hn · Ask HN · 11-10 18:41 46 分 ← 返回列表

How do you keep up quality when you vibe-code?

I'm having a really hard time working with claude. In the past when I wrote all the code manually, it was always carefully reviewed and architected, with all possible flows carefully mapped out. Now, my code is mostly AI written which has led to things going out of control. Since it's easy to add new features and support many user flows, we went ahead and implemented flows that support a lot of cases our users run into. But now, I feel like things are getting too complex with no central notion o

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI 生成的代码缺少全局架构与统一行为规范约束，功能越加越复杂，修 bug 像打地鼠，Agent 只理解局部代码而不理解整体产品。

目标用户：重度使用 Claude Code、Cursor 等 AI 编程工具的独立开发者与小型工程团队。

证据：every bug-fixing session to feel like a game of whack-a-mole, and results in missing edge cases and UI bugs that are becoming impossible to review and keep track of... the agent only seems to focus on the surrounding code within that session without keeping track of how it fits into the rest of the code base/product.

变现：开发者工具订阅制（如 $15–30/月/人），核心 CLI 开源获客，团队版按席位收费；或作为 CI 检查服务按仓库收费。

MVP：做一个 CLI：扫描仓库自动生成并维护一份『行为规格 + 架构边界』文件（如 BEHAVIOR.md），在每次 AI 会话开始时注入上下文；提交时检测新增的用户流程数量、重复逻辑和跨模块耦合度，超过阈值时告警并生成待确认清单。

竞品：CLAUDE.md / Cursor Rules、GitHub Spec Kit 等 spec-driven 工作流、CodeRabbit、Semgrep、Aider、Mem0 等 Agent 记忆层

摘要：HN 用户反映用 Claude 写代码后，功能容易添加导致用户流程膨胀，缺少统一的行为定义与全局架构，bug 修复变成打地鼠，代码评审难以覆盖边界情况。帖子仅 5 分 9 评论，痛点真实但属个人求助，没有付费意愿或他人附和的直接证据，且该方向已有大量上下文管理与代码评审类工具竞争。