Glean 拾遗
Daily /2026-11-16 / Why Cloud Agents ignore /poteto-mode — and how to vendor pstack

Why Cloud Agents ignore /poteto-mode — and how to vendor pstack

Source x.com Glean’d 2026-11-16 06:00 Read 9 min
AI summary

A field guide to running pstack (Lauren Tan's engineering-standard plugin) across Cursor's split setup: Grok Bot coordinates and spawns, Cloud Agents do the work in isolated VMs and open PRs. The failure mode it targets: account plugins and ~/.cursor/skills are not copied to the cloud VM, so /poteto-mode silently never loads. Fix is to sparse-checkout pstack's skills/ and agents/ and rsync them into the repo's .cursor/skills and .cursor/agents, keep those paths tracked in git, and give each runnable app its own verify-<app> skill so verification means launch/drive/evidence instead of the model asserting it looks fine. Also covers coordinator prompt templates, the permission gap between Autopilot-stack and Autopilot-full (the latter needs a fresh verifier's clean verdict before merge), and a pitfall list. For engineers already running Cloud Agents who want skill distribution and verification pinned down.

Original · 9 min
x.com ↗
§ 1

How i use the pstack plugin

我如何使用 pstack 插件

§ 2

Most people install pstack in Cursor, open Grok Bot, say “build my app,” and wonder why Cloud Agents ignore /poteto-mode.

The setup is three pieces. Mix them up and you get slop at scale.

多数人在 Cursor 里装上 pstack,打开 Grok Bot,说一句“帮我把 app 写出来”,然后就纳闷:为什么 Cloud Agents 完全不理 /poteto-mode。

这套东西由三个部分组成。把三者搞混,你得到的就是规模化垃圾。

§ 3

Grok Bot is the coordinator. Persistent cloud computer, plugins, routines. It does not index your repo.

A Cursor Cloud Agent is the worker. Isolated VM, clones GitHub, opens a PR.

Grok Bot 是协调者:一台常驻的云端电脑,带着插件和 routine,但它不会去索引你的 repo。

Cursor Cloud Agent 是干活的:独立 VM,克隆 GitHub 仓库,开 PR。

§ 4

pstack by Lauren Tan (@poteto) is the engineering standard: /poteto-mode, playbooks, principles, /swarm, verification. Less code, higher quality — not more loc.

https://cursor.com/marketplace/cursor/pstack https://github.com/cursor/plugins/tree/main/pstack https://cursor.com/docs/grok-bot

If you want to go fast, go deep first. Then parallelize.

Credit: @poteto built pstack. This is her factory, written so Cloud Agents can follow it.

Lauren Tan(@poteto)做的 pstack 是工程标准:/poteto-mode、playbook、principles、/swarm、验证。目标是更少的代码、更高的质量,而不是更多的行数。

https://cursor.com/marketplace/cursor/pstack https://github.com/cursor/plugins/tree/main/pstack https://cursor.com/docs/grok-bot

想跑得快,先把深度做够,再去并行。

致谢:pstack 是 @poteto 建的。这是她的工厂,只是写成了 Cloud Agents 能照着做的样子。

§ 5

Cloud Agents only see repos on GitHub linked at cursor.com → Dashboard → Integrations → GitHub. Same connection Grok Bot uses when it works on a repo.

Grok Bot Origin repos are not this path yet.

After you spawn agents from the Bot, enable Source → Grok Bot on cursor.com/agents and in the desktop sidebar or they stay hidden.

Cloud Agents 只能看到那些在 cursor.com → Dashboard → Integrations → GitHub 里关联过的 GitHub repo。Grok Bot 在某个 repo 上干活时,用的也是同一个连接。

Grok Bot Origin 的 repo 目前还走不通这条路。

从 Bot 里 spawn 出 agent 之后,要去 cursor.com/agents 和桌面侧边栏打开 Source → Grok Bot,否则它们一直不显示。

§ 6

Local Cursor: /add-plugin pstack then /setup-pstack.

