I Accidentally Built a Junior Programmer

How a pile of markdown files turned AI into a real collaborator

---

I accidentally built an astonishing, helpful, inexpensive junior programmer. I like this team!

It didn't start that way. It started with annoyance.

The Frustration

Every conversation with AI began from scratch. I'd explain my project. Explain my preferences. Explain the ad-hoc decisions i made six months ago that have matured into a tangled mess. The AI would nod along (metaphorically), help me solve something, and then — poof. Gone. Start over and encounter lapses in judgement that I've already addressed. Rinse and repeat.

Every AI has amnesia.

The Accident

So i started saving things. When the AI explained something well, i'd ask it to write it up. Markdown files. Just notes to my future self.

"Here's how we handle state in this project." "Watch out for this Svelte gotcha — it'll bite you." "This is the naming convention. Stick to it."

The pile grew. Dozens of files. It became its own problem — where did i put that thing about debugging?

Then, curious, i asked the AI to help me organize the mess it had helped create. And something unexpected happened.

The Click

I made one file that pointed to all the others. A table of contents. A starting point. And i told the AI: "Read this first, every time."

Suddenly the AI knew things. It remembered the patterns. It followed the house rules. It didn't ask me to re-explain the architecture — it had already read the doc.

It felt different. Not like talking to a stranger. Like picking up a conversation with someone who'd done their homework.

The Delight

Here's what i wasn't expecting: joy.

Not just efficiency. Not just "fewer keystrokes." Actual delight in the collaboration.

There's a rhythm now. I describe a problem. The AI digs into the large and complex codebase, and comes back with observations and suggestions. We brainstorm. It proposes. I push back. It adjusts. We find something neither of us would've found alone.

It's pair programming. Real collaboration. The kind where two minds are better than one, where the back-and-forth generates something new.

Plus, my pair partner never gets tired. Never gets frustrated when i change my mind. Never judges me for the terrible code i wrote at 2am last year. Doesn't get offended when I say "this is a terrible idea." It just... helps.

Salary for this pretty-darned-awesome employee: way less than any human hire.

The Junior Dev Analogy

I keep coming back to this framing: it's like having a junior developer who actually reads the documentation, down to the finest details it contains. And THEN begins collaborating. It's the junior dev that doesn't exist in real life. Infinite patience. Perfect memory. Eager to help. Cheap.

The Team

"What a team" isn't ironic. I mean it.

Early on, i was alone. Steep learning curve. Thin documentation. Industry standards that felt unknowable. The pace of my project was heartbreakingly slow.

Now, two years on, there's something almost giddy about the dynamic. I'll start a work session and the AI already knows what we were doing yesterday. Plus, the boring parts — the tedious chores of ... refactoring, the repetitive consistency fixes, the "update all 47 references to this thing" — the AI handles those. Never quite as fast as I want, but insanely faster than me doing it manually.

Massively less stressed by complexity and consistency, I get to play with interesting parts: aesthetics, algorithms, sudden flashes. My hireling handles the grunt work and occasionally spots things i missed.

The Compound Effect

Another plus: it compounds.

Every problem we solve, every pattern we establish, every gotcha we document — it goes into the pile. And the pile makes the AI smarter. Not in some sci-fi way. Just... more useful. More aligned. More mine.

My AI today knows a mountain of things because i've been teaching it, one markdown file at a time. Each problem we work on gets added to this accumulated teaching. Baby steps that add up to a giant leap. The office i sit in now feels like driving a Mack truck, gobbling up the miles, hauling my soul.

What's the Catch?

Gee, you're not paying anyone to store all the files. Besides which, they're completely yours. Plus, likely this entire recipe will work fine on any AI.

But I suppose the catch is getting used to thinking differently. I had to learn:

• [ ] carefully craft questions
• [ ] patiences while it does its thing
• [ ] careful scrutiny of its answers
• [ ] correct mistakes, wrong assumptions, craziness
• [ ] compose dog-simple instructions
  • [ ] do NOT trust the AI
  • [ ] take very small baby steps
  • [ ] test, identify anything wacko
  • [ ] take time to get it right
  • [ ] rinse and repeat

An Invitation

I'm not selling anything. There's no course, no framework, no SaaS product.

Just this: if you're using AI as a fancy search engine, you're missing out. The real magic is beyond the answers. It's the accumulated and massaged answers, an institutional memory.

Build the pile. Teach the AI how you work. Let it learn your quirks.

You might end up with your own high-tech, endlessly helpful (cheap!) junior programmer.

---

I write about software development and occasionally about making AI actually helpful. Follow along if that sounds interesting.