The archive · Developer & Business Tools · Product decision · 2026
Cangjie Skill: distilling books into Agent Skills hit 9.5k GitHub stars
kangarooking's MIT pipeline distills books, videos and podcasts into callable Agent Skills; under five months after launch the repo passed 9.5k stars.
Cangjie Skill
What the business is
Cangjie Skill is a free MIT-licensed open-source system that distills high-value content — books, long-video subtitles, podcast transcripts and courses — through a seven-stage RIA-TV++ pipeline into installable, callable Agent Skill packs, with an official registry website, trilingual READMEs and companion repos in the nuwa–cangjie–darwin skill ecosystem.
How it started
kangarooking, an indie developer and AI blogger (WeChat official account 'Kangarooking AI Inn (袋鼠帝 AI 客栈)', Xiaohongshu, Douyin and X audiences), saw nuwa-skill make 'distill a person into an AI skill' — Elon Musk skills, Warren Buffett skills — a viral idea, with darwin-skill evolving skills automatically. He built the complementary move: distill the methodologies people express systematically in books, interviews, podcasts and long videos, because summaries and notes are compression rather than structured reuse. The GitHub repo was created on 2026-04-16.
What happened
The project shipped a methodical stack fast: v2.5.0 introduced a Capability Bundle as the single source of truth (stable capability cards before any installable output), two deterministic delivery modes, a unified scripts/cangjie.py toolchain with diagnostics, rollback and benchmarking, plus Registry v2 and website support; a DeepSeek Harness plugin and simplified Chinese, English and Japanese READMEs widened distribution. Published packs included Buffett shareholder letters (20 skills), Poor Charlie's Almanack (12), Mao's Selected Works (25), Huangdi Neijing (22), Andrew Ng's AI for Everyone video course (25) and X-growth resources (15), while external contributors added book2startup, book2skill and six qbdx-hub example repos. By the September 2026 crawl the repo had passed 9.5k stars and 1.1k forks; no company, funding or paid tier is visible in the material.
No ending yet — it is still running.
Background
Cangjie Skill is a free MIT-licensed open-source system by kangarooking (袋鼠帝), a Chinese AI blogger and indie developer, that distills methodologies from books, long-form videos, podcasts and courses into executable, installable Agent Skills. Its RIA-TV++ pipeline runs seven stages — whole-content comprehension on Mortimer Adler's method, five parallel extractors, triple verification with a promotion gate, RIA++ capability cards, Zettelkasten linking, pressure testing and deterministic compilation into a single router Skill or a compact pack.
The bet extends a viral category: nuwa-skill had proven people could be distilled into AI skills, so kangarooking gambled that systematically expressed content deserved the same treatment and that one rigorous, spec-first implementation could own the book-to-skill format. The repo ships a DeepSeek Harness plugin, simplified Chinese, English and Japanese READMEs, a registry website (cangjie-skill.com), and an ecosystem position between nuwa-skill (people), darwin-skill (evolution) and third-party contributors who added example repositories.
The distribution bet worked quickly: the repo was created on 2026-04-16 and passed 9.5k stars and 1.1k forks by the 2026-09-04 crawl, with generated packs spanning Buffett's letters, Poor Charlie's Almanack, Mao's Selected Works, Huangdi Neijing and Andrew Ng's AI for Everyone course. As of September 2026 it remains a free open-source project with no company or monetization visible in the material — a bet on ecosystem position rather than on revenue.
What has to be true
- The idea rode a proven wave: nuwa-skill had already made 'distill people into skills' viral, so content distillation entered with a familiar mental model and an existing audience.
- Rigor became the differentiator: triple verification, promotion gates and bait-question pressure tests made packs feel engineered, not scraped from a book.
- Deterministic packaging plus a registry gave users a repeatable install path, which is what a reference implementation needs to become the default.
- The creator's Chinese developer audience across WeChat, Xiaohongshu, Douyin and X gave the launch an initial push that a nameless repo would not have had.
- The open question is monetization: with 9.5k stars but no company, funding or paid tier visible, the bet that ecosystem position later converts remains unproven.
What can be applied
Compete on output quality, not category novelty: a verified, pressure-tested pipeline turned a cloneable idea into the reference implementation the ecosystem contributed to.
Aftermath
As of 2026-09-05 Cangjie Skill is live and actively maintained: v2.5.0 with its Capability Bundle architecture, a live registry at cangjie-skill.com, and 9.5k stars, 1.1k forks and 58 commits on GitHub. The material shows no company, funding or paid product behind the project — it is a creator-led, open-source bet on the Agent Skills ecosystem, with contributors from the broader Chinese AI community expanding its example repositories.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card