neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
翻譯準備中
pstack,已本土化:Lauren Tan 的嚴謹工程工作流程技能(poteto 模式手冊、23 條原則、審查與調查技能),在 Claude Code、Cowork、Codex 與 ChatGPT 上原生運行。
AI 翻譯
pstack, nativized: Lauren Tan's rigorous engineering-workflow skills (poteto-mode playbooks, 23 principles, review and investigation skills) running natively on Claude Code, Cowork, Codex and ChatGPT.
開發工具TypeScript · JavaScriptGitHub @ahmedbenaw上線中營運者未確認
試用過了嗎? 點一下就好,不需要登入。
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權,自行管理介紹,無需管理員核准。
登入並產生驗證檔案,在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案,即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入,將前往登入頁面。登入後請回到此處,再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify,即可驗證所有權並自行管理介紹。
一分鐘完成驗證pstack,已本土化:Lauren Tan 的嚴謹工程工作流程技能(poteto 模式手冊、23 條原則、審查與調查技能),在 Claude Code、Cowork、Codex 與 ChatGPT 上原生運行。移植自 cursor/plugins。
AI 翻譯
pstack, nativized: Lauren Tan's rigorous engineering-workflow skills (poteto-mode playbooks, 23 principles, review and investigation skills) running natively on Claude Code, Cowork, Codex and ChatGPT. Ported from cursor/plugins.
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
This is a Claude-Code port of pstack, a plugin originally built for Cursor by Lauren Tan. It is MIT licensed (see LICENSE); the copyright and license are unchanged from upstream.
pstack is a set of rigorous engineering-workflow skills — poteto-mode's playbooks, the twenty-three principle skills, review and investigation skills (how, why, interrogate, architect, arena, swarm, ...), and the poteto-agent / Comment Sicko subagents. See docs/guide/README.md for the full getting-started guide (also ported from upstream) and the top-level skill/agent files for what each one does.檢視完整 README ↗This is a Claude-Code port of pstack, a plugin originally built for Cursor by Lauren Tan. It is MIT licensed (see LICENSE); the copyright and license are unchanged from upstream.
pstack is a set of rigorous engineering-workflow skills — poteto-mode's playbooks, the twenty-three principle skills, review and investigation skills (how, why, interrogate, architect, arena, swarm, ...), and the poteto-agent / Comment Sicko subagents. See docs/guide/README.md for the full getting-started guide (also ported from upstream) and the top-level skill/agent files for what each one does.
This port keeps the methodology, playbooks, and principles content verbatim — that's the valuable part. What changed is the plugin-runtime plumbing that's specific to Cursor:
.cursor-plugin/plugin.json → .claude-plugin/plugin.json, rewritten to Claude Code's flat plugin manifest shape (name, version, description, author, homepage).SKILL.md (skills, references, benny automations) kept only name and description, which is what Claude Code's skill loader reads. Cursor-runtime-only fields were dropped: disable-model-invocation, icon, color, paths. Two fields carried operational meaning worth keeping, so instead of silently dropping them they were folded into a short prose note in the skill body:
poteto-mode's mode: true (Cursor's "sticky mode") and its reminder field — noted in skills/poteto-mode/SKILL.md.typescript-best-practices's paths: ["**/*.ts", "**/*.tsx"] auto-attach glob — noted in skills/typescript-best-practices/SKILL.md.agents/poteto-agent.md and agents/comment-sicko.md keep name/description. poteto-agent's is_background: true has no Claude Code frontmatter equivalent (backgrounding is a dispatch-time choice, not a per-agent declaration), so it was dropped and replaced with a one-line note in the agent body instead.AskQuestion (Cursor's structured-question tool) was renamed to Claude Code's AskUserQuestion everywhere it's referenced in prose (poteto-mode, automate-me, setup-pstack skills, and the autonomous-run playbook).cursor-team-kit dependencies resolved. A few skills and playbooks called out to deslop, control-cli, and control-ui from Cursor's separate cursor-team-kit plugin. They now point at capabilities this repo actually has: control-cli/control-ui → create-verification-skill and maintain-verification-skill, which generate and maintain a project-local skill that drives the real app. deslop has no single equivalent — it strips slop from code, where this repo's unslop handles prose — so the outcome is written inline (narrating comments, unsupported guards, dead compatibility paths, unrelated edits), with principle-subtract-before-you-add as the rule and Claude Code's /simplify as a one-pass accelerator.automations/benny dormant bug-triage pack, docs/guide/**, assets/logo.png, all playbooks, references, and scripts — was copied over unchanged (only the same SKILL.md frontmatter cleanup applies to automations/benny/skills/**).Beyond the plumbing above, the ported content has been taken off Cursor-specific
tooling so an instruction names something the running host actually has:
gh is the single forge (Origin CLI and Graphite are gone), skill/plugin/transcript
paths resolve per host, the model config lives at ~/.pstack/models.md, /loop and
agent runtimes are described by capability rather than by vendor, and Bugbot triage
is generalized to automated review of any kind.
docs/nativization-map.md is the full audit: every
Cursor reference, its equivalent, and which of the four hosts it works on.
Two things genuinely do not port, and the map says so rather than pretending:
swarm, arena, architect, interrogate, both
autopilots — is degraded there. That is a property of the host.benny runner. Cursor hosts the automation that watches a repo and fires
on a webhook. benny's skills are host-neutral and run anywhere; its runner has
no equivalent, and on Claude Code the nearest thing is a scheduled task.hookify/ carries four guards for configuring and releasing pstack — two on the model config (~/.pstack/models.md and /setup-pstack's confirmation gate), two on the release path (a version-keyed cache that silently ignores content-only pushes, and status rows that record what was published rather than what is installed). Each encodes a failure that actually happened, named and dated in the rule itself.
They are templates. hookify globs .claude/hookify.*.local.md relative to the session's working directory only — there is no global rule location — so a rule fires only once copied into the project's own .claude/. hookify/README.md has the detail.
This fork tracked upstream's version exactly up to 0.15.2. Nativization makes the content deliberately diverge, so from 1.0.0 the fork keeps its own version line. Upstream releases are still tracked and merged; they just no longer set the number.
MIT, Copyright (c) 2026 Lauren Tan. See LICENSE.