baoyu-skills: Specialized Agent Tools for Content Generation

The baoyu-wechat-summary skill extracts highlights from WeChat group chats and converts them into structured, readable digests. It maintains long-term group history, user profiles, and group-level fact memory across multiple runs.

Key Features:

• Two Digest Modes: Generates a normal digest by default, with an optional roast (毒舌) version for a more critical/humorous tone.
• Memory Management: Tracks per-group history (history.json, history-digests.jsonl), per-user profiles, and group facts (memory.md).
• External Dependency: Relies on the wx-cli binary (the wx command). It does not bundle the script itself.

⚠️ Sandbox Warning: Because wx-cli accesses ~/.wx-cli/ and WeChat's data directory (e.g., ~/Library/Containers/com.tencent.xinWeChat/ on macOS), all wx commands must be executed with dangerouslyDisableSandbox: true in environments like Claude Code.

The baoyu-post-to-wechat skill allows you to publish content to a WeChat Official Account (微信公众号) using either the WeChat API or a Chrome browser (via CDP). It supports two main formats:

1. Articles (文章): Long-form content using HTML, Markdown, or plain text. Markdown input automatically converts external links into bottom citations for a cleaner WeChat look (unless --no-cite is used).
2. Image-Text Posts (图文): Short posts consisting of multiple images (up to 9) and text.

Publishing Methods

• api (Recommended): Fast; requires API credentials and your local IP to be in the WeChat allowlist.
• browser: Slower; uses Chrome with a logged-in session via CDP.
• remote-api: Fast; used when your local IP is not allowlisted but a remote server's IP is. It tunnels outbound HTTPS calls to WeChat through an SSH SOCKS5 dynamic port forward. No files or secrets are stored on the remote host.

When content exceeds the chunk threshold in Refined mode, the system uses a distributed approach between a Main Agent and multiple **Subagents":

1. Main Agent performs analysis on the entire document $\rightarrow$ 01-analysis.md.
2. Main Agent assembles the global translation prompt $\rightarrow$ 02-prompt.md.
3. Content is split into chunks in the chunks/ directory.
4. Subagents (one per chunk) are spawned in parallel. Each subagent reads 02-prompt.md to ensure context consistency and produces a chunk draft.
5. All chunk results are merged into 03-draft.md.
6. Main Agent performs the Critical Review $\rightarrow$ 04-critique.md.
7. Main Agent performs Revision $\rightarrow$ 05-revision.md.
8. Main Agent performs final Polish $\rightarrow$ translation.md.
9. Final Step: A cross-chunk consistency check is performed on terminology, narrative flow, and transitions at chunk boundaries.

The baoyu-wechat-summary skill produces two types of digests:

1. Normal version (Default): A sober, professional summary of group chat activity. It serves as the source of truth for profile updates and incremental mode.
2. Roast version (Opt-in via include_roast=true): A sarcastic, sharp, and provocative critique of the chat activity. It uses the same structure as the normal version but with a different tone, leaderboard annotations, and user portraits.

Key differences:

• Tone: Normal is professional; Roast is sharp and mocking (but not malicious).
• Title: Roast version includes a "毒舌版" suffix.
• Pain-points: The Pain-point section is omitted in the Roast version.
• @bot Q&A: In Roast mode, this becomes the "毒舌值班版" (Roast Duty version) with a mocking tone.
• Footer: Normal uses a standard footer; Roast uses a disclaimer about being an emotionless AI.
• User Portraits: Roast uses a "不留情面版" (Merciless version) that highlights flaws and contradictions.

The baoyu-article-illustrator skill analyzes articles to identify where visual aids are needed and generates consistent illustrations using a Type × Style × Palette approach.

Triggering the skill: You can invoke this skill by asking to:

• "illustrate article"
• "add images"
• "generate images for article"
• "为文章配图"

Core Concept: The Three Dimensions To ensure visual consistency, the skill uses three dimensions to define an image:

1. Type: The information structure (e.g., infographic, scene, flowchart, comparison, framework, timeline).
2. Style: The rendering approach (e.g., notion, warm, minimal, blueprint, watercolor, elegant).
3. Palette: The color scheme (e.g., macaron, warm, neon).

You can combine these manually or use a --preset flag (e.g., --preset edu-visual) to apply a pre-defined combination.

# Example manual combination
--type infographic --style vector-illustration --palette macaron

# Example using a preset
--preset edu-visual

The baoyu-electron-extract skill allows you to decompile or inspect the source code of an installed Electron application. It extracts the app.asar bundle and attempts to reconstruct the original source tree using .js.map files (if they contain sourcesContent). If source maps are unavailable, it falls back to extracting the raw files and formatting the minified JavaScript/CSS with Prettier.

Key Features

• Source-map Restoration: Rebuilds the original directory structure (e.g., src/main.ts) inside a restored/ directory.
• Prettier Formatting: Automatically formats minified code if source maps are missing.
• Platform Support: Works on macOS and Windows. It automatically searches standard installation directories (like /Applications or %LOCALAPPDATA%\Programs).
• Noise Reduction: Always skips node_modules and bundler runtime files to focus on the application's actual source code.

# Example: Extracting the 'Codex' app
${BUN_X} {baseDir}/scripts/main.ts Codex