开发者与大模型生态里被反复提出、平台迟迟没做的能力缺口。Capability gaps raised again and again inside AI developer communities.
我最近在几个 AI 编程工具之间切换，上下文和记忆的部分开始让我烦了。Claude Code、Codex、Cursor、Windsurf 处理项目上下文、规则、记忆和历史的方式都不一样。Claude Code, Codex, Cursor, Windsurf all seem to have slightly different ways of handling project context, rules, memories and session history.
调 LLM 时我们经常要先从文档里抽出文本，而文档可能有 20 种格式、有时还要 OCR。有没有一个 API 直接帮你做完，好让我们直接喂给模型？We need to extract a text string from documents… 20 different formats and sometimes needs OCR. Is there an API that does this for you?
要让我的聊天机器人真正可用，我需要一种方式让用户预购额度：类似买一张只能在你这使用的 ChatGPT 礼品卡/兑换码。I really need some way for users to purchase usage for my chatbot… a way to buy a ChatGPT Card or Gift Code that can only be used by my service.
这是大家自发 brainstorm 的地方：更清晰的文档、更详细的日志与错误诊断、更高的 token 上限……Share ideas like clearer documentation, detailed logging and better error diagnostics, expanded token limits…
用户会输入猫狗品种名，我要在 400 多个已知品种里找最接近的。比如输入「bull dog」，就要命中对应的标准品种。I want to resolve it against a large list of known breeds (400+) and pick the closest one from the list.
功能上 Cursor 做得很好，数据、比较逻辑、组件、PDF 生成、测试都没问题。问题在视觉设计——无论我把 prompt 写得多细。Functionally, Cursor is doing a good job… The problem is visual design. No matter how detailed I make the prompt.
我们的服务器调 API 再把结果渲染给用户，合作方无法判断这些内容是我们自己写的还是模型生成的，而这些结论是有后果的。3rd parties we're working with have no way of knowing if we generated those responses ourselves… there are implications.
建议开放基于邮件的交互：用户发一封邮件就能开启对话，模型把回复整理好后以邮件发回。Users could send an email to initiate a conversation, and the model would respond via email once a reply is formulated.
我经常在手机的语音模式和 Codex 之间来回切换，把语音里想清楚的需求复制过去。真希望不用来回切。I find myself switching between ChatGPT voice mode on my phone and Codex a lot to copy and paste requirements I've brainstormed.