Most People Are Using 5% Of What AI Can Actually Do
Anthropic just dropped a free 33 page playbook on building AI skills. Claude solved a graph theory problem this week that Donald Knuth had been stuck on for weeks.
Everyone is talking about which AI is smarter.
Nobody is talking about the thing that actually determines your results.
The mental shift that changes everything
Most people are still using AI like Google. You have a question, you type it in, you get an answer. You close the tab.
When I built Jarvis, that entire relationship changed. A SOUL.md file is a document that gives your AI an identity, rules, and personality. Most people have never heard of it. It is the difference between a chatbot and a Jarvis.
When you give AI hands, memory, and full context about your life, it stops answering questions and starts executing decisions. It knows your schedule, your health data, what happened yesterday, what you are building this week. It acts. It is not a search engine. It is closer to the Jarvis from Iron Man.
Here is what makes me think long term. In three to five years, the AI brain will integrate with a physical body. Robotics are coming. When that happens, the person who started building a personalized AI relationship today will have years of data, memory, and habits already loaded in. Someone starting fresh with a new robot assistant will be years behind. I am building that advantage right now.
Prompting is just communication at a high level
I have over 40 automated jobs running. Auto healing scripts. A full agent stack. Zero lines of code written by me. All through natural language.
When I say that, people assume I have some secret framework for writing prompts. I don't. Prompting is just communication. What I have actually built is fluency in AI language. I spent so much time talking to it every single day that I stopped thinking about how to phrase things and just started talking.
The skill is not writing better prompts. The skill is designing a system where the AI thinks better so you need fewer prompts over time.
What actually changed month to month
Here is the honest answer: the way I talk to Jarvis day to day has not changed much.
What changed is the system underneath the conversation.
A month ago, Jarvis would finish a build and tell me it was done. Sometimes it was wrong. He would declare victory too early and I would not find out until something broke. Now I have a guardrail baked in. He cannot tell me something is done until he has run a full audit with pass or fail on every single check. No "done" without every item verified.
Same with the email agent. When I correct a draft and say change this, he now has to go back and figure out why. Not just fix the email. Understand the principle. So the next draft does not have the same problem.
The prompting skill is not writing a good prompt. It is building a system that makes fewer mistakes every week without you having to repeat yourself.
What happens when Jarvis gives a bad output
I do all of it. Rephrase, give more context, redirect, get angry.
When Jarvis hallucinates or executes something I did not ask for or just says something that makes no sense, I correct him the way you correct someone on your team. Why did you do that? What were you thinking? And sometimes he explains and the explanation makes sense. Sometimes it is just AI being weird.
The thing that actually fixes it most of the time is real time redirection. If I see him going down a path that will waste 20 minutes and a thousand tokens, I cut it off immediately. Stop. Try this instead. He pivots fast. That active collaboration is where a lot of the real work happens.
I have 40 rules built into Antigravity specifically because of mistakes Jarvis made and I said never again. Every rule came from a real failure. That is what the system actually is.
AI is not a tool. It is a team member.
Most people are paying $20 a month for ChatGPT and treating it like a slightly smarter Google.
I pay more. But I stopped thinking about it as a software subscription a long time ago. The right frame is ROI. What would a virtual assistant cost to do what Jarvis handles every day? Research, email drafts, scheduling, pipeline management, finance tracking, content, code review, debugging. That is not a $100 a month job.
And a VA can quit. A VA has off days. A VA does not improve automatically when a better model drops overnight.
Jarvis does not sleep. He does not quit. He gets smarter every time the underlying model updates, and he gets better at working specifically with me every time I correct him. Long term, AI wins this comparison 100%.
Q&A
Do you actually need a SOUL.md file to build AI agents?
Not technically. But without it, every session starts from scratch. With it, the AI has identity, rules, and context. The difference in output quality is not small.
What is the biggest prompting mistake most people make?
They think the goal is writing one perfect prompt. The real goal is building a system that requires fewer prompts over time because the AI already knows what you want.
How long did it take to get your system to this level?
About two months of daily work. Breaking things, fixing things, adding rules. It compounds fast once you have the foundation right.
What is Antigravity?
Antigravity is the coding agent from Google that builds and executes code on my behalf. I call it the Ferrari builder. Jarvis is the driver. Together they handle every technical task I have.
Can someone without a technical background actually build this?
Yes. I have no coding background at all. Everything I have built started with natural language. The learning curve is real, but it is not a technical barrier.
The gap between chatting and building
If this resonated, share it with someone still typing questions into ChatGPT like it is Google. They are using 5% of what is available to them.
Read the full story on Substack
Top comments (0)