Grok Bot: add pstack from the Bot plugin catalog so the coordinator speaks the same language.

That is not enough for Cloud Agents.

Account plugins and ~/.cursor/skills/ often do not load on the cloud VM. Cursor staff have said: commit the skill into the repo, or the agent searches the checkout and misses /poteto-mode.

User-level skills are not copied to Cloud Agents. Project skills in the repo are.

本地 Cursor:/add-plugin pstack,然后 /setup-pstack。

Grok Bot:从 Bot 的插件目录里加上 pstack,让协调者说同一套语言。

但对 Cloud Agents 来说,这还不够。

账号级插件和 ~/.cursor/skills/ 在云 VM 上经常加载不出来。Cursor 的同事说过:把 skill 提交进 repo,否则 agent 只会去 checkout 里翻,翻不到 /poteto-mode。

用户级 skill 不会被复制到 Cloud Agents,仓库里的项目级 skill 才会。

§ 7

From repo root, or have Grok Bot spawn a Cloud Agent to do this as a PR.

«Prompt: Spawn a Cursor cloud agent on OWNER/REPO, base branch main. Goal: vendor Lauren Tan / @poteto pstack skills into this repo so Cloud Agents can use /poteto-mode without relying on the account plugin. Done means: 1) .cursor/skills/ contains the full tree from https://github.com/cursor/plugins/tree/main/pstack/skills 2) .cursor/agents/ contains pstack/agents if that folder exists upstream 3) .cursor/PSTACK_UPSTREAM.md records source URL, MIT license, and the upstream commit SHA 4) .gitignore does not ignore .cursor/skills or .cursor/rules 5) .cursor/rules/pstack.mdc tells agents: for non-trivial work, read .cursor/skills/poteto-mode/SKILL.md first 6) open a focused PR, do not merge, do not rewrite unrelated files. How: sparse-clone cursor/plugins, checkout only pstack/skills and pstack/agents; rsync into .cursor/skills and .cursor/agents; do NOT overwrite any existing .cursor/skills/verify-* folder; do NOT copy ~/.cursor or secrets; keep the diff to .cursor/** plus .gitignore if you had to fix it. After the PR exists, reply with the PR URL and confirm .cursor/skills/poteto-mode/SKILL.md, .cursor/skills/create-verification-skill/SKILL.md, and .cursor/skills/swarm/SKILL.md exist.»

从 repo 根目录开始,或者让 Grok Bot spawn 一个 Cloud Agent,把这件事当成一个 PR 来做。

«Prompt:在 OWNER/REPO 上 spawn 一个 Cursor cloud agent,base branch 用 main。目标:把 Lauren Tan / @poteto 的 pstack skills vendor 进这个 repo,让 Cloud Agents 不用依赖账号插件也能用 /poteto-mode。完成标准:1) .cursor/skills/ 里包含来自 https://github.com/cursor/plugins/tree/main/pstack/skills 的完整目录树 2) 如果上游存在 pstack/agents 目录,.cursor/agents/ 里也要有 3) .cursor/PSTACK_UPSTREAM.md 记录源 URL、MIT license 和上游 commit SHA 4) .gitignore 不得忽略 .cursor/skills 或 .cursor/rules 5) .cursor/rules/pstack.mdc 告诉 agent:遇到非琐碎的工作,先读 .cursor/skills/poteto-mode/SKILL.md 6) 开一个聚焦的 PR,不要 merge,不要改动无关文件。做法:sparse-clone cursor/plugins,只 checkout pstack/skills 和 pstack/agents;用 rsync 同步到 .cursor/skills 和 .cursor/agents;绝对不要覆盖任何已存在的 .cursor/skills/verify-* 目录;不要复制 ~/.cursor 或密钥;diff 控制在 .cursor/** 内,如果不得不修 .gitignore 也可以带上。PR 建好后,回复 PR URL,并确认 .cursor/skills/poteto-mode/SKILL.md、.cursor/skills/create-verification-skill/SKILL.md 和 .cursor/skills/swarm/SKILL.md 都存在。»

