Let me be honest upfront: I’m not a complete stranger to code. I took some programming in college. I can read code and understand what it’s doing. But write it from memory? Build something from scratch? No.
I can’t sit down and type out a function. I don’t remember syntax. I couldn’t tell you the difference between a Promise and a callback without looking it up.
Over the past few weeks, I shipped this:
- SQLite database with 15+ tables
- FastAPI backend with WebSocket real-time streaming
- React dashboard with multiple views
- Multi-agent coordination system with 4 AI personas
- Automated test suite
- Cross-platform installers for Mac, Linux, and Windows
By directing an AI.
## The Problem
I use Claude Code daily. It’s powerful. But every session starts from zero.…
Let me be honest upfront: I’m not a complete stranger to code. I took some programming in college. I can read code and understand what it’s doing. But write it from memory? Build something from scratch? No.
I can’t sit down and type out a function. I don’t remember syntax. I couldn’t tell you the difference between a Promise and a callback without looking it up.
Over the past few weeks, I shipped this:
- SQLite database with 15+ tables
- FastAPI backend with WebSocket real-time streaming
- React dashboard with multiple views
- Multi-agent coordination system with 4 AI personas
- Automated test suite
- Cross-platform installers for Mac, Linux, and Windows
By directing an AI.
## The Problem
I use Claude Code daily. It’s powerful. But every session starts from zero.
You spend an hour debugging something, finally fix it, close the terminal... and next week Claude might make the exact same mistake. It has no memory. You end up re-explaining your project’s quirks over and over.
What if Claude could remember what worked and what didn’t?
That’s what I decided to build.
## What I Built
ELF (Emergent Learning Framework) - persistent memory for Claude Code.
It records failures and successes to a local database, then injects relevant history into Claude’s context before each task. Patterns that keep working gain confidence. Eventually they become "golden rules" Claude always follows.
It’s not AI magic. It’s structured note-taking with automatic retrieval.
## How It Actually Works
A typical session looked like this:
Me: The dashboard should show heuristics with confidence scores.
Claude: writes code
Me: Make confidence a progress bar, not just a number. Sort by highest first.
Claude: updates code
Me: That broke the other tab. Fix it.
Claude: debugs and fixes
Multiply that by hundreds of exchanges over a few weeks. That’s the process.
I couldn’t have written any of it. But I could:
- See when something looked wrong
- Know what "done" should look like
- Catch when a fix broke something else
- Decide what to build next
## What Skills You Actually Need
Reading code - Not writing it. Just enough to follow along and spot obvious problems.
Knowing "done" - The AI doesn’t know your standards. You have to recognize when something is right.
Product thinking - What to build, what to skip, what order to build it in.
Persistence - Things break constantly. You keep going.
Taste - Knowing when something is overengineered or when it’s too hacky.
## The Hard Parts
Context limits are real. Long sessions lose coherence. Claude forgets what it built two hours ago. I learned to work in focused chunks.
Debugging without understanding - When something breaks and you can’t trace the code yourself, you’re dependent on the AI to figure it out. Sometimes it goes in circles.
You can’t fully verify - I have to trust the tests, trust the behavior, ask probing questions. I can read the code but I can’t always evaluate if it’s good code.
It’s probably not faster - A skilled developer would have built this quicker. But a skilled developer wasn’t building it. I was.
## The Meta Twist
The system I built solves the exact problem I kept hitting while building it.
Every time Claude forgot what we’d done, I thought: "This is why I’m building this."
By the end, the framework was recording its own development. Claude was working better because it had history to draw from.
## The Project
ELF (Emergent Learning Framework)
- Persistent memory across Claude Code sessions
- Pattern tracking with confidence scores
- React dashboard to visualize what’s been learned
- Multi-agent coordination for complex tasks
Open source. MIT license.
GitHub: https://github.com/Spacehunterz/Emergent-Learning-Framework_ELF
## Who This Is For
If you can read code but can’t write it - you’re closer to building things than you think.
The gap between "understands programming" and "builds software" just got smaller. Not gone, but smaller.
I’m not a developer. But I shipped something real.
Built with Claude Code. I can read every line - I just couldn’t have written them.