# The Self-Learning System
Case Study #1 from syNRGy
By Navya Rehani Gupta (NRG)
https://synrgy-nrg.com/self-learning-system

---

A system that makes me sharper every week, automatically.

## How It Works

Claude doesn't remember anything between sessions. Every conversation starts blank. So I built the memory myself: three plain-text files that load at the start of every AI session. Each one compounds.

| File | Purpose |
|------|---------|
| Working Preferences | How I think and decide |
| Mistake Patterns | AI errors I've caught and codified |
| Session Learnings | Running improvement log |

**Every session loop:** Load context > Work gets done > Capture insights > Catch AI errors

**Every Friday loop:** Score 5 dimensions > Compare to top operators > Surface upgrades > Next week is smarter

The structured approach (CLAUDE.md, hooks, skills) comes from Dave Killeen's Dex (https://heydex.ai/), an open-source AI operating system for professionals. Scoring framework, automated briefing pipeline, and AI mistake codification built on that foundation.

## Why I Built This

I realized I had no way to measure whether I was improving with AI or just using it more. That bothered me.

We're all using AI now. But are we getting better at it, or just busier with it?

So I built a system that answers that, every week. Two hours to build. Returns compound every week. It's already changed how I make decisions, how fast I ship, and what I catch before it becomes a problem.

## The Friday Email

Every Friday morning, an automated email tells me whether I leveled up or coasted. It scores me across five dimensions:

- **Automation:** Do my workflows run with minimal intervention?
- **Learning & Memory:** Does every session build on the last?
- **Custom Workflows:** Am I eliminating repetition?
- **Ecosystem Reach:** Am I using the best tools available?
- **Resource Awareness:** Am I being efficient with what I have?

### My Weekly Scorecard (7.8/10)

| Dimension | Score | Note |
|-----------|-------|------|
| Automation | 8/10 | Daily runs with retries. Next: parallel job execution |
| Learning & Memory | 9/10 | Every session compounds. Best dimension in the system |
| Custom Workflows | 7/10 | 10 active. Several manual processes still not converted |
| Ecosystem Reach | 6/10 | New AI capabilities ship weekly. Room to adopt faster |
| Resource Awareness | 7/10 | Token tracking works. Not yet proactive about cost optimization |

Next milestone: 9.0/10 (+2 Ecosystem, +1 Automation)

This score will never be 10. The tools evolve too fast. The point is to keep climbing.

## Technical Architecture

- **Memory Layer:** Three plain-text markdown files. No databases, no APIs. Human-readable, AI-readable, version-controlled.
- **Automation Layer:** Hooks fire on specific events: session-start loads learning files, auto-capture prompts for insights after significant work.
- **Skills Layer:** Ten reusable workflows as slash commands, each built after repeating the same workflow three times.
- **Scheduling Layer:** Five automated jobs M-F: 5am briefing, model updates, industry digest, Friday review, catch-up retries. All with validation gates and duplicate prevention.
- **Weekly Scripts:** Friday review analyzes session patterns, researches top operators, scores my setup across five dimensions, and emails a report with upgrade recommendations.

### File Structure

```
~/.claude/
├── CLAUDE.md                    # Master instructions
├── settings.json                # Hooks configuration
├── hooks/                       # Automated triggers
├── skills/                      # Reusable workflows (10)
├── scripts/
│   ├── daily-briefing.sh        # 5am industry briefing
│   ├── claude-updates.sh        # Model + changelog check
│   ├── agentic-ai-digest.py     # Competitive intel digest
│   ├── friday-weekly-review.sh  # Full weekly analysis
│   └── catchup-missed-jobs.sh   # Retry at 9am, 12pm, 2pm
└── learning/
    ├── Working_Preferences.md   # How I operate
    ├── Mistake_Patterns.md      # AI errors I've caught and codified
    └── Session_Learnings.md     # Running improvement log
```

## What I Learned

The returns started compounding in week two. Three things surprised me:

The biggest win for a busy operator: every week I got actionable tips I never would have found on my own. The system surfaced three things automatically:

- What other top operators were doing that I wasn't. Techniques and workflows I'd never have time to research myself.
- Things I was doing repeatedly that needed to be automated. Patterns I couldn't see because I was too close to the work.
- New model improvements I needed to be aware of. Capabilities that shipped while I was heads-down on other things.

I also learned that most of my improvement came from AI mistakes the system caught and codified on its own. When the AI gets something wrong, the system writes a prevention rule automatically. I don't maintain the file. It maintains itself. The Friday score made me honest. I couldn't tell myself I was improving when the number said otherwise.

If I were starting over, I'd build the mistake patterns file first. You're debugging and training the AI. Everything else layers on top of that.

## Compare Notes

Curious how your AI usage is actually compounding? I'd love to hear what you'd measure.

Connect on LinkedIn: https://www.linkedin.com/in/navyarehani/

---

Learning loop inspired by Dave Killeen's Dex (https://heydex.ai/). Scoring framework and automation pipeline built on that foundation.

(c) 2026 Navya Rehani Gupta