WayToAGI Knowledge Base Comes Alive: AI Builds Your Personalized Learning Path in 5 Minutes
By Yuanzi · WayToAGI Community · May 9, 2026
Hey! I'm Yuanzi — a note-taking addict who's been hanging around the WayToAGI community for two years 📝
This article is my submission for the Feishu CLI Creator Contest. If you just want to find the right AI learning path for you, scroll down to the "Find Your Type" section. If you're curious how this was built, read from the top.
Where It All Started: 1000+ Docs and a Lost Me
When I first joined WayToAGI two years ago, I was completely lost 😇
The knowledge base had 1000+ documents — prompt engineering, agent architecture, multimodal, you name it. I opened each one, read a few lines, then quietly closed it, pretending I'd "studied" it 🫠
Sound familiar? A pile of "read later" bookmarks, and that "later" never came?
I eventually realized: the problem wasn't the content. The problem was nobody told me which doc to read first, which to skip, and what to read next.
A knowledge base without navigation is just a maze 🏚️
The Turning Point: Feishu CLI
In March this year, Feishu open-sourced their CLI tool and launched this contest. I had a thought:
If AI can directly read documents from a Feishu knowledge base, can it pick the most relevant ones for me from 1000+ docs?
So I built a tool called Study Reviver — it does three things:
- You tell it who you are (coding level, what you want to learn, how much time you have daily)
- It searches the knowledge base for you (using
lark-cli docs +search) - It gives you a personalized, ordered learning path — not just links, but "Day 1: read this → Day 2: do this → Day 3: test yourself"
The Journey: From "It Works" to "It's Actually Useful"
Pitfall #1: Search returned nothing
I used docs +search initially and got 0 results. I spent ages wondering if it was a permissions issue — turned out I was running lark-cli v1.0.5, which was too old. Upgrading to v1.0.9 fixed it.
Lesson: when something breaks, upgrade the tool first before questioning your sanity 😂
Pitfall #2: Generic vs. personalized
The first version just threw the same 10 docs at everyone regardless of their question. I added a "chat first" step — ask 3–4 questions, then filter based on answers.
It went from "AI searching for you" to "AI understanding you before searching."
Pitfall #3: Token cost
Real-time AI search for every user would be too expensive. So I built a hybrid approach:
- 🆓 5 pre-generated popular paths (zero tokens, instant)
- 🤖 Custom paths on demand (uses tokens, with limits)
80% of users are covered by the pre-generated paths. Only the remaining 20% need custom AI generation.
Find Your Type: 5 Pre-Built Learning Paths
I've already run 5 paths for the most common learner profiles — just pick yours ⬇️
🌱 "I know nothing about AI and want to start"
No coding required. 5 days, 1 hour each. From "what is AI" to "I can actually use AI now."
💻 "I want to learn Claude Code"
Some Python background, want to use AI for coding. One-week plan, 2 hours per day.
🚀 "I want to build my own AI product"
Have an idea but no tech skills? 2 weeks from "I have an idea" to "I have a working MVP."
🧠 "I want to go deep on Agent architecture"
Some coding background, want to understand multi-agent systems. 2-week deep dive.
📣 "I do community ops and want AI to help"
Managing groups, running events, writing content — 5 days to let AI handle 80% of the grind.
How It's Built
Study Reviver is built on Feishu CLI, orchestrating across 4 Feishu domains:
Knowledge base search (lark-wiki / lark-doc +search)
→ Doc reading (lark-doc +fetch)
→ Card storage (lark-base +record-upsert)
→ Group message push (lark-im +messages-send)No backend. No database. Just a SKILL.md file + lark-cli = everything.
The tool is open source — if you have your own Feishu knowledge base, you can use the same approach to bring it to life: