LLMs can't think. It's pretty much useless to try.

Cursor + Claude

You can't ask a nailgun to build a house.

Learn to code? 😂

Pretty much the best solution I've found is to break down your project in the small tasks. And tackle it one by one. Also, claude is much better than GPT for this. You can upload your project files and documentation so it can reference your code. Also, if you're comfortable with it, you can run Claude code directly on your computer so it can reference your entire project.

outline and planning.

chatgpt needs all your requirements well documented, you can't just start asking for shit, its gonna make up what it hasn't been told is your goal. Typically devs don't plan because they have to learn the plan as they develop the real requirements as they go along.

llms work best with too much input and if you already know 99% of the tech stack and patterns and deployment env. you'll use.

they save you the typing, not the knowing what is needed.

I feel this so much. Occasionally the thing will lose all context and you end up with duplicate files, random weird folder structures and files that don’t have any consistency to them. I have yet to figure out how to make it do as I ask.

also, chatgpt seems to give me 1000 words no matter what its always 1200 words or less

Hopefully, "vibe coders" don't become like the cashier that can't make change without a cash register. nostr:nprofile1qqstnem9g6aqv3tw6vqaneftcj06frns56lj9q470gdww228vysz8hqpzdmhxue69uhkzmr8duh82arcduhx7mn9qy2hwumn8ghj7etyv4hzumn0wd68ytnvv9hxg8mv947

nostr:nevent1qqsfeuhmaawx9q8wtkkt3x5ys4nq37u486hmeum4e6jd6260hyj3r0gpp4mhxue69uhkummn9ekx7mqzyzu7we2xhgry2mknq8v7227yn7jguu9xhu3g90n6rtnjj3mpyq3acqcyqqqqqqghmle7x

I’ve had better luck with nostr:npub17304velluajf6lylvjynpj2f3ndg396w063gj2gef5qk0nwtcyjqfj9yky and the Deep Seek model. But I only use it for very simple coding (not a coder).

Have you tried Replit? I just started experimenting there. It writes the codes in sections and you can revert back to prior milestones in the project. I used ChatGPT to build my initial prompt for Replit.

Don’t use chatGPT, for one…

Second, use an IDE like windsurf

Third, you should highlight code or mention files you want adjusted, then explicitly write what actions you want it to take. Your prompt is the most important part. Also, reject code and tell it that it fucked up, and try again

Still looking to tame all of this in Visual Studio.

I've been treading lightly but am optimistic it WILL happen, but doubtful it's available now.

I’ve found it necessary to ask very specific questions regarding atomic pieces of functionality; “write a typescript function that does X.” And then a few follow up iterations usually gets me within spitting distance of what I’m looking for.

Raycast has some useful prompts for code questions you can copy/ paste even if you don’t use Raycast. https://ray.so/presets/code

Us o3 in ask mode to plan the stuff out, then use gemeni 2.5 to implement . If your using cursor flip buttons and switches to widen chat context, setup auto-run and enable mcp tool use. Create project rules to limit hallucinations, create a custom agent to limit hallucinations, and create an agent rule which references an md file in your project which tracks success's and failures.

Either that or just use o3 for everything with really good prompts.

as others have suggested use an agent based ide as cursor/windsurf/vcode

cursor out of the box gives you decent performance, if you wanna go deep vcode+roo is a deep rabbit hole specially if you dig into advanced setups like the SEC playbook https://github.com/jodobear/sparcaroo

Sounds to me like it's trying to be independent.

Cline w/claude, Cursor, or copilot agents. There are probably better answers for how this works, but with copilot agents it breaks work down into small pieces and feeds it back to the LLM so it itterates on smaller components isntead of trying to do the job all at once. In lay terms: I think the models understand the constraints of the output windows size and try to get entire thing done in one sweep, they currently have no concept of history or memory, that's where tools come in to help break the work down and keep feeding it pieces.

Lol vibe coding, because we had such a shortage of technical debt on any web application already

First assess problem.

Break down tasks

Architect solution first.

Define spec.

Implement to the spec.

When instructing for modularity, focus on input, output expectations not the 30000 ft view

Design is separate from writing code. When writing code ask it explicitly for single functions with specific inputs and outputs.

Write a function called fun that takes inputs x, and y. Does blah, and outputs z.

Did you try Roo code with the nostr addons?

https://github.com/jodobear/sparcaroo

Code writing I think we are still 10 years behind before AI can partially replace developers

Try Replit

No idea what works well with medium+ size. Pretty much you need to be very specific and still it struggles.

goose

Just prompt for it not doing that.

use goose

Cline (open source vs code extension)