I make content, and I build the AI tools I make it with. Started with cameras, drones, and DaVinci Resolve; now I build tools for creators — typesetting, video pipelines, design systems an agent can pick up and use — and the agent infrastructure underneath them, which has to survive contact with real devices and real workflows: phones, watches, Windows. I care less about demos than whether the thing still works after deployment, user input, and edge cases.
芦苇 (Reed) — One Markdown draft becomes a WeChat article, Xiaohongshu cards and a long image, plus interactive WeChat stories. An AI can edit the open article through MCP without owning the document. 227 themes, 3 written here and 224 ported with author and license kept. Live.
Yoru Design Systems — Nine self-contained design systems, each one an installable Agent Skill: design guide, CSS tokens, components, and specimen cards for web, video, slides, and long-form content.
Blender Story Video — Agent Skill: one idea → 4K H.265 through Blender + Resolve, with delivery checks for frame count, encoding, and checksum.
Yoru Studio — A self-hosted execution studio for solo creators: projects, storyboards, field work, retrospectives, backups, and MCP-connected AI agents.
LyricLens — Turns lyrics into language-learning cards while you listen. One product, three hosts: a NetEase Cloud Music plugin, a standalone Windows app, and a phone review PWA, with end-to-end encrypted sync between the last two.
BGM Search Cheatsheet — Search phrases for free music libraries across 20 video scenarios (in Chinese). Library tags describe a feeling, not a use: background music finds filler, soft ambient underscore finds a usable bed.
Claude Desktop Kill-Switch — Caught Claude Desktop connecting directly 25 times in 12 seconds when the proxy dropped. Fail-closed firewall rules, plus the six "fixes" that looked right and weren't.
Akari Pulse — Reverse-engineered health data off a vivo WATCH GT. The official path won't export it, so: watch RPK → phone APK → own database → MCP.
Listening Bridge — Gives an AI fresh listening context from Android or Windows: active track, synced-lyrics context, and explicit playback controls over MCP.
chatgpt-mcp-connect — Verified ways to get custom MCP servers into ChatGPT, including direct servers and runtime/aggregation setups.
is-ai-down — "Is it just me?" One page. Status + reachability probes on Cloudflare. Live.
40 merged pull requests outside my own repositories, in 7 projects.
- Operit — Android AI agent · 21 merged PRs: model/provider integration, tool-call and history correctness, Compose/DSL state & timing, crash and UX fixes
- cli2api — multi-account API proxy for coding CLIs · 11 merged PRs: provider adapters (including an experimental Huawei Cloud CodeArts one), earliest-expiry-first account scheduling, model catalog persistence, tool-name preservation on streamed tool calls
- Smaller ones — gzh-design-skill (WeChat clipboard serialization), open-kimi-ppt-skill (browser launch and download races, #5 #6), awesome-ai-companion, H440-FPV frame



