reddit · SideProject · 10-17 15:50 47 分 ← 返回列表

I got tired of AI's "yes-man" behavior, so I built a 3-step workflow and a unified gateway to fix it.

The more I use AI (mostly DeepSeek), the more I realize a brutal truth: AI is just a mirror of the user's logic. If your thinking is all over the place, the AI will just give you fancy-sounding garbage. If you don't know how to verify what it spits out, you aren't building anything—you're just using a glorified, expensive search engine. I used to get burned all the time by code that looked perfect but completely broke in production. So I stopped trusting the "yes-man" behavior and force

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 会迎合用户、给出看似正确却难以验证的答案，容易导致代码在生产环境出错；同时，多个模型 API key 和插件也让人难以管理。

目标用户：使用 DeepSeek/Doubao 等模型进行编码的个人开发者、独立开发者和小团队

证据：I used to get burned all the time by code that looked perfect but completely broke in production. / Without this framework, AI just makes you a faster producer of bad decisions. / I built my own unified gateway (NieToken) ... I'm throwing in $0.5 in free credits for anyone who wants to test this workflow out.

变现：统一模型网关按 token 抽成或订阅；把三步验证流程做成 prompt 模板包、浏览器插件或 IDE 插件，按席位收费

MVP：两周内做一个轻量 Web 或 CLI 工具：输入代码或方案后，工具自动生成“事件反方（Devil's Advocate）提问 + 依据文档并附引用 + 输出置信度打分的三重校验结果”，并支持填入自有 API key 对接 2-3 家模型，先免费收集使用数据

竞品：OpenRouter、LiteLLM、Portkey、Cursor 内置规则/自定义 instructions

摘要：作者用 DeepSeek 时发现 AI 会一味迎合用户，导致代码看似正确却在生产环境出错，于是总结出三步验证流程，分别为事件反方复核、仅依据官方文档作答、输出 1-10 置信度打分。他还顺带推广自己做的统一网关 NieToken，提供 $0.5 免费额度，用于在文本任务中调用 DeepSeek、在图像任务中调用 Doubao。目前没有付费或收入数据，付费意愿只能靠免费额度试探。