Comment by englishrookie
4 days ago
Would you mind sharing your setup (LLM model, IDE, best practices)? Personally, I'm struggling to get value out of Continue.dev in VSCode (using Gemini 2.0 Flash by default, with the option to switch to more advanced models). I still revert to pasting code into ChatGPT chat window (using the website), frequently.
Are you using agentic features, given that you have not just one but two PMs?
The biggest tip I can give you is to stay in the framework you are most convenient in, and have the most experience. Start building stuff the way you would by yourself, but then start delegating the repetitive tasks to an agent. My best recommendations would be using Cursor in Agent mode, and switching to VScode in Agent mode when your credits with Cursor run out. The reason why I like Cursor more is because of the Checkpoints. And VScode Copilot Agent checkpoints suck; but you can still use git to create your own checkpoints (git add / git stash, etc).
I don't even use completions, really just agent mode. I do planning, wireframing, creating specs all with agents. Even small MVPs created in 5 minutes, deployed in 10, during a meeting to just brainstorm. As for the models. Go with Claude 3.5 or 4.0, GPT5. Use sequentialthinking and Taskmaster MCP. I could write a book about it... but the best way to go about it is to dive into, get frustrated, push through and then learn it the hard way. I started delegating a lot of my programming work the day ChatGPT came out; just copy and pasting, and since that day, my reliance on AI has just been increasing, and I have been getting better at it (and now I am at this stage.. with 2 PMS).
Also to add another point is that if you felt like an agent did not help you correctly, or way overshot, did too much edits, etc. Go back to the original prompt, rephrase it - sometimes you need 1-2 times. Sometimes the model just don't work for your workflow. It can become quite delicate.
One of the bigger things is when you introduced some bug, start working backwards with the agent, simplifying whatever you build to its bare necessities, and the moment it dissapears, start a new chat, and build it back up to what it was before (in the desired non-bugged state). This often works if you then also switch to a completely different model.