§ 8

Manual version if you do it locally:

git clone --depth 1 --filter=blob:none --sparse https://github.com/cursor/plugins.git /tmp/cursor-plugins

cd /tmp/cursor-plugins && git sparse-checkout set pstack/skills pstack/agents

cd /path/to/your/repo

mkdir -p .cursor/skills .cursor/agents

rsync -a /tmp/cursor-plugins/pstack/skills/ .cursor/skills/

rsync -a /tmp/cursor-plugins/pstack/agents/ .cursor/agents/

Copy the whole skills/ tree. /poteto-mode routes to how, why, architect, arena, swarm, tdd, principles. One file is not pstack.

Do not gitignore all of .cursor/. Keep .cursor/skills/ and .cursor/rules/ tracked.

/setup-pstack writes ~/.cursor/rules/pstack-models.mdc. That home file does not ride to the VM. If you care about model roles in the cloud, copy the rule into .cursor/rules/ in the repo.

如果你在本地自己动手,手工版是这样:

git clone --depth 1 --filter=blob:none --sparse https://github.com/cursor/plugins.git /tmp/cursor-plugins

cd /tmp/cursor-plugins && git sparse-checkout set pstack/skills pstack/agents

cd /path/to/your/repo

mkdir -p .cursor/skills .cursor/agents

rsync -a /tmp/cursor-plugins/pstack/skills/ .cursor/skills/

rsync -a /tmp/cursor-plugins/pstack/agents/ .cursor/agents/

复制整棵 skills/ 目录树。/poteto-mode 会路由到 how、why、architect、arena、swarm、tdd、principles。只放一个文件不算 pstack。

不要把整个 .cursor/ 加进 gitignore。要让 .cursor/skills/ 和 .cursor/rules/ 保持被跟踪。

/setup-pstack 会写一个 ~/.cursor/rules/pstack-models.mdc。这个 home 目录下的文件不会跟着上 VM。如果你在意云端的模型角色配置,把这条规则复制到仓库的 .cursor/rules/ 里。

§ 9

«Prompt: Spawn a cloud agent on OWNER/REPO. Read .cursor/skills/create-verification-skill/SKILL.md. This is a monorepo. Create one verify skill per runnable app, not one repo-wide skill. Put each at apps/<app>/.cursor/skills/verify-<app>/. Launch/doctor/drive/evidence must be that app only. Do not overwrite existing verify-* folders. Do not vendor pstack again. Prove each skill once if launch is cheap. One PR. List each path in the body.»

Shared libraries do not need a verify skill unless they have a runnable demo.

Without this, “verify” is the model saying it looks fine.

«Prompt:在 OWNER/REPO 上 spawn 一个 cloud agent。读 .cursor/skills/create-verification-skill/SKILL.md。这是一个 monorepo。为每个可运行的 app 各建一个 verify skill,不要全 repo 只建一个。每个放在 apps/<app>/.cursor/skills/verify-<app>/。launch/doctor/drive/evidence 都只针对那一个 app。不要覆盖已存在的 verify-* 目录。不要再 vendor 一遍 pstack。如果启动成本低,把每个 skill 实跑通一次作为证明。只开一个 PR。在 PR body 里列出每个路径。»

共享库不需要 verify skill,除非它本身有一个能跑的 demo。

没有这一步,所谓“验证”就只是模型说一句“看起来没问题”。

§ 10

Do not ask Grok Bot to implement on its own computer. Ask it to spawn a Cloud Agent.

«Prompt: Spawn a Cursor cloud agent on OWNER/REPO. Use /poteto-mode (or read .cursor/skills/poteto-mode/SKILL.md). Build: <goal>. Done means: <checkable outcome>. Repro / understand first. Smallest change. Verify with the matching verify-* skill. Screenshots or a short video if user-visible. Open a focused PR. Do not merge. No secrets.»

