Working with AI
Last week, I built a CLI tool for wrapping some AWS functionality, using Go. While I knew what I wanted to build, I wasn't deeply familiar with AWS's Go SDK. Instead of spending hours reading docs, I let GitHub Copilot scaffold a prototype. After fixing a few AI-generated quirks, I had a working tool in minutes rather than hours.
This experience, combined with Mitchell Hashimoto's recent Backend Banter episode on AI and our CTO's survey about AI productivity, made me reflect on how I work with these tools.
A Balanced Approach
In my daily TypeScript work, I usually know exactly which APIs I need. But when questions arise, I often turn to AI assistants like Claude before traditional search engines. They provide quick guidance that I always verify - no blind copy-pasting here. During debugging, they help identify relevant APIs and potential solutions.
When learning Go recently, I took a different approach. I disabled AI assistance initially to better learn the basics. Only later did I leverage GitHub Copilot for more complex tasks.
I have a clear line tho: if explaining something to AI takes longer than solving it myself, I tend to do it manually, instead of giving Copilot "context" by way of adding files etc. (Although this may be overcome with an initial context setup, as per this article from the VSCode Blog.
The AI Dilemma
While AI boosts my productivity, I share Namanyay's concerns about creating "illiterate programmers." The risks are real - from weakened debugging skills to lost learning opportunities. I've seen this in code reviews: AI-generated configurations that work but completely miss the requirements or aren't up industry best practices.
Then there is the movement of vibe coding - where you don't actually code, but ask AI to solve you problems. Whilst this may work for pet projects, and initial prototypes, I feel this isn't an approach that can be applied solving new problems. At least not yet.
I maintain a balanced approach: leveraging AI to accelerate development while ensuring I deeply understand the solutions I implement. These tools enhance my workflow without compromising my fundamental programming skills.
Sources:
- https://x.com/karpathy/status/1886192184808149383
- https://nmn.gl/blog/dangers-vibe-coding