I've been using Claude code in my terminal workflow, where I write, compile, and test code. However, I've noticed that with my $100/month subscription, Claude tends to stop working effectively after a short period—it delivers only basic results or suggests I ask for improvements. Even when I run it in dangerous mode, I expect it to work for longer than just 20 minutes. I often find myself needing to prompt it to try again or enhance its unfinished work.
I'm thinking about a few options to improve the situation:
1. Upgrading to the $200/month plan to see if it performs better.
2. Using Claude code alongside another Claude code MCP server instance to create a loop where it checks its own work and retries if necessary, while also getting feedback from the server.
3. Exploring alternatives, like Aider, though I'm concerned about whether it will provide better results within my budget.
4. Considering a task MCP server to increase the number of steps and improve performance.
5. Running a script that lets one Claude code instance check the output of another; if the second instance finds the results lacking, it would prompt a retry.
What do you all think of these ideas? Any additional suggestions for achieving long-lasting and refined performance on a budget?
1 Answer
I think a lot of your issues might stem from how you're prompting the AI. Just upgrading your subscription may not solve things if your prompts aren't hitting the mark. Have you tried refining your prompts further? The 200/month plan will increase your limits, but if you're not using those limits fully, it might not be worth it. Focus on improving your prompting techniques to get the most out of what you have!
I see what you're saying, but I really believe I've put a lot of effort into crafting my prompts and even experimenting with meta-prompting. I've noticed the lower tier tends to hit the Opus limit quicker, which can make a difference in the results you get.