產品經理，用 AI agent 自建工作流。
I'm a product manager who uses Claude Code and other AI agents every day. I treat them as a system to design, measure and improve, and I like turning repeated work into small, checkable pipelines.
- A personal AI agent operating system: a task inbox that routes work automatically, background sub-agents that execute it, and independent reviewer agents that verify the results.
- A personal knowledge base with a fact-checking loop: facts are extracted with sources, then checked against the originals before they are written down.
- Multi-agent project management: PM, PO and executor layers, with quota control so long-running work stays within budget.
- Token ROI and model routing: estimating whether a task is worth the spend, then assigning the right model to it.
- Cross-CLI handoff: moving work between different AI command-line tools without losing context.
- Privacy-first design: personal-data gates,
private/separation, local-first storage. - A LINE chatbot powered by Gemini, and a hardware prototype, the Smart Speaking Glove.
- pm-init: a Claude Code skill that sets up local project tracking with an editable HTML dashboard and a Markdown mirror.
- Smart-Speaking-Glove: an ESP32 glove that turns finger presses into speech.
- Privacy by default. Personal data stays local and separated from code; risky actions are off until confirmed.
- Verify before trust. AI output gets an independent check against the source before it becomes a fact.
- Small, reversible steps. Dry runs first, snapshots before edits, nothing destructive without a review.
- Measure ROI. Track what each automation costs and what it returns, then adjust.