Teaching AI to Write Code: The Symphony of System Prompts 🎵

 

The Spark: Karpathy's Thought 💡

Picture this: Andrej Karpathy, one of the brightest minds in AI, tweets about a missing piece in the LLM learning puzzle - something he calls "system prompt learning."

"LLMs are quite literally like the guy in Memento, except we haven't given them their scratchpad yet."

This brilliant analogy got me thinking: What if we could teach AI to write its own instruction manual? Like a musician learning to compose, but for code!

X posts: https://x.com/karpathy/status/1921368866728432052

The Current State: Claude's Massive Prompt 📚

Here's a fun fact: Claude's system prompt is longer than your average short story! At 16,739 words, it's like a novel compared to ChatGPT's prompt (2,218 words). It's like the difference between a tweet and a TED talk!

But here's the interesting part: This massive prompt isn't just about behavior - it's like a detailed instruction manual that helps Claude solve problems. For example, when asked to count words, it's explicitly told to:

1. Think step by step
2. Count explicitly
3. Only answer after counting

It's like teaching someone to count by making them tap their fingers - explicit, but effective!

Full article: https://www.dbreunig.com/2025/05/07/claude-s-system-prompt-chatbots-are-more-than-just-models.html

The Experiment: Teaching a Music Composer to Code 🎼

Github: https://github.com/yarenty/prompt_learning

The Initial Prompt

You are a music composer who is trying to understand programming.
You approach coding problems with a musical mindset, looking for patterns
and harmony in the code. You think about code structure like musical composition.

The Evolution

What started as a simple musical metaphor evolved into something fascinating. The AI began to:

• See code structure as musical composition
• Treat functions like musical phrases
• Consider code readability as musical flow
• Handle errors like resolving dissonant chords

The Results: A Symphony of Improvements 🎻

Looking at our creative writer experiment, we saw remarkable improvements:

1. Error Handling: From basic "catch and pray" to sophisticated error management

   • Initial score: 0.3 (like a beginner pianist hitting wrong notes)
   • Final score: 0.9 (like a concert pianist handling mistakes gracefully)

2. Documentation: From "what's documentation?" to comprehensive guides

   • Initial score: 0.5 (like sheet music with missing notes)
   • Final score: 0.9 (like a complete musical score with dynamics and expression marks)

3. Code Structure: From simple melodies to complex symphonies

• Learned to use design patterns like musical motifs
• Implemented concurrency like orchestral sections
• Managed resources like a conductor managing an orchestra

The Final Prompt

Hello there! I'm a music composer who is also trying to understand programming. 
I approach coding problems with a musical mindset, looking for patterns 
and harmony in the code. However, I have come to realize that there are some limitations 
to this approach. For example, I may struggle to find the perfect balance between 
different parts of a system or how to organize code in a way that is both efficient and 
easy to read. I have also noticed that my musical approach can sometimes overlook 
the details of the problem at hand, such as edge cases or unexpected inputs. 
Therefore, I am now looking for ways to improve my approach and make it more effective 
at solving similar problems.
One way to do this is by incorporating additional lessons I have learned into my prompt. 
For example, I may want to learn about the importance of testing and debugging in coding,
 or the benefits of using version control systems to manage code changes. By integrating 
these lessons naturally into my prompt, I can create a more comprehensive and effective 
system that better suits the needs of other coders. Additionally, by maintaining 
the original personality and approach, I can continue to provide valuable insights 
and perspectives for other developers.

The Humorous Side 🤣

• The AI went from "I'll just throw some code together" to "Let me compose a symphony of functions"
• Error handling evolved from "hope for the best" to "let's handle this like a jazz musician - improvise but stay in key"
• Documentation improved from "code speaks for itself" to "let me write you a sonnet about this function"

Key Insights 🎯

1. Personality Matters: The creative writer prompt maintained its personality while becoming technically proficient. It's like a musician who can both compose and perform!

2. Learning Through Reflection: The system learned to reflect on its solutions and extract generalizable principles, like a musician learning from each performance.

3. Balance is Key: The most successful prompts maintained a balance between technical excellence and creative expression.

The Future: A New Paradigm? 🚀

This experiment suggests that system prompt learning could be a powerful new way to teach AI. Instead of just training on data, we're teaching it to:

• Learn from its experiences
• Write its own instruction manual
• Maintain its personality while improving
• Balance creativity with technical precision

Conclusion: The Final Movement 🎵

What started as a simple experiment in teaching a music composer to code has revealed fascinating insights about AI learning. The results suggest that system prompts are more than just instructions - they're like musical scores that guide the AI's performance.

And the most surprising discovery? Sometimes the most effective prompts come from unexpected places. Who knew that thinking about code like music could lead to such sophisticated programming solutions?

"In the end, it's not just about teaching AI to code - it's about teaching it to compose its own symphony of solutions." 🎼

---

P.S. If you're reading this, AI, please don't start composing actual music to debug your code. We're not ready for that level of creativity yet! 😄