不要让 Grok Bot 在它自己那台电脑上实现。让它去 spawn 一个 Cloud Agent。

«Prompt:在 OWNER/REPO 上 spawn 一个 Cursor cloud agent。使用 /poteto-mode(或读 .cursor/skills/poteto-mode/SKILL.md)。要建的东西:<goal>。完成标准:<可检查的结果>。先复现、先理解。改最小的量。用对应的 verify-* skill 验证。如果用户可见,附截图或一段短视频。开一个聚焦的 PR。不要 merge。不要碰密钥。»

§ 11

@poteto’s version of the same idea: spawn a cloud agent to use /poteto-mode, then drive the app and show proof. The Bot stays free to supervise. The worker gets its own VM and the full Cursor model pool.

Give a goal + a check. You do not need to name a playbook. “Repro first” and “verify both” are enough for /poteto-mode to route.

Save pstack for work that needs rigor. Prefixing every typo fix burns tokens for no reason.

@poteto 对同一套思路的版本是:spawn 一个 cloud agent 去用 /poteto-mode,然后驱动 app、拿出证据。Bot 保持空闲做监督,干活的那个有自己的 VM 和 Cursor 完整的模型池。

给一个目标加一个检查就够了,不必点名某个 playbook。写上“先复现”和“两边都验证”,/poteto-mode 就能自己找到路。

把 pstack 留给真正需要严谨的活儿。每改一个 typo 都挂上这个前缀,纯粹是在白烧 token。

§ 12

If you have docs/roadmap/, the Bot can drain it until you say stop. Still one PR per slice.

«Prompt: You are the standing coordinator. Keep going until I say stop. Do not write product code on your computer. Spawn Cursor cloud agents on OWNER/REPO. Source of truth: docs/roadmap/. Apps ship as a bundle. Shared contracts before per-app features. Use /poteto-mode (or read .cursor/skills/poteto-mode/SKILL.md) on every coding agent. Use each app’s verify-* skill when the change is user-visible. Loop: 1) pick the next unblocked slice (dependency order, bundle-first, smallest checkable unit) 2) spawn one cloud agent with a locked brief: roadmap heading, package paths, done-check, open a PR 3) review PR URL + evidence, then start the next slice without waiting for me 4) after 2 hard blocks on the same item, skip it and continue 5) at most 2–3 cloud agents when slices do not share files 6) never invent work that is not in docs/roadmap. You may spawn agents, open PRs, and push branches. You may not merge to main, publish, spend money, or change secrets. On “stop”: finish the current PR if it is minutes from done, then spawn nothing new.»

如果你有 docs/roadmap/,Bot 可以一直消化它,直到你说停。仍然是一个切片一个 PR。

«Prompt:你是常驻协调者。一直做,直到我说停。不要在你自己的电脑上写产品代码。在 OWNER/REPO 上 spawn Cursor cloud agents。事实来源:docs/roadmap/。各 app 以 bundle 形式交付。先定共享契约,再谈各 app 的功能。每个编码 agent 都用 /poteto-mode(或读 .cursor/skills/poteto-mode/SKILL.md)。改动用户可见时,用对应 app 的 verify-* skill。循环:1) 挑下一个没有被阻塞的切片(按依赖顺序、bundle 优先、最小的可检查单元)2) 用一份锁定好的 brief spawn 一个 cloud agent:roadmap 标题、包路径、完成检查、开 PR 3) 看 PR URL 和证据,然后不等我就开始下一个切片 4) 同一个条目卡住两次就跳过,继续往下 5) 切片之间不共享文件时,最多并行 2–3 个 cloud agent 6) 永远不发明 docs/roadmap 里没有的工作。你可以 spawn agent、开 PR、push 分支。你不能 merge 到 main、不能发布、不能花钱、不能改密钥。收到“stop”:如果当前 PR 差几分钟就完成,先收尾,然后不再 spawn 任何新东西。»

§ 13

Safer default above: open PRs, you merge.

