返回首页
最新
My company wants to organize and classify their documents primarily to guard against people uploading sensitive documents into AI or emailing them to the wrong recipients. Need to go beyond just credit card or ID detection, many documents are sensitive because of what they are e.g. Legal agreements, NDAs. Any good tools out there to recommend?
# Hello World(valid terse; for Show HN)<p>I'm terse(greetings)<p>## Container<p>"A container may have one text block, single or multi-line"<p>Object(Semi-colon separated attributes; variables_like: 123; "Traditional unicode strings allowed here too.")<p>Another object with a name(multiple objects under a container)<p>## Further description of TERSE<p>"""<p>TERSE is a unified way to store state, with a defined way to query and mutate it that has been relentlessly refined for overall simplicity and token efficiency not just at rest, but in operation.<p>We realized state access was a low-level, common pattern that keeps getting reinvented. It goes deeper than just AI memory. Consider that Anthropic memory and others are a bunch of markdown files and in some respects, they got it right... simpler is better.<p>So REALLY the problem is settling on a format that is flexible to cover 95% of a domain. We can do so MUCH better than Markdown; TERSE is opinionated on that but it has very good reasons.<p>By being a line-ordered and line-identifiable structure (with special treatment for free-form blocks like this), now the whole state can be idempotently re-declared. Literally, this text of this post is a TERSE state declaration that an AI or you could make. Or, duh, just copy this to a .terse file!<p>So let's say you did that. To query this state:<p>> ? Hello World<p>To copy this whole state somewhere:<p>> # My new container<p>> ? Hello World //inline queries physically expand into the declaration<p>Directives are a [tail] with composable set of operations. For example, let's clean that up.. I declare thee removed!<p>> # My new container [REMOVED] // all contents too!<p>There's more. TERSE makes it exceeding easy for the AI to do stuff with state.<p>The MCP has one main op where these two things are sent *together*<p>1) state declarations<p>2) state queries (post-application of declarations and results unioned)<p>BTW, to query all state, wait for it...<p>> ? // Not that you want do this often!<p>TERSE has a full object model for (de)serialization.
A reference Python implementation of the full specification is also included.<p>"""<p>## Example apps included(to help get you started; in the monorepo)<p>### TERSE MCP("store anything you want locally for your AI")<p>### TERSE Memory(a generalist; modify your own)<p>### TERSE Brain<p>"A Karpathy-brain compatible API but with TERSE as the backing store. 1/6th token use and 1/8th the number of tool calls!"<p>### Misc<p>TERSE Browser(query-response UI; great for testing)
VS Code syntax highlighter(vsix included)<p>## Reminders<p>Terse is pre-release(alpha)
我不是软件工程师，但我一直在进行一个实验，看看代理开发是否能够在现有的SaaS代码库上生成一个有用的原型。<p>这个代码库有超过100万行，已有15年历史，托管在Azure上，主要使用C#和React编写。<p>原型需要在九月份提供给客户进行测试。虽然没有开发人员能够全职参与，但我偶尔可以获得一些特定技术问题的帮助。一名工程师将在八月份评估实现情况，并决定我们对AI生成代码的信心程度，以便我们决定如何将代码“转换”为生产级别。<p>我的问题是，在八月份之前，我可以使用AI工具和手动测试做些什么，以使代码尽可能稳健和可审查。我希望增加AI生成的代码质量，以便将其推向生产级别的路径更接近“复制粘贴”，而不是从头开始重新构建一切。<p>环境
原型正在一个独立的分支上开发，部署到一个独立的内部环境，并连接到自己的数据库和架构。<p>规划过程
我们在六月份采访了客户，并将结果转化为MVP规格。<p>这个过程大致如下：
1. 编写PRD（产品需求文档）。
2. 使用大型语言模型（LLM）将PRD转换为架构文档，并由架构师进行审查。
3. 创建产品设计，包括屏幕图像和包含交互细节及边缘案例的md文件，使用Claude Design。
4. 使用代理将工作分解为史诗任务，使用PRD和架构文档作为指导。<p>这些史诗任务是最细粒度的规划文档，经过我和架构师的审查。<p>开发过程
开发流程旨在尽量减少干预：
1. 规划代理将史诗任务转换为md故事文件和Jira故事。
2. 编码代理实现故事，包括测试，并打开PR（拉取请求）。
3. 审查代理审查PR，要求更改，并将其合并到原型分支中。<p>编码代理会轮询PR以获取审查意见，并可以将问题升级回规划者。<p>代理还有一个“停止并询问”的列表，用于他们不被允许自主做出的决策。我（以及几次一名工程师）通过解决这些升级问题并每天手动测试累积的更改来参与其中。<p>大多数实现和初步审查都是通过基于Claude的代理完成的。对于风险较高的PR，我还使用Codex作为审查者。第二个模型的审查在Claude生成的代码中发现了更多相关问题，但令牌配额限制了使用。<p>我还对代码进行了单独的重构和加固运行。<p>结果
规划、设置环境和代理流程大约花了两周时间，然后代理在大约两周内构建了整个MVP。从规模上看，它包含了13000行功能代码，以及相同数量的测试代码。<p>我希望得到的建议
假设在审查之前我无法获得实质性的开发人员参与，我该如何增加代码尽可能接近生产级别的可能性？<p>以下是我一直在思考的一些问题：
1. 在工程师审查代码之前，哪些检查或开发循环会带来最大的信心提升？
2. 你会如何使用独立的代理或模型来降低编码者和审查者做出相同错误假设的风险？
3. 测试是否应该由与实现代码不同的代理生成？
4. 什么文档或证据会使最终的工程审查更快且更可靠？
5. 如果你只有几周时间来改善这个原型，然后交给工程师，你会优先考虑什么？