Learning verification through the Socratic Method — powered by AI, driven by your own thinking.
I maintain a digital garden (a personal wiki built with Quartz + Obsidian) with 380+ notes covering LLMs, signal processing, mathematics, physics, and more. Many of these notes were expanded with AI assistance — which creates a trap:
Writing a note with AI help ≠ understanding the material.
I found myself staring at beautifully structured notes and wondering: "Do I actually know this, or did I just organize information well?"
A global skill for Claude Code and Codex:
quiz: turns the agent into a Socratic learning partner.
Instead of asking the AI to explain concepts to me, /quiz does the opposite: the AI primarily asks questions, and I must answer from my own understanding. It probes deeper with each exchange, following a 5-level depth progression:
Level 1: Basic Recall — "Can you say it in your own words?"
Level 2: Relationships — "How does X relate to Y?"
Level 3: Application — "How would you use this in a new scenario?"
Level 4: Critical View — "What are the limits? When does this fail?"
Level 5: Synthesis — "What new thing can you create from these ideas?"
The AI stays in Socratic mode by default: it asks, tracks mastery across the core ideas, escalates misconceptions through focused prompts, and closes when understanding is demonstrated. If I explicitly leave Socratic mode, it can provide a concise correction before resuming.
In Claude Code or Codex:
/quiz me on LoRA
test my understanding of scaling laws
check if I really know the Chinchilla paper
use the Socratic method on my antenna notes
The skill works with any source — local files, URLs, pasted text, or just a topic name. The agent proactively hunts down the material and begins questioning.
User: /quiz me on LoRA
AI: I've read your LoRA notes. Let's explore this together.
To start — what problem does LoRA solve that full
fine-tuning creates? In your own words.
User: Full fine-tuning needs gradients and optimizer states
for all 7B parameters — that's ~84GB VRAM. LoRA
freezes the base model and only trains tiny adapter
matrices.
AI: Good. Now — LoRA makes a key assumption about the
weight update ΔW. What is that assumption, and why
is it reasonable?
User: It assumes ΔW is low-rank... because the intrinsic
dimension of task adaptation is much smaller than
the full parameter space.
AI: Interesting — "intrinsic dimension." If a 7B model only
needs to change a few hundred degrees of freedom to
learn a new task... what does that imply about how
knowledge is stored in the pretrained weights?
... (continues probing deeper, never giving answers)
| Flashcards / Quizzes | Socratic Method |
|---|---|
| Test memorization | Test understanding |
| Right/wrong answers | Depth of reasoning |
| Fixed question set | Adaptive follow-ups |
| You vs. the test | Dialogue with a mentor |
| Surface-level recall | Exposes hidden gaps |
The Socratic method reveals unknown unknowns — the gaps you didn't even know you had.
Paste this into Claude Code or Codex — the agent will install the quiz skill:
Install this as a global skill for both Claude Code and Codex:
https://raw.githubusercontent.com/PinkR1ver/socratic-learning/main/skills/quiz/SKILL.md
The agent reads the skill file, creates the right directories, and sets it up. No manual commands needed.
socratic-learning/
├── README.md
└── skills/
└── quiz/
└── SKILL.md # The Socratic questioning skill
- Claude Code or Codex
- A knowledge base to quiz yourself on (works with any markdown notes, not just the author's digital garden)
- No API key needed — the skill runs entirely within your existing AI agent
"I know that I know nothing." — Socrates
The goal is not to feel smart. The goal is to find the edges of your understanding — and then push them outward.
This is a method, not a tool. The skill file is just the prompt that keeps the AI in character. The real work happens in your mind, one question at a time.