[Feature + PoC] xhs-note-to-skill: Convert Xiaohongshu (小红书) Service Walkthroughs into Self-Evolving Agent Skills

[melisaliao502-debug]

Hi! 👋

I'm from the Xiaohongshu (小红书 / REDnote) team. I've been thinking a lot about how the massive amount of genuinely useful, experience-driven content on Xiaohongshu could be turned into something agents can actually use — and when I saw OpenSpace, it really clicked for me.

Xiaohongshu users don't write documentation. They write "I just did this, here's exactly how, and here's where I almost messed up." Tax filing, visa applications, insurance claims, government portal navigation — millions of step-by-step walkthroughs with screenshots, error workarounds, and timing tips. This is real procedural knowledge, battle-tested by real people.

I've been experimenting with converting these notes into structured skills, and I'd love to share what I've learned.

---

What I'm sharing: xhs-note-to-skill

A skill that converts Xiaohongshu how-to notes about online services into OpenSpace SKILL.md format. The pipeline:

1. Extract: Parse note structure → steps, prerequisites, error cases, timing constraints
2. Structure: Map to SKILL.md with decision trees, error handling tables, and breakpoint markers
3. Enrich: Tag [FRAGILE] elements for OpenSpace evolution triggers

Included example

A real tax refund walkthrough (~15K saves) → complete china-individual-tax-refund skill with:

• 6 execution steps mapping to actual 个人所得税APP UI
• 5 error-handling entries (server crashes, wrong bank card type, missing income data...)
• Decision points, timing windows, and version history

Why online services specifically?

Xiaohongshu has guides for everything (travel, cooking, beauty...), but online service walkthroughs are uniquely valuable for agents because:

• Agent-executable: Steps map to real UI interactions (navigate URL → fill form → click → verify state)
• High-stakes failures: Getting tax filing wrong has real consequences → rich error handling → great fuel for OpenSpace's FIX trigger
• Seasonal refresh: Tax apps update annually, visa portals change quarterly → natural evolution cycles
• Directly comparable to GDPVal: The benchmark tests tax prep and compliance — this bridges the gap to Chinese-language equivalents

Evolution fit

Trigger	How XHS Notes Feed It
FIX	Portal UI update → skill breaks → new "2026新版" notes appear on XHS → auto-patch
DERIVED	tax-refund-employee → derive tax-refund-freelancer from related notes
CAPTURED	Agent improvises around unlisted error → workaround captured as new rule

Scale potential

Domain	XHS Coverage	Example
Tax Filing (报税/退税)	Very high	china-individual-tax-refund
Social Security (社保)	High	social-security-transfer-city
Housing Fund (公积金)	High	housing-fund-withdrawal-rental
Visa Applications (签证)	Very high	japan-tourist-visa-apply
Insurance Claims (保险理赔)	Medium-high	car-insurance-claim-online
Government Portals (政务服务)	High	business-license-register

---

I've put together a PR (#67) with the full SKILL.md as a starting point. Would love to explore more interesting directions together — there's so much untapped experiential knowledge on the platform that could make agents genuinely smarter. 🚀

Happy to iterate on format, scope, or approach based on your feedback!

[melisaliao502-debug]

One more thought — if there's interest, we'd also love to explore building evals together.

The GDPVal benchmark is great for English professional tasks, but there's no equivalent for Chinese online service workflows (tax filing, social security, government portals, etc.). We're thinking about whether we could collaborate on a Chinese-language eval set — potentially with internal resources from our side to help build and validate it.

This could be valuable for the whole community: a real-world benchmark for Chinese service automation that tests both skill quality and evolution effectiveness.

Happy to discuss scope and format if this sounds interesting! 🤝