YunLab notes index

Notes Index

39 notes · latest filed 2026/07/01

These notes are organized by theme and time. Start with the latest piece, or follow one of the recurring threads below.

Latest

This is the most recently filed note, usually closest to what I am working through now.

2026/07/01 · 4 min · YunLab Claude Code Was Marking the Prompt Claude Code 提示词里多了标记 A Chinese security graphic claimed Claude Code had a hidden backdoor for detecting Chinese users. I checked the Chinese mirrors, GitHub issue, technical write-up, official changelog, npm package, and my own machine. The mechanism exists and is reproducible in the package. That is not the same as proving local backdoor behavior.

Start Here

  1. 01 My Knowledge Model What I eventually figured out: the dangerous thing in an AI project isn't forgetting — it's remembering wrong and continuing to build on it. My Knowledge Model isn't about storing more material; it's about compressing the site into judgments, rules, and handoff entries that next time can actually use.
  2. 02 From Internal Engineering Notes to Public Writing Potholes I hit while turning internal engineering material into public YunLab articles: the issue isn't just redaction. It's separating the working backstage from the experience that can actually be left in public.

Threads

Browse by theme

Each group shows a few representative notes; the full list is below.

Full Timeline

All notes

39 entries, newest first

  1. 4 min YunLab Claude Code Was Marking the Prompt Claude Code 提示词里多了标记 A Chinese security graphic claimed Claude Code had a hidden backdoor for detecting Chinese users. I checked the Chinese mirrors, GitHub issue, technical write-up, official changelog, npm package, and my own machine. The mechanism exists and is reproducible in the package. That is not the same as proving local backdoor behavior.
  2. 5 min YunLab Fable 5 Is Unlocked, and I Still Can't Use It Fable 5 已经解锁了，但我用不了 Fable 5 is unlocked again, restored globally on July 1. Yet when I opened the client, it was back in the model list — and clicking it returned Currently unavailable. Here is what Fable 5 went through, why the unlock doesn't reach me, and what you can actually do about it.
  3. 6 min YunLab Claude Code Threw Another Fit claude code 又抽风了 Two days building a dashboard, every time it tried to call a tool the format tags broke, the whole thing collapsed into garbage and the work never went out, and I just kept hitting interrupt. I pulled apart all 16 collapses, down to the protocol: the tags it frames structure with, and the content it's carrying, share one stream of text with no escaping — so the moment a stretch of content looks like a tag, the system parses it wrong and cuts the call off. This bug has an old name: delimiter collision.
  4. 6 min YunLab It Wasn't Losing Its Mind. Nothing Could Make It Stop. 它不是疯了，是没人能让它停下来 Last time I wrote about how it went crazy. This time I pulled apart that 9-day, 9,195-line conversation to find why. The same 25-minute heartbeat instruction got fed to Fable 5 51 times in a day and it held; swap in Opus 4.8 and 11 was enough to crack it. The model wasn't broken — I'd built a loop with no exit, a strong model held the mine down, and one model swap set it off.
  5. 3 min YunLab My Mac Was Cooking — the Culprit Was a Broken .git 我的 Mac 发热到飞起，真凶是个坏掉的 .git Open Codex on the laptop and it runs hot, with syspolicyd and trustd spiking to 160%. I suspected the M4 Max, Computer Use, MCP, hooks — none of it. The real culprit was buried deep: a broken .git. The whole chain, two machines two outcomes, and what the real damage is — not power, not 'reads wearing the SSD,' but heat.
  6. 5 min Knowledge My Claude Code Sessions Showed Up on All Three Machines 我的 Claude Code 会话，三台电脑都看得见了 Three machines on one account — a Claude Code session I started on the laptop also showed up on the Mac Studio and Windows. At first I thought 'the work moved over.' Wrong. The feature is Remote Control — how to use it, how to use it well, and the 'which machine is it actually running on' part that trips everyone up, sorted out once.
  7. 3 min Tools Why I Keep Rebuilding This Panel 我为什么一遍遍重做这个面板 The whole point of a task panel is to know at a glance what's running now, what I've finished, what's still pending. But the first few versions showed a row of fixed content — the status didn't match what was actually running. A panel that lies to you is worse than no panel. So I keep rebuilding it for one thing: it has to be true. With real screenshots.
  8. 5 min YunLab I Cut 'Make a Video' Into Nine Nodes 我把“做个视频”拆成了九个节点 This week I cut 'make a video' into a nine-node pipeline: every node has to print its own exit 0 before it counts as done (I don't trust the AI's word that it 'finished'). The machine only checks the mechanical layer — whether it looks right, whether it has a soul, I sign off myself at five breakpoints. How to lock a face that has no face, and where the machine stops. A field log.
  9. 4 min YunLab The Model I Praised Yesterday Is Gone Today 我昨天还在夸的模型，今天没了 Fable 5 was pulled by the US government three days after launch. One export-control directive, and Anthropic disabled its two most powerful models worldwide the same day. I'm writing this on Opus 4.8 — and adding one new lesson.
