AI That Codes Smarter? Andrej Karpathy’s “Skills” Explained for Non-Developers

AI tools like ChatGPT, Claude, and Cursor can now write entire pieces of software. But anyone seriously using them quickly notices the same problem: the AI often does too much.

It changes code that didn’t need changing. It invents extra features. It refactors half the project. Or it makes assumptions that are completely wrong.

That is why a new trend is emerging among developers: teaching AI discipline through Skills and behavioral rules.

One interesting example is this GitHub repository:

https://github.com/forrestchang/andrej-karpathy-skills

What Exactly Is This?

The repository andrej-karpathy-skills is not a software library, not a framework, and not an AI model.

It is essentially a collection of behavioral guidelines for AI coding tools.

Think of it as:

• a training manual for AI programmers
• a “think before coding” protocol
• a safety framework for AI agents
• a workflow for making AI code more intelligently and predictably

The name references ideas and observations from Andrej Karpathy, former Director of AI at Tesla and one of the best-known AI engineers in the world.

The repository itself was created by Forrest Chang, who translated Karpathy’s ideas into practical AI development rules.

Why Is This Needed?

Every developer using AI tools recognizes these frustrations:

❌ Problem 1 — The AI Changes Too Much

Prompt:
"Make this button blue"

Result:
- New component structure
- CSS refactor
- Extra helper functions
- 12 files modified

AI often tries to be “smart,” but ends up creating unnecessary technical debt.

❌ Problem 2 — The AI Makes Assumptions

Prompt:
"Add authentication"

AI:
- chooses JWT automatically
- creates middleware
- builds refresh tokens
- changes database structure

Without first asking what you actually meant.

❌ Problem 3 — Over-Engineering

AI tends to create solutions that are “too professional”:

• unnecessary abstraction
• design patterns nobody requested
• generic systems for simple problems

So What Do These Skills Actually Do?

The repository tries to make AI agents:

• more careful
• more focused
• simpler
• more controllable
• less chaotic

when generating code.

The 4 Core Principles

1. Think Before Coding

The AI should think first before writing code.

Instead of:

"I know what the user means"

It should say:

"These are my assumptions"
"This part is unclear"
"There are multiple possible interpretations"

Example

Bad:

User:
"Build a login system"

AI:
- creates JWT
- uses Redis
- builds OAuth
- creates a complete auth infrastructure

Good:

AI:
"Do you want:
- simple session login?
- JWT API authentication?
- social login?
- Laravel Breeze/Fortify?
"

2. Simplicity First

Keep solutions as simple as possible.

Many AI tools build software as if they are designing systems for NASA. But in reality, most teams want:

• readable code
• maintainable code
• minimal changes

Example

Bad:

AI creates:
- service layers
- repositories
- adapters
- factories
- dependency injection
- interfaces

Good:

Just:
- one function
- clear naming
- directly working code

3. Surgical Changes

Probably the most important rule.

The AI should only change what is necessary.

That means no:

• unnecessary refactoring
• project restructuring
• random formatting changes
• unsolicited “improvements”

Example

Prompt:

"Add a phone number field"

Bad AI output:

• rewrites the entire form component
• adds a new validation layer
• switches UI libraries

Good AI output:

+ input field
+ database migration
+ validation rule

Done.

4. Goal-Driven Execution

The AI should work toward a clearly defined goal.

Meaning:

• what exactly should work?
• how do we verify it?
• when are we finished?

Example

Success criteria:
- user can log in
- session persists
- logout works
- tests pass

This prevents “half-working magic.”

Why This Matters

We are moving toward a world where developers increasingly:

• guide AI
• supervise AI
• correct AI
• orchestrate AI

instead of typing every line manually.

That means the skillset is shifting from:

"How do I code?"

to:

"How do I manage AI programmers?"

What Are “Skills” Anyway?

OpenAI describes Skills as:

“Reusable workflows with instructions, examples, and optional scripts, allowing AI to perform tasks consistently.”

In other words:

You are teaching AI how to behave.

Exactly like mentoring a junior developer.

Think of It Like a Senior Developer

Junior AI AI with Skills Makes assumptions Asks clarifying questions Refactors everything Changes only necessary code Over-engineers solutions Keeps things simple Acts impulsively Works toward clear goals Produces “magic” Produces predictable behavior Why Developers Are Excited About This

Especially in larger projects like:

• Laravel applications
• React frontends
• Python automation
• microservices
• legacy codebases

you do not want AI randomly touching everything.

This approach makes AI:

• safer
• more predictable
• better at collaboration
• less likely to create technical debt

A Real-World Example

Without Skills

Developer:
"Fix this bug"

AI:
- changes 17 files
- updates dependencies
- rewrites components
- breaks tests

With Skills

AI:
"The bug is probably located here.
I recommend changing only this function.
These are the side effects.
These tests verify the result."

That difference is massive.

Useful Links

GitHub Repository

https://github.com/forrestchang/andrej-karpathy-skills

Andrej Karpathy

https://karpathy.ai/

OpenAI Skills Documentation

https://help.openai.com/en/articles/20001066-skills-in-chatgpt

Claude Code

https://www.anthropic.com/claude-code

Cursor AI

https://cursor.com/

Final Thoughts

This repository may look small and simple, but it actually represents a major shift in software development.

We are moving from:

“AI that writes code”

to:

“AI that learns professional engineering behavior”

And that is probably the future of AI development.

Not just smarter models. But smarter workflows. Smarter limitations. Smarter discipline.

Because the best AI is not the AI that does the most.

The best AI is the AI that does exactly what is needed — and nothing more.