skillens 0.2.0

AI-powered CLI tool that evaluates whether a learning resource is worth your time

🔍 SkiLens

Is this worth learning? Ask before you invest 40 hours.

One command. Instant verdict: LEARN · SKIP · CONSIDER ALTERNATIVES.

pip install skillens
skillens "https://www.coursera.org/learn/machine-learning"

English · 中文

---

😤 The problem

You find a course. It looks promising. 40 hours later:

• 🕰️ Content is 3 years outdated, pre-transformer era
• 📉 The framework it teaches is getting automated away
• 📚 The same material exists in a 2-hour YouTube video
• 🎯 It doesn't match where your career is actually heading

"The hardest part of learning isn't learning — it's deciding what to learn."

SkiLens answers "should I learn this?" before you commit. Paste a URL. Get a verdict in 3 seconds. Optionally find better alternatives.

---

⚡ 30-second demo

$ skillens "https://www.coursera.org/learn/machine-learning"

╭──────────────────────── 🔍 SkiLens ─────────────────────────╮
│                                                              │
│   Machine Learning — Stanford (coursera)                     │
│   by Andrew Ng · Updated: 2024-03 · 60h                      │
│                                                              │
│   Market Demand    ████████░░  78    ↗ growing               │
│   Skill Half-life  ██████████  ~10y  ✦ evergreen             │
│   Info Density     ██████░░░░  62    → moderate              │
│   Freshness        █████░░░░░  48    ↘ aging                 │
│   Effort vs Return ███████░░░  71    → solid                 │
│                                                              │
│   Overall          ███████░░░  68                            │
│                                                              │
│   ⚡ Verdict: LEARN (with caveats)                           │
│   → Foundational ML is timeless, but content predates        │
│     the transformer era                                      │
│                                                              │
│   ✅ Strengths                                               │
│   · Exceptional pedagogy from Andrew Ng                      │
│   · Covers math foundations thoroughly                       │
│                                                              │
│   ⚠️  Concerns                                               │
│   · No coverage of LLMs / transformers                       │
│   · TensorFlow 1.x examples are outdated                     │
│                                                              │
│   💡 Better Alternatives Found                               │
│                                                              │
│   1. Deep RL Course (Hugging Face)     82  +14 ↑            │
│      ├─ Fresher (2025), hands-on code                        │
│      └─ youtube.com/...                                      │
│                                                              │
│   2. Spinning Up in Deep RL (OpenAI)   79  +11 ↑            │
│      ├─ Excellent info density, free                         │
│      └─ github.com/openai/spinningup                         │
│                                                              │
╰──────────────────────────────────────────────────────────────╯

Zero config. No API key needed. The killer feature — discovering better alternatives — works out of the box.

---

🚀 Quick start

# 1. install
pip install skillens

# 2. point it at anything
skillens "https://..."

# 3. there is no step 3

SkiLens auto-detects the platform and extracts the right metadata. No flags, no setup.

---

🎯 What gets evaluated

Dimension	What it measures	How
Market Demand	Is this skill actually in demand?	Bundled skill-demand dataset + popularity proxies
Skill Half-life	How long until this knowledge is stale?	Topic classification (foundational / established / fast-moving)
Info Density	How much signal per hour invested?	Syllabus analysis, unique-topic ratio
Freshness	Is the content up-to-date?	Published / last-updated dates, deprecation detection
Effort vs Return	Is the time investment justified?	Duration × demand cross-reference
Profile Match (optional)	How well does it fit your background?	Token-overlap scoring against your skills / target role

Scores are combined into a single 0–100 overall, mapped to one of three verdicts: LEARN, SKIP, CONSIDER ALTERNATIVES.

---

🌐 Works on everything

# URLs — platform auto-detected
skillens "https://www.coursera.org/learn/machine-learning"
skillens "https://youtube.com/watch?v=..."
skillens "https://github.com/openai/gym"
skillens "https://arxiv.org/abs/2301.00001"
skillens "https://any-blog-or-tutorial.com/..."

# Skill / topic — market analysis via web search
skillens topic "reinforcement learning"
skillens topic "rust programming"

# Side-by-side comparison
skillens "https://courseA" --compare "https://courseB"

# JSON output for scripting
skillens "https://..." --json | jq .verdict

Supported platforms

Platform	Status	What's extracted
🎓 Coursera	✅	Syllabus, rating, enrollment, instructor, institution (via JSON-LD)
📺 YouTube	✅	Title, views, duration, description, tags (via yt-dlp)
💻 GitHub	✅	Stars, activity, topics, README, language
📄 arXiv	✅	Abstract, authors, categories, dates
🌍 Generic webpage	✅	Open Graph / JSON-LD / meta fallback
🎨 Plugins	✅	Third-party providers via entry points — see below

---

🧠 Deep mode (optional LLM)

Quick mode (default) needs zero config. For nuanced analysis, plug in any LLM:

# OpenAI (default)
skillens config set llm openai
skillens config set api-key sk-...
skillens "https://..." --deep

# Anthropic Claude
skillens config set llm anthropic
skillens config set api-key sk-ant-...

# Local Ollama (100% private, $0)
skillens config set llm ollama
skillens config set model llama3.2

All three backends use structured outputs (OpenAI tool-calling, Anthropic tool_use, Ollama JSON schema) — the LLM can't go off-rails and the scores always validate against a Pydantic schema.

Deep mode falls back to quick-mode scoring automatically if the LLM call fails, so --deep is always safe to enable.

---

👤 Personalize

Tell SkiLens who you are once, and every evaluation gets a personal-fit score:

skillens profile set --skills "python,pytorch,rl,robotics" \
                     --role "ML engineer" \
                     --years 5

Market Demand    ████████░░  78
Freshness        █████░░░░░  48
Profile Match    █████████░  85   ← new: strong alignment with your stack

The matcher lives at skillens/profile/matcher.py — it's pure Python, no LLM needed. Sweet spot is 10–15% token overlap (new but not redundant); heavy overlap gets penalized as "you already know this".

---

🔌 Plugin ecosystem

Add a provider for any platform without forking SkiLens. Third-party packages register via entry points:

# your-package/pyproject.toml
[project.entry-points."skillens.providers"]
udemy = "my_package.udemy:UdemyProvider"

# my_package/udemy.py
from skillens.providers.base import BaseProvider
from skillens.core.models import ResourceMeta, SourceType

class UdemyProvider(BaseