I've run into these problems when building with AI: - AI loses context - need a bunch of project rules, cursor rules, system prompts but the AI still doesn't follow them sometimes. - AI can't fix bugs - telling it "that didn't work" 100 times doesn't work - stuck at 60% - hard to make progress once project gets more complex, it add one feature but breaks 3 other. So i've decided to build a cursor clone (which is really a vs code clone) that has - better context management - AI maintains context of project scope, feature requirements, task breakdown, and testing. - visual debugging - AI traces code execution and isolates issues much better (equivalent of putting console logs everywhere, but you can see where the bug is) - native integration with supabase - a lot of the issues I run into are database related, so this should help. It's working very well with sonnet 3.7 & deep-seek, but it could just be my own test cases. I'm looking to have a few people test this. Also, should I open source this?? If you are interested, leave a comment :)

I have spent the last few evenings finishing a site I have had for ages. This is a complete revamp and it is all written in Typescript. I didnt know Typescript until I took Brandons Full stack AI Marketing Platform course ( I really should finish it, im staill at 43% and have been for months) So I have created the site infops.vercel.app using React, Typescripot. Hosted on Vercel with a Neon database backend. Drizzle-kit for ORM. Drizzle is so nice, it's simply a case of modifying your drizzle.config.ts to switch to a different DB. The articles are created via a python backend which intefaces to Crews and used API calls to LLM's for deep research. It then packages them up, feeds the Markdown to the DB, and et voila, we have a working site with content. I just need to get a little script to run to see what the news of major news story of teh day is and i will feed that into the script to fully automate the site. All in all the whole site has been created utilising AI, from Cursor for code. Crew for Information retrevial, Python for stuffing the database. Im not 100% happy with it as it is text heavy and image light. But hey, evwery product has a lifecycle. Hope you like the content. Influence Operations is close to my CyberSecurity heart.

I'm thinking of building a MCP server to dynamically retrieve latest docs. Which libraries would you like to see supported/indexed first? I find that even with the latest models, AI always incorrectly changes my OpenAI or NextJS related code. For example: - it ALWAYS changes gpt-4o-mini to gpt-4. - it often changes client.chat.beta.completions to client.chat.completions.create - it always uses Next 14 conventions when my project is using Next 15. You could put it all in .rules file, but that will inject it into every single prompt which could decrease the answer quality because its not always relevant and bloats the .rules file when you have many libraries you are working with.

Ask AI to ONLY keep the necessary changes after a few turns to fix something This only applies if there were multiple edits to fix it. Chances are, most of the code the AI just changed weren't necessary, keeping them will bloat the project and worse, confuse the AI in the future.

Hey all. I have just thrown up a site that i am working on. it would be handy if some of you could just take a look. Im not looking for much. any feedback would be great. And yes I know the landing page images are dark. I have an issue logged in my git hub to address them . https://nc500-mvp.vercel.app/ Thanks UPDATE: Added a pic of some Analytics that have come through. I geek out on this stuff