on weekends:
I can now build software for almost any small inconvenience in my life.
Auto-reorder things I buy every month. Extract insights from long YouTube videos I don’t have time to watch. Automate repetitive tasks built specifically for me.
Personal software is becoming almost free to create.
I wonder what this looks like in 5 years.
My honest AI workflow today:
I use Superset as my IDE. It keeps multiple projects organized on a single screen.
I run several SPECS in parallel. Each spec lives on its own branch, pulled from the project I’m building.
Each branch opens a Claude Code session and it can run any LLM at the same time. That’s how I keep 10+ agents working simultaneously.
Every spec still needs my time. Not to read everything Claude writes, but to give the software the TASTE I want. I skim for what matters and let it ride.
At the end, I push the PRs from my local branches one by one. It’s a line: staging first, production after.
Build in parallel. Ship in series.
Have you noticed the power of repetition?
Our brains need 8 repetitions of information before considering it for memory storage.
One message repeated 8 times is better than 8 different messages sent only once.
Repeat your message and customers will listen.
I worked my entire life with a clock in the wrist as a medical doctor.
Now as an entrepreneur, I feel like there is fewer obligations with the ticking clock. However, there is much more responsibility with the result itself.
Changing mentality from 9-5 to 168h to alocate/week.
Reading “Ready, Fire, Aim” as suggested by @AlexHormozi.
It’s incredible how a few text lines can bring CLARITY to make decisions in each business level.
From the book:
0-1 million: SELL
1-10 million: BUILD
+10million: LEAD
Very accurate so far! Product mode ON!
It's been 9 months since I quit my job as an MD anesthesiologist, to pursue the entrepreneurs journey.
Enjoying the ride so far to build an 100 million dollar business.
Just started sharing the insights along the way.
Follow for hidden knowledge acquired in the front lines.
@startupschool#BuildSprint one-liner: We’re TodoMed. An E-commerce for educational ebooks, to help inexperienced or outdated medical doctors to improve their medical skills.