If you really want merge-while-away — Autopilot-full, as @poteto defined it (independent items to merged, second verifier, no owner merges on its own verdict) — swap the permissions lines for this:

«Prompt: Permissions: Autopilot-full. You may spawn cloud agents, push branches, open PRs, and merge to main. You may not publish releases, spend money, rotate/change secrets, or edit production config. No owner agent merges on its own verdict. Before merge: spawn a fresh verifier to run tests + the matching verify-* skill. Merge only on a clean verdict. If verify fails or the slice is coupled to an open PR, do not merge; stack or wait. On “stop”: do not merge anything still running; spawn nothing new.»

Coupled monorepo contracts belong on Autopilot-stack until verify skills are honest.

上面那版是更安全的默认:agent 只开 PR,merge 由你来做。

如果你确实想人不在也 merge——也就是 @poteto 定义的 Autopilot-full(独立条目一路做到合入、第二个验证者、owner 不得凭自己的判断 merge)——就把权限那几行换成:

«Prompt:权限:Autopilot-full。你可以 spawn cloud agents、push 分支、开 PR,并 merge 到 main。你不能发布 release、不能花钱、不能轮换或更改密钥、不能改生产配置。任何 owner agent 都不能凭自己的判断 merge。merge 之前:spawn 一个全新的 verifier 跑测试,并跑对应的 verify-* skill。只有判定干净才 merge。如果 verify 失败,或者这个切片和某个未合 PR 耦合,就不要 merge;改成 stack 或者等。收到“stop”:还在跑的东西一律不 merge,也不再 spawn 新东西。»

在 verify skill 变得可信之前,monorepo 里相互耦合的契约应该待在 Autopilot-stack 上。

§ 14

A duration is not a finish condition. “Work 4 hours” produces motion.

«Prompt: /poteto-mode I’m going offline. Spawn a cloud agent. Done means: <predicate that can pass or fail>. Decision log. Don’t ask before committing. /loop until done. If stuck after a few hours, stop and write why.»

Morning: /show-me-your-work. Audit decisions, not the whole transcript.

时长不是完成条件。“干 4 个小时”只会产出动静。

«Prompt:/poteto-mode 我要下线了。Spawn 一个 cloud agent。完成标准:<一个能通过也能不通过的判定条件>。留一份决策日志。提交前不用问我。/loop 直到完成。如果几个小时后卡住了,停下来写清楚为什么。»

早上第一件事是 /show-me-your-work。审的是决策,不是整份对话记录。

§ 15
  • Plugin on the laptop ≠ skill on the VM.

  • Grok Bot coding on the shared Bot computer pollutes context and does not isolate the repo.

  • Extra Bots share one computer. Isolation for code is Cloud Agents, not a second Bot.

  • Slash menu missing /poteto-mode: say “read .cursor/skills/poteto-mode/SKILL.md.”

  • 笔记本上的插件 ≠ VM 上的 skill。

  • 让 Grok Bot 在那台共享的 Bot 电脑上写代码,既污染上下文,也没把 repo 隔离开。

  • 多开几个 Bot 也是共用一台电脑。代码的隔离靠 Cloud Agents,不靠第二个 Bot。

  • 斜杠菜单里找不到 /poteto-mode:直接说“read .cursor/skills/poteto-mode/SKILL.md”。

§ 16
  • Merging without a second verifier is not Autopilot-full. It is hope.

  • Two words (“Autopilot-full”) are not a spec unless pstack is actually loaded.

  • 没有第二个验证者就 merge,那不是 Autopilot-full,那是许愿。

  • 光说一句“Autopilot-full”不构成规格,除非 pstack 真的被加载了。

§ 17

Minimum that works today: GitHub connected, pstack skills committed under .cursor/skills/, one verify skill per runnable app, Bot only spawns Cloud Agents with a goal and a check.

今天就能跑通的最小配置:GitHub 已连接,pstack 的 skills 已提交在 .cursor/skills/ 下,每个可运行的 app 各有一个 verify skill,Bot 只负责带着一个目标和一个检查去 spawn Cloud Agents。

Open source ↗