Re: excessive comments. I have a prompt that lists a few criteria that agent must follow "otherwise human will not accept your work" and one of them is to clean up comments and it works fairly well.
Thanks for sharing this. I've been using Claude Code extensively and based on your example prompt, maybe I should spend even more time curating prompts ;-)
A few things:
You mention having it come up with a plan, this is indeed essential.
Claude code can write Github issues or just write a MD document with the plan, which you can refine together.
Cutting it loose with the plan has unlocked more complex usecases for me where it will do big refactorings or implement features in a more contained way
You mention that heavy users spend about $50 per month in tokens. I feel seen. I've spent that much in a day or so, but I admit I'm not using it purely for coding but also for creating workshops, testing Kubernetes cluster ideas (those logs add up in tokens).
I'm a solopreneur, but the cost is totally worth it to me!
Refactoring is a double edged sword. Sometimes it amazes me, especially in frontend / JS tasks, but when it comes to static typed languages like Java it often leaves me annoyed how it won't even compile and check if it got everything.
Also, I feel some kind of MCP to handle the refactoring itself would save so many tokens here.
Great post! Random question - how do you handle having both Cursor rules and Claude.md? Do you have a way to share context between them? Or do you find different agents need slightly different context?
When I want to add new rules, I have the agent write them to both places. If others aren’t doing that consistently, you could set up a GitHub Action that watches changes to the Cursor rules or CLAUDE.md, then runs Claude Code with the --print flag in non-interactive mode. It can one-shot an update using the allowed tools you pass in, then open a PR and ping you for review.
They are basically duplicates. I think the Cursor rules are more fleshed out but that's just a side effect of the people using it being more into that stuff. Whereas I find with Claude it's often easier to just tell it what to do than it is to write lots of rules.
The part about the coding style always makes me wonder. Reading of the code what's always big (if no the biggest) part of the development. The easier it was to parse, understand, the faster people would be able to change it. Dense logic and huge-classes / methods are not user-friendly. Also, some layout metrics rules are a clear indication, that some stuff is getting too big and some kind of extraction/refactors is needed. Both of those things seems to be hurting right now.
I like to learn to use it. Do you have any online resources for it? I use Cursor currently. However, I like to Sublime Text interface a lot. Have been trying to stick with it. Currently juggling between the two.
Oh great so just fuck junior devs right? Yea real exciting and not dystopian time we are living in. I hope you get laid off so you can just SEE how much AI has DESTROYED the job market!
@Alex, this is by a wide margin the most practical and best-written AI use in coding article I've read. Hat tip from me.
Thanks Tom! If you implement anything from the article and still think that, let me know 😂
Re: excessive comments. I have a prompt that lists a few criteria that agent must follow "otherwise human will not accept your work" and one of them is to clean up comments and it works fairly well.
Thanks for sharing this. I've been using Claude Code extensively and based on your example prompt, maybe I should spend even more time curating prompts ;-)
A few things:
You mention having it come up with a plan, this is indeed essential.
Claude code can write Github issues or just write a MD document with the plan, which you can refine together.
Cutting it loose with the plan has unlocked more complex usecases for me where it will do big refactorings or implement features in a more contained way
You mention that heavy users spend about $50 per month in tokens. I feel seen. I've spent that much in a day or so, but I admit I'm not using it purely for coding but also for creating workshops, testing Kubernetes cluster ideas (those logs add up in tokens).
I'm a solopreneur, but the cost is totally worth it to me!
Refactoring is a double edged sword. Sometimes it amazes me, especially in frontend / JS tasks, but when it comes to static typed languages like Java it often leaves me annoyed how it won't even compile and check if it got everything.
Also, I feel some kind of MCP to handle the refactoring itself would save so many tokens here.
My 2c
Great read
Great post! Random question - how do you handle having both Cursor rules and Claude.md? Do you have a way to share context between them? Or do you find different agents need slightly different context?
When I want to add new rules, I have the agent write them to both places. If others aren’t doing that consistently, you could set up a GitHub Action that watches changes to the Cursor rules or CLAUDE.md, then runs Claude Code with the --print flag in non-interactive mode. It can one-shot an update using the allowed tools you pass in, then open a PR and ping you for review.
They are basically duplicates. I think the Cursor rules are more fleshed out but that's just a side effect of the people using it being more into that stuff. Whereas I find with Claude it's often easier to just tell it what to do than it is to write lots of rules.
"Courage to ask" is not optional anymore. Need to drop the "ego" now.
The part about the coding style always makes me wonder. Reading of the code what's always big (if no the biggest) part of the development. The easier it was to parse, understand, the faster people would be able to change it. Dense logic and huge-classes / methods are not user-friendly. Also, some layout metrics rules are a clear indication, that some stuff is getting too big and some kind of extraction/refactors is needed. Both of those things seems to be hurting right now.
Sublime Text user here. How do you use it with Sublime Text? Not able to find any official plugins for the integration.
There's no plugin needed. It runs in your terminal and edits your files; you read them in Sublime.
I like to learn to use it. Do you have any online resources for it? I use Cursor currently. However, I like to Sublime Text interface a lot. Have been trying to stick with it. Currently juggling between the two.
Oh great so just fuck junior devs right? Yea real exciting and not dystopian time we are living in. I hope you get laid off so you can just SEE how much AI has DESTROYED the job market!
Good post! thanks for sharing your experiences. mine are similar.