作者做了一个叫 salt.md 的工具,把每个任务做成一个页面,智能体读取页面上的规则、完成工作并在页面上留下笔记,这样从 Claude 切换到 Codex 时,Codex 能知道 Claude 之前做了什么。作者主要用 Claude Code 开发,有时用 Codex,并称最大的经验是让智能体写下为什么这么做,而不只是做了什么。项目已开源在 github.com/saltmd/salt.md。
In February I hooked up 11 MCP Servers to my coding agent. Just the tool list in an empty session ate up 34,000 tokens, with Datadog alone contributing over two hundred tools. The agent got slower and kept picking the wrong tools.
Why it matters: I'm writing this up because 11 MCP Servers of my own blew up my context, and it showed me that tools and knowledge belong in different containers.
文章从认知科学角度分析 Vibe Coding 的兴起:自然语言提示词消除了编程语法门槛,让任何人都能把脑中的想法直接变成可运行的应用。但作者指出,用户脑中的心智模型越模糊,对技术需求的描述就越不精确,也就越依赖 AI 厂商补全细节,这正是厂商获利的关键。作者认为“先澄清需求”(Clarify First)才是 AI 工程的未来。
The author open-sourced claudemanager, a local daemon that Claude Code points to via ANTHROPIC_BASE_URL. It only changes the request's Authorization header to route sessions to the Max account with the most remaining capacity in its 5-hour, weekly, and per-model windows, switching at custom thresholds before those windows fill up.
Why it matters: The author also open-sourced a local proxy that automatically distributes Claude Code traffic across multiple Max accounts based on remaining quota, and logs requests along the way.
作者发布 windvane,一个 MIT 许可、无依赖的 Claude Code 插件,用 Python 引擎在会话中自动看护上下文:recorder 根据任务列表、编辑、提交和上一条回复起草检查点,模型一次调用即可接受或改一个字段;插件监控上下文填充,在检查点落盘后于下一个回合边界压缩,并自行发一条提示让模型从检查点继续。
The author has open-sourced the Project Athena v9.9.9 kernel (MIT License), a local-first memory and governance framework built to solve the problems of Claude Code needing repeated corrections, sub-agents overstepping their bounds and modifying files, and losing state when it hits the quota limit mid-task.
Why it matters: The author validated a CLAUDE.md slimming and local-memory approach across 1900 sessions, and provides a directory structure and verification rules you can reuse directly.
The author points out that CLAUDE.md and built-in memory in tools like Cursor and Claude Code only cover a single machine or a single repository, so they can’t share team decisions across tools and members. As a result, the same architectural trade-offs get explained over and over again.
The author built the MCP server pulpie-mcp so AI coding agents can scrape web pages and docs into complete Markdown themselves, keeping tables, code blocks, links, and images intact.
The author released the cache-warmer plugin to fix a problem: after you step away from Claude Code for more than 5 minutes, the next prompt has to rewrite the prompt cache at full price, so costs pile up over long sessions and the first reply gets slower.
A developer built a small Mac app for personal use with Claude, relying on readme.md and session_state.md to track project state. After each code change, they had Claude run git commit push main to push to GitHub, and frequently used /clear.