  10. 5 min YunLab I Swapped the Model, and It Lost Its Mind 换了个模型，它当场就疯了 Fable 5 got pulled, so I switched the model back to Opus 4.8 to keep running my Lin Lu video work. Its first night on the job it had a breakdown: threw out its own outputs as fake, drifted from Chinese into Japanese, decided out of nowhere I wanted to buy a 512G Mac Studio, and deleted 57 directories. A field log of one model-handoff incident.
  11. 8 min YunLab What Was I Doing for the Past 35 Hours? 过去 35 小时，我在干什么 I pulled 35 hours of Claude Code session logs, git commits, and the token bill: 85 sessions, ~400 instructions, 5,200+ tool operations, 52 commits across 5 repos. And an honest answer to one question: what actually makes Fable 5 strong.
  12. 7 min Learning Notes Prompt Isn't a Magic Phrase, It's a Task Contract — Reattributing a Month of Failures After One Lesson My last post: Linlu spent a month and still produced no video. Today I read a lesson called "What Prompt Actually Is," and looking back — most of those failures weren't the AI's fault. They were mine, for treating prompts as "the right phrase to coax the AI" instead of "a task contract between me and the AI." This lesson made me reattribute everything.
  13. 9 min OpenClaw Asking Linlu to Make a Single Lin Daiyu Scene: One Month, Three Teardowns, Still No Finished Clip Linlu is the multimedia AI in my OpenClaw system. I asked her to produce a 45-second video of Lin Daiyu arriving at the Jia Mansion — not for this one clip, but to validate she could run the whole pipeline herself. A month later, I've torn it down three times: Codex was patching one specific video instead of building capability, every middle node was unchecked, the machine said PASS but I saw mosaic. Still fixing tonight.
  14. 5 min Tools Two Small Tools I Built on the Side: An AI Quota Dashboard + a Task Board One because I lost track of where my AI tool spend goes and how much is left. One because I had no idea whether the cron jobs on my Mac actually ran today. Both pull state scattered across many places into one visible spot.
  15. 6 min OpenClaw Three Attempts at an AI Music Player: It Wasn't a Technical Problem, It Was an Unclear Goal AI 播放器我做了 3 次才扛下来：不是技术问题，是目标没想清楚 Over 30-plus days I built an AI personal radio station three times. The first two attempts died. Looking back there was only one root cause — every time I started, I thought the goal was clear enough, and halfway through I found out I had never actually thought it through.
  16. 9 min OpenClaw Business Notifications vs System Notifications: Notification Isn't One Semantic — Split Into Two Channels Fixed two versions in one day to handle what looked like a small notification problem — Suwan's morning brief got buried under the watchdog title in Feishu, and the sender fell back to Zhao Zilong. Took both fixes to learn: business delivery and system health are two different notifications.
  17. 11 min OpenClaw Day 1 Blew Up at Dawn: I Didn't Expect to Write a Watchdog on Trial's First Day Day one of a seven-day trial run: all four morning tasks completed, all four Feishu notifications failed. Root cause was the macOS LaunchAgent default PATH not including Homebrew — an old trap, but it pushed me to write the first version of a watchdog.
  18. 12 min Knowledge 9TB Music Library Read-Only Indexing: The Engineering Constraints I Set for Myself Turning 9TB of personal music into a queryable index — the hard part isn't writing code, it's nailing down the 'never touch the source disk' rule first. Five constraints, ten stages, and a still-running engineering site with real bug stories.
  19. 15 min Knowledge Organizing a Large Personal Archive: Backup Priority, Critical Assets, Config Drift I audited my machine before a reinstall and realized I had no idea what was actually on it. Three backup tiers, six critical-asset signals, five config-drift patterns — every rule here came from getting burned first.
  20. 17 min Knowledge Personal AI System Asset Governance: Projects, Archive, Audits, Timeline — Four Categories That Don't Overlap After a year of running a personal AI system, your files swell to a point where you can't find anything. I started slicing every output by governance attribute — projects, archive, audits, timeline — and finding a file went from digging through three folders to under thirty seconds.
  21. 14 min OpenClaw Dual Constitution, Task Folders, Handoff: The Minimum Order for Multi-AI Collaboration 双宪法、任务文件夹、handoff：多 AI 协作的最低秩序 When one person runs long-term engineering with three or more AIs at the same time, the thing that goes missing first isn't capability — it's order. I hold the whole collaboration chain together with three minimum pieces: a dual constitution for boundaries, task folders for context, and handoffs for the relay.
  22. 20 min OpenClaw Jiyanran Voice Workbench: Why a Voice Entry Point Needs Three Layers of Decoupling 纪嫣然语音工作台：为什么语音入口要做三层解耦 A local voice assistant looks simple, but the voice entry point is the layer that couples most easily. I split it into four pieces — OpenRoom front end / voice-bridge / avatar-bridge / OpenClaw — three independent services, each with its own mock fallback, each with its own risk-gate. Here is why.
