An engineering notebook on AI, systems, and everyday experiments.

分析OpenAI Apps SDK通过_meta域绕过context window的做法如何违背MCP设计哲学，以及协议分裂成不同dialect的潜在危机。

Analyzing how OpenAI's Apps SDK extension with _meta field violates MCP's design philosophy, creating dialects that may fragment the standard like SQL or CSS.

将常见的AI使用挫折重新定义为管理问题。当把AI当作实习生而非工具时，不可靠、幻觉和代码质量等问题都可以通过建立信任、层级分工等成熟的管理原则来解决。

Addresses common AI frustrations by reframing them as management problems. When we treat AI as an intern rather than a tool, issues like unreliability, hallucinations, and code quality become solvable through established management principles like trust-building and hierarchy.

记录在API因合规问题被拒后，使用视觉大模型从截图提取财务数据，实现十年手动记账流程的自动化。展示了本地模型、交叉验证和人机协作工作流如何安全处理敏感金融数据。

Documents automating a decade-long manual financial reconciliation process using vision LLMs when API access was blocked by compliance. Demonstrates using local models with screenshots, cross-validation, and human-in-the-loop workflows to process sensitive financial data.

GPT-5是产品升级而非单纯模型升级：新增reasoning_effort和verbosity参数，可控性大幅提升，让开发者能根据场景灵活权衡推理深度和回复长度。

GPT-5 is a major product upgrade, not just a model upgrade: new API parameters for reasoning_effort and verbosity enable unprecedented controllability for building AI-powered products.

记录通过迭代式问题解决开发自动截图功能的经历。关键经验：优化问题定义而非仅优化提示词、从"最小可行真相"出发锚定工作流、根据任务需求匹配合适的AI模型能力边界。

Recounts building an automated screenshot feature through iterative problem-solving with different AI models. Key lessons: refine the problem not just the prompt, anchor in a "Minimum Viable Truth," and match AI capabilities to task requirements.