ChatGPT-4: The Developer's bionic Upgrade 👩💻+🦾

ChatGPT-4 excels not only at assisting with boring, straightforward development tasks, it could do much more. However, to tap into its full potential, it may be beneficial to consider some ideas beyond OpenAI’s “GPT best practices”. Here's a brief summary of my personal experiences and learnings from the recent workshop (thanks again Rinat Abdullin ), that may help you too.

1. 🎯 Write precise instructions

Instead of just writing “check this code”, tell it exactly what you want.

“Tell me where this code does not follow SOLID principles.”        

1. 📦 Specify Requirements

Detailed specifications yield better results.

“This JSON represents a USER DTO that consists an email address and a date of birth. Return a JSON where date of birth is before 1990-01-01. Do not translate parameter names.”        

1. 🧑 Use Personas

Make ChatGPT your coding buddy.

“Act as a professional C# senior developer that masters SOLID principles and produces clean, maintainable and readable code.”        

1. 📘ChatGTP Custom Instructions

Use the "Settings/Custom Instructions" in ChatGPT to consistently guide its behavior. You could set something like:

“I'm a senior software engineer, do not explain everything like I'm a five year old.”

Additionally you may tune the respond with something like:

“Take a deep breath before responding. Don't be chatty and remain neutral.”

1. 🤔 Make it “think”

ChatGPT works best when you give it time to “think” about the problem. Just asking “Is this code correct?” will eventually result in “Yes, it is correct.”, but not explain why it thinks it is correct. However, further asking can be really useful:

“Explain to me why this code is correct, line by line.”         

1. 🎯 Tell it exactly what you want

Like specifying requirements, telling ChatGPT exactly what you want it to do leads to better results.

“Please refactor this code to comply with SOLID principles and clean code standards. Take a deep breath and make it maintainable and readable. Explain the result to me.”        

1. 🐡 Memory-of-a-goldfish (or context is limited)

ChatGPT's context isn't infinite. If the context exceeds its limit, it might lose prior information and "forget" context already provided. E.g. if you start with an Persona at some point it will forget about this, especially when pushing large prompts (e.g. pasting large code blocks). As of this writing ChatGPT4 supports max 8192 tokens. According to OpenAI-Help this is about ~= 6000 words.

1. #⃣️ Different patterns

You may use 

• “completion patterns” (e.g. Copilot line-by-line code completion), 

• “chat patterns”

“suggest refactoring for this code block”        

• “advanced patterns”

“transform this format-x data to format-y data”
“generate a Mermaid diagram from this hand drawing”
“given this date, analyse and explain patterns with charts”
"generate a visually appealing chart from the provided JSON data"        

1. 🐍 Code generation works best with Python?

I’m not confident about this, but in the recent workshop, there was a sentiment that tools like ChatGPT and Copilot work best generating Python code. One approach to overcome this might be to first produce code in Python and, once satisfied, translate it to the target language, such as Go.

I like my bionic upgrade. What’s your take on ChatGPT-4 for software developers?