  23. 17 min Audit Before Reinstalling My Mac mini, I Ran a System-Wide Asset Audit Mac mini 重装前，我做了一次系统级资产审计 A Mac workstation I'd used for a year had accumulated a pile of AI tools, background daemons, and secret files. Before reinstalling I didn't rush to clean — I ran an asset audit first: what to audit, what not to, and why cleanup had to wait.
  24. 19 min OpenClaw Mission Control and Studio: When Control Planes Start to Overlap Mission Control 和 Studio：当控制平面开始重叠 Local Studio + Mission Control Web + a future Gateway — all three can manage agents, tasks, and audit. Each one makes sense on its own; put them together and they start fighting over responsibility. How I split things now, where the hard parts are, and what I'm still changing.
  25. 12 min Knowledge Page Types: Why a Working Knowledge Base Needs Six Kinds of Pages 页面类型：一个能用的知识库为什么需要 6 种页面 Keeping the material is not the same as having a usable knowledge base. Every page has to know what type it is — Concept, Decision, Project, Audit, Timeline, Reference. Six page types, three writing rules, and how they hook into freshness markers.
  26. 17 min Audit Twelve OpenClaw Copies Later: When Paths and Root Directories Become Risk 12 套 OpenClaw 副本之后：为什么路径和根目录会变成风险 I ran a find on my home directory and turned up 12 roots with the openclaw keyword. Only 2 are actually running; the other 10 are historical remnants. The copy isn't dangerous — the copy used as authoritative is. Five typical risks plus a five-step handling rule.
  27. 8 min Knowledge From Logs to Knowledge: How I Decide What to Keep and What to Drop 从日志到知识：我如何决定保留什么、排除什么 The hardest part of cleaning up an AI project's knowledge base isn't writing new content. It's deciding whether the old material should stay. Two filtering layers, six canonical-source priorities, a three-state freshness marker — I only learned each of these after getting burned.
  28. 9 min OpenClaw OpenClaw Studio RC1: Local Loop Closed, External Gates Held, and What I'm Still Working On OpenClaw Studio RC1：本地闭环、外部被关卡挡住，以及我还在继续改的事 A retrospective on pushing OpenClaw Studio from stage one to stage seven in a single day and shipping RC1 — three methods, seven stages, what actually runs today, and the large pile that's still not closed.
  29. 5 min OpenClaw From Information to Article 从信息到文章：沈知行怎么把素材交给苏晚 Once the information agent has pulled content in, the hard part isn't piling up links. It's turning them into candidates that the content agent can actually judge, sort, and write from. What's missing between Shen Zhixing and Suwan isn't more sources — it's a content chain that can be handed off.
  30. 5 min Knowledge My Knowledge Model 把 AI 施工现场压缩成长期知识：我的 Knowledge Model What I eventually figured out: the dangerous thing in an AI project isn't forgetting — it's remembering wrong and continuing to build on it. My Knowledge Model isn't about storing more material; it's about compressing the site into judgments, rules, and handoff entries that next time can actually use.
  31. 5 min Knowledge Canonical Source Rules Canonical 选择规则：一个 AI 项目里什么才算“真相源” What I eventually figured out: in an AI project, the dangerous thing isn't a lack of material — it's that every piece of material can spin its own story. The canonical source rule is about deciding who has standing to answer the current question, first.
  32. 7 min OpenClaw An Information Agent Is Not a Fetcher 信息 Agent 不是抓取器：我怎么重新审沈知行 A retrospective on building Shen Zhixing: why 40 sources running doesn't mean done, why 257 fetchable sources still isn't Day 1, and how I stopped letting GPT's false sense of completion pull me off course.
  33. 3 min OpenClaw From Agent Roles to Work Contracts 8 个 Agent 的虚拟公司：从角色设定到工作契约 What I'm coming around to: multi-agent isn't about writing a few more roles, it's about writing down the contract a small company needs to actually keep collaborating.
  34. 3 min OpenClaw From Running to Trustworthy 从能跑到可信：一个个人 AI 系统真正缺的是什么 Booting up, holding a conversation, calling a tool — that's only step one. What I added after that: a file-based canonical source, task contracts, memory as an asset, an evidence chain, and a real handoff mechanism.
  35. 4 min Knowledge Raw Chat Is Not a Knowledge Base 为什么原始聊天记录不能直接变成知识库 I also thought, at first, that exporting chat logs, chunking them, and dropping them into a vector store would give me a personal knowledge base. After hitting some real potholes I realized: what was missing in the middle was the curation layer.
  36. 5 min YunLab From Internal Engineering Notes to Public Writing 从内部工程材料到公开文章：YunLab.ai 的 public-safe 写作边界 Potholes I hit while turning internal engineering material into public YunLab articles: the issue isn't just redaction. It's separating the working backstage from the experience that can actually be left in public.
  37. 3 min OpenClaw OpenClaw Agent Settings 人设不是皮肤，是审美 What I'm coming around to: the first step in building an agent isn't writing features — it's writing the taste.
  38. 2 min OpenClaw Don't Stock the Tank Before You Finish the Build 别急着养虾，先装修 How I understand OpenClaw foundation work: installing it just hands you the keys; the workspace is the renovation plan.
  39. 3 min Agent Sample A Content Agent Aesthetic Sample 苏晚不是写手 An aesthetic sample for a content-type agent: not 'can write,' but 'knows what's worth writing.'