AI coding has made me dramatically faster. But I’m starting to think we’re creating a completely new category of problems

开源代码

热度趋势

趋势数据积累中

百分比基于当前可用热度信号，而非评论数或独立用户人数。

AI 摘要

AI 编码代理显著提高了开发速度，但也带来了一系列新问题。开发者面临的挑战包括在新 AI 会话中需要反复解释产品、代理无法理解某些设计选择背后的原因，以及需求在不知不觉中发生变化。其他令人沮丧的问题还包括代理推翻或与之前会话中的决策相矛盾、不确定项目是否真正完成、半成品或已被放弃，以及在数周后回顾代码时难以理解当初的实现方式。…

Hi everyone,

I’ve been building more and more of my products with Claude Code, Codex and other AI coding tools. The speed is ridiculous.

Something that might have taken me a week before can sometimes be built in a day. But the longer a project survives, the more I notice the problems changing. The hard part isn’t always writing the code anymore.

Things that bug me a lot are like:
 - Starting a fresh AI session and having to explain the product again.
 - The agent understanding the codebase but not why something was built a certain way.
 - Requirements slowly changing without you realizing it.
 - One agent undoing or contradicting decisions made in another session.
 - Forgetting whether something is actually finished, half-built, abandoned, or just discussed.
 - Coming back to something you built weeks ago and thinking: **why the hell did I do it this way?**

AI confidently changing something that technically makes sense but is completely wrong for the product. Context windows filling up right when the agent finally understands what you’re building. Having decisions scattered between chats, prompts, commits, notes, and your own memory.

And I’m curious whether this is just how I build, or whether other people building heavily with AI are running into the same thing.

So I want to ask the vibe coders / AI builders here: What are the problems you repeatedly face when building real products with AI?

I’m especially interested in the annoying problems that start appearing after you’ve been working on the same product for weeks or months. The stuff that makes you think: “There has to be a better way to do this.”

Could be context loss, debugging, security, testing, requirements, managing multiple agents, remembering decisions, deployment, maintaining old AI-generated code — anything.

I’m currently collecting these problems because I’m building something, but I don’t want to build based on assumptions about what vibe coders need. I’d rather hear it directly from people actually doing this every day.

What’s your #1 frustration with building software using AI coding agents right now?