I don't write the code. I make sure it is right.
An AI generalist from Chhapra, Bihar: what the work actually is, how I started, why I took a Python course I never intend to use, and a running log of what I learned and when.
I'm an AI generalist. That is not a borrowed job title, it is what the work actually is: moving between prompt, context, loop and graph depending on what the job needs, and picking whichever free tool tests best for that job.
I started a year ago with ordinary chat. The first thing I built was my own agent: a scheduled worker that checks whether the job is actually done, creates follow-up tasks when it needs more time, wakes itself and keeps going, and writes notes into a GitHub repo so the next run starts with context instead of from zero. It ran on free models with no API key and cost me nothing.
I don't write code and I don't intend to. I direct it, and I took a Python course for one reason: so I can describe a job well enough that the AI cannot fill the gaps with guesses, read the structure it hands back, and catch its mistakes. A team leader who cannot do the work cannot get the work done. Same rule here.
For now I build for myself, not for clients. Feedfy is next: instead of scrolling for an hour to find content worth keeping, a feed that goes and finds it, in about two minutes.
- Role
- AI generalist
- Focus
- Prompt · context · loop · graph
Automation · agent building - Stack
- Whichever free tools test best for the job
- Location
- Chhapra, Bihar · India
IST (UTC+5:30) - Status
- Building for myself
What I learned, when, and what I built with it.
A running log, newest at the bottom. Every entry names the thing I learned and the tools I actually used to make it, because "I know AI" means nothing without the receipts.
Started with plain chat
No agents, no connectors, no code. Just asking, reading the answer, and noticing how often the confident answer was the wrong one.
My first agent — a scheduled worker that kept its own notes
It checks whether the job is done, creates follow-up tasks when it needs more time, wakes itself and continues, and writes what it learned into a GitHub repo so the next run starts with context instead of from zero. Free models, no API key, no VPS.
TruePick — my first real product
A nine-page affiliate storefront driven entirely by JSON, with a blog that reads itself aloud and YouTube reviews on product pages. It taught me that data outlives design.
Social Media Manager — my first public skill
An agent skill that discovers which connectors actually work before it promises anything, then keeps a knowledge base of brand voice so the writing sounds like the person, not the model. MIT licensed, because a skill nobody can fork is a brochure.
One small tool a day
Twenty days, twenty tools: timezone overlap, CSV triage, SQL guard, SLA clock, contract drift, spec ambiguity. None of them impressive alone. All of them together are a habit, and habit is what I was actually building.
Employee Hub — AI employees with a written standard
The biggest thing I have built: five new AI employees every morning, 57 skills, and one rule that matters more than the rest — never fake the part you cannot do.
This site
Nine pages became eight, one hero video became 47 frames, and every word on it came out of a conversation rather than a template. The scroll animation is the only clever thing here, and it exists because a still image felt dead.
Feedfy — designed, not built yet
Instead of scrolling for an hour to find content worth keeping, a feed that goes and finds it in about two minutes. I have the data and the plan. When it exists, this line gets a date.