That was a long preamble to this question: any senior devs out there (20+ years) who enjoy using Cursor? What’s the trick?
I started programming professionally around java 1.3, and the greybeards around that time were talking about how they had moved to OOP. Most disliked it but the people who didn't adapt got stuck in dead ends. (Stay in a dead end long enough and it becomes lucrative - see Oracle DBAs and COBOL developers - if you can stay there long enough!)
You absolutely have to treat coding with LLMs as a new skill. It's just like learning a new editor to the same precision that you know Emacs or Vim. It's learning the rough edges - and the rough edges keep changing. As a senior, it's frustrating! However, as I learn more, I learn the necessary concepts to utilize what is good and try the frustrating part in a month or two when it has gotten better and so have I.
I've spent a career reading others' code, so that doesn't bother me.
Now, I prompt through TDD for code that matters and that has helped stop many of the problems I see others face. It's a little slower as an iteration loop but it generates the test data well, it allows me to make sure I understand the problem and the prompt, and it forces the changes to be small, which increases success.
If I have to make a non-trivial change to the code, I know it will try to rewrite it if I don't start a new agent session. So, I'm liberal with creating new sessions.
It is much better at small problems than large problems, which makes me cognizant to keep my problems small. Not only is that good for the LLM, it's good for the codebase.
In some cases, it's better to have the LLM create a script to make the change than try and have it go through the code with a prompt. It's very good at simple scripts.
And I'm not afraid to discard it when it's not valuable at the moment.
But the keyword here is "simple and concrete". I do not understand why people expect those tools to be good at tasks that involve having years of context on something. It really makes you think that the people that say that AI will replace SWEs have never actually used any of these tools.
Then paste that output into Cursor with Claude 3.7 and have it make the actual code changes and ask it to build/fix errors along the way with yolo mode enabled.
The 2-step process is a lot better since Grok can refer to the entire context of your codebase in one shot and come up with a high quality implementation plan, which is then handed off to Cursor to autonomously make the code changes.
Claude 3.7 is excellent, and better at coding but I appreciate the context size of Grok and feel like I get better bang for my buck for general purpose research too.
Mostly you have to guide it a lot in verbose, planning methodology, often chained via other LLMs. It is like working with a junior developer - no doubt about that. But that's still really good for $20/month.
Coding is not my full time job and it really hasn't been more than 20 hrs/week since my time in FAANG; I do a lot of statistical work, IT, and other stuff too. Maybe that's why I like the mercurial help of Cursor since I am not using it 50 hours/week.
- Gemini: refuses to work 20-30% of the time - Sonnnet: does 50% too much work
Ask: previously was chat, and just tries to answer questions. Does not have the capability of editing your code directly (although if it provides a snippet, you can always click to apply it to the code).
Manual: previously was composer in standard mode. Can edit code across multiple files, but only works one prompt at a time. So if you ask it to edit tests, it will do that and then wait for your next input.
Agent: previously composer with agent mode enabled. Same as manual, but can figure out next steps and automatically execute them. For example, it can edit the tests, then run the CLI command to run the tests, then edit the code again if there are test failures, and repeat.
I find agent to be most helpful when you know the end goal but you need to be clear about what you want. Tell it things like “run the tests to make sure they’re working” and “search the codebase for where this class is used”.
I find manual best for when you know what small steps to do. Like, “create a helper class for managing permissions”, followed by “write tests for the profile view that checks permissions”, followed by “refactor the profile view to use the helper class”.
Cursor is magic. Two massive use-cases for me:
1. Autocompletion of a copied function.
So I have some function, modX(paramX), and I need to duplicate it to modY(paramY). They're near enough the same. I love that I can copy/paste it, rename a variable or two, and Cursor then intuits what I need. Tab tab tabbity-tab later, job done. This might not be the most amazing use of AI but is sure as shit helps my RSI.
(I know I should abstract my functions yada yada.)
2. Generation of a new function where I have no idea how to start.
I tell the prompt what I need. "I'll give you a bunch of Markdown files and I'd like you to take the frontmatter properties and put them on an object then take each of the headers and its following content and put that in to this property on the object".
It'll do that, to 90%. I'll fix the 10%, making sure that I understand what it's created for me.
This would have taken me 4 hours. With Cursor it takes 15 minutes.
In my experience, the solution to low traction languages and frameworks with lots of boilerplate and busy work coding is to use higher traction languages. I much prefer to grapple the problem with my brain in a tighter way than to attempt shaping a loose language with a looser one...
My prompts usually resemble actions I could tell a college student — they just have a better understanding of concepts and professional lingo.
The benefit of this approach is that you know the code fairly well. You are staying with the LLM in developing a deeper understanding of the code you'll ultimately create a PR for. Then when there's an incident, you have enough deep knowledge of the code that you can be tactical.
I have found that until I trust AI to develop the code unsupervised, I have to have an equally good mental model of everything AI makes.
Other times, I’m just jumping into a quick feature of bugfix ticket that’s nothing particularly fun or interesting - so it’s more fun to basically roleplay overseeing a junior dev, without any of the social pressure of interacting with a real person, and not have to get my hands dirty with boring tasks.
It’s all about finding the opportunity to have fun with it.
I feel like a lot of time is being wasted debating how serious and how real it is - but are you having any fun with it?? Cause I kinda am.