How is AI increasing your productivity?

I'm using ChatGPT and Cursor right now. Overall, they work decently for generating code patterns that are common.

I have noticed that Cursor doesn't work as well if you are trying to generate code for certain libraries (Firebase in my case) or laying out UI through CSS Tailwind classes. It's not perfect, but when it gets it right, it can shave off minutes when it generates code that can handle different edge cases that you might have missed. This can really add up throughout the course of a day.

I'll use ChatGPT to bounce off higher level ideas about system design to see if I'm missing any details. It's been pretty solid in verifying my thought process and even suggesting things that I might not have thought of.

Like Sai and Charlie mentioned, it takes a certain amount of effort to make AI actually work for you when coding. AI is good for:

• extremely common code patterns
• shuffling simple stuff around (variable name changes, small function refactors)

Larger design decisions that span multiple functions/scripts don't work very well. Similar to how Alex mentions breaking down units of work, breaking down work even more with atomic tasks helps you get a way better ROI on AI tools. I use this process:

• break down tasks to the function level
• insert some comments to annotate what I want Cursor to do
• use the chat interface and feed in context for each mini-task
• audit your code by looking at it/testing it

Rinse and repeat.

The way I like to think about it is that AI is 100% useful on 20% of your work as a SWE (doing mundane stuff like regex, etc). Cursor can't help you distill requirements or make a great design. Cursor also won't help you manage expectations with your boss.

ChatGPT and Copilot are the big ones. There are folks using Gemini, Claude, and Perplexity as well. Most recently, Cursor (AI-powered IDE) has been popping off.

I am using ChatGPT a decent amount as I pick up Node.js back-end development again. It's been really helpful as a faster Google effectively where I can quickly familiarize myself with basic syntax and mechanics (e.g. how to write the cleanest loop, how to check for null). It gives code that I can usually copy-paste (after I verify it of course) alongside good explanations.

When it comes to the higher-level stuff though like code structure, I am relying on my own fundamentals (all programming languages are effectively the same) and doing that myself. AI is generally pretty bad at this stuff as code quality is very context-dependent, and there isn't a lot of information about deep optimization for it to train itself on (this is the kind of stuff that's gatekept by working at top companies like FAANG).

As for 20% - 40% boost, I am skeptical, but it depends on the organization. If your engineering organization doesn't really care about quality and the blocker really is just code velocity instead of something like lack of alignment, I could see that. Otherwise, not really.

Here's some other good resources to go through:

• "What are the top 3 generative AI tools/use cases that can boost productivity?"
• "How Can Junior/Mid-level Engineers Effectively Integrate AI into Their Coding Workflow?"
• "Are there any AI tools that can validate design approaches?"

One thing I am having success with is looking for making a tradeoff decisions with technologies I am not familiar with. For example, trying to evaluate react redux. Or tradeoff decisions between different types of rendering.