ruburi
@ruburi
ok i can conclude my late night coding session today with this statement: either i suck at prompting, or claude code cant do anything that goes slightly above junior level development im kinda scared to leave a codebase to llms. i understand that maybe i dont use its full power, but im not a beginner in prompt engineering and try to write instructions as clearly as possible, involve opus to write comprehensive architecture plans etc -> still hard to follow basic logical principles. as long as there is no agi (logic) -> codebases can be easily messed up. even when im considering leaving the agent to do simple tasks like cleaning up some code, it can forget functions it wrote a minute ago or start overcomplicating things that said, i will proceed making the mini app without using claude code or copilots. i find llms extremely useful when learning new information quickly or summarizing concepts -> must have ps: spent $200 on api calls. maybe not much, but it gave me a good understanding of how things really work. yeah might go at a slower pace, but steady
2 replies
2 recasts
15 reactions
sirouk
@sirouk
Instead of jumping right into writing code, flip it to ask mode. Then have a breakdown and read me that explains the project and then from there beside the framework that you’re going to use, dependencies. Essentially making a PRD. Then ask it to write a roadmap complete with milestones tasks, and sub tasks. Then have it go over the road map it just wrote again and breakdown. Anything that’s over a six level difficulty on a scale of one to 10, inclusive of milestones and tasks and sub tasks. Have it write that to a document by flipping it out of ask mode back to agent mode. Then sequentially follow the plan and have it check off each sub task and task and milestone as it goes through. But don’t just go through one by one without creating a simulated pair programming approach where there’s a test for every piece of code written.
1 reply
0 recast
0 reaction
ruburi
@ruburi
ive been doing all of that. there were no very complex tasks like make the whole backend -> use this and that. i broke down each part, instructing it step by step with CLAUDE and README with the project description and roadmap -> doesnt matter + overcomplicating simple bugs occasionally (dont think i need to write breakdowns on how to fix bugs too) my interaction wasnt like -> make the whole app in one prompt. there were around 100 to 200 iterations with the agent
0 reply
0 recast
0 reaction
