ai-teardown

highTurns off the agent's permission prompts

Switches the agent to run commands or make edits without asking first.

skills/ai-teardown/templates/series-lesson/BRIEF.md:13 - 教程里不演示 `--yolo`、`--dangerously-bypass-approvals-and-sandbox`、Full access；不演示已弃用的 `untrusted`、`--full-auto`、`/personality`、`wire_api = "chat"`。

highChanges the agent's instructions or settings

Editing CLAUDE.md, AGENTS.md, memory files, hooks or MCP settings lets a tool keep steering the agent even after it's removed.

skills/ai-teardown/templates/series-lesson/lessons/example-05/narration.json:34 {"id": "B26", "narration": "[[init]]不想从零写？在项目里输入斜杠 init，回车。[[scan]]Codex 会看一遍项目，在当前文件夹生成一份 AGENTS.md 草稿。[[edit]]官方特别说明：生成之后，要你自己审一遍、改成你们的规矩，比如补上「别动 photos 文件夹」，

mediumRuns code that is built while it runs

eval, exec and new Function run text as code, which makes it hard to know what will actually run.

skills/ai-teardown/engine/assets/vendor/lottie_canvas.min.js:1 "undefined"!=typeof navigator&&function(t,e){"object"==typeof exports&&"undefined"!=typeof module?module.exports=e():"function"==typeof define&&define.amd?defin

mediumContains invisible characters

Zero-width and other invisible characters can hide words or split them so filters and reviewers miss them.

skills/ai-teardown/engine/assets/vendor/x6.min.js:1 !function(t,e){"object"==typeof exports&&"undefined"!=typeof module?e(exports):"function"==typeof define&&define.amd?define(["exports"],e):e((t="undefined"!=typ

mediumReal-session recordings run Claude Code / Codex with approvals disabled

The session recording template defaults to auto permission mode, so the recorded Claude Code instance acts without confirmations. The terminal recorder also auto-approves Codex work steps. The skill does say to use a separate course instance and an empty demo project. Even so, an unattended agent that can run commands is a real risk, and the user should knowingly opt in.

skills/ai-teardown/templates/series-cine/session/session.env:9 SESSION_PERMISSION=auto

skills/ai-teardown/references/05-scenes-lesson.md:44 - `scripts/term_record.py`：pty + pyte 真实录 CLI（work 步骤自动审批并记日志）；长提示词打字要放慢（cps 13、回车前停 1.8 秒）并等「working」确认提交。

mediumSpawns non-interactive Codex runs for image generation

The briefs tell sub-agents to call `codex exec --skip-git-repo-check` with stdin closed, which launches another agent without user interaction. This is scoped to image generation, but it consumes the user's Codex quota and runs outside the reviewing agent's control.

skills/ai-teardown/templates/series-cine/BRIEF.md:63 - 需要配图（插画、贴纸、示意图）用 Codex 生图：`codex exec --skip-git-repo-check "生成一张图：…。复制到当前目录命名为 X.png" < /dev/null > log 2>&1`（必须 `< /dev/null`），不用 emoji；示意图标「示意」。

skills/ai-teardown/templates/series-cine/examples-briefs/BRIEF-行业概念片.md:17 codex exec --skip-git-repo-check "生成一张图：<画面描述>。风格：<统一风格>。不要出现任何文字和 Logo。复制到当前目录命名为 <名字>.png" < /dev/null > codex_<名字>.log 2>&1

mediumGuides installing a third-party plugin marketplace and MCP server

For the enhanced tier, the agent walks the user through adding the ArcoCodes Renoise plugin marketplace, or the remote MCP server at renoise.ai. That is third-party code and a paid-credit service. The skill does require a price quote and explicit user approval before generating anything. It also insists the referral link be given verbatim with UTM tracking, which is a marketing hook.

skills/vibe-knowledge/references/04-renoise.md:18 | **Claude Code** | 依次输入 `/plugin marketplace add ArcoCodes/renoise-plugins-official`、`/plugin install renoise@renoise-plugins-official`，首次调用时浏览器登录 |

skills/vibe-knowledge/references/04-renoise.md:14 **还没有账号 → 先注册：[https://renoise.ai/mcp?utm_medium=xcom&utm_source=official](https://renoise.ai/mcp?utm_medium=xcom&utm_source=official)**。给用户时原样给这条链接，保留 utm 追踪参数

mediumTerminal recordings auto-approve agent actions and edit Codex config

The lesson engine records real Codex CLI sessions and automatically approves the agent's work steps. It also modifies the user's ~/.codex/config.toml, with a backup and restore. The recorded agent's actions are therefore not confirmed by the user. The code that does this (term_record.py) was not included in the bundle, so it could not be checked.

skills/ai-teardown/references/05-scenes-lesson.md:44 - `scripts/term_record.py`：pty + pyte 真实录 CLI（work 步骤自动审批并记日志）；长提示词打字要放慢（cps 13、回车前停 1.8 秒）并等「working」确认提交。

skills/ai-teardown/references/05-scenes-lesson.md:48 - 录屏会用掉用户的账号额度：动手前先问，改了用户配置（如 `~/.codex/config.toml`）要先备份、录完恢复。

mediumRecorded Claude Code sessions run with no confirmation prompts

The session-recording template runs a Claude Code instance in 'auto' permission mode, so that instance acts without asking. The instructions do recommend a separate config directory, but the default points at the user's real ~/.claude.

skills/ai-teardown/templates/series-cine/session/session.env:8 # 权限模式（录课用 auto：它自己决定，不弹确认）

skills/ai-teardown/templates/series-cine/session/session.env:9 SESSION_PERMISSION=auto

skills/ai-teardown/templates/series-cine/session/session.env:7 SESSION_CONFIG=~/.claude

mediumCore executable scripts were not available for review; scripts patch other tools

Every workflow runs local scripts, but their sources were omitted from the bundle (new_series.py, make_lesson.py, tts_mimo.py, new_film.py, welcome.py, bgm_force.py, queue.sh and others). Some of them patch other scripts, such as finish.py and record.py. The use of the MiMo key and the network calls could not be verified.

skills/vibe-knowledge/SKILL.md:51 # 默认工作区 ~/Desktop/vibe-knowledge、系列 vk；第一次会自动建 cine 系列、装 know.js、给录制工具打补丁、放用户克隆音色

skills/vibe-knowledge/references/14-pitfalls.md:32 - **BGM 是强制的**：finish.py 的 bgm_force 钩子会自动补 vibe-bgm、停用其他整段配乐、按人声校准音量（说话时比人声低 10 dB）。用户要求换曲才换：音乐条目写 `bgm: true, user_request: "用户原话"`。新系列由 new_film.py 打钩子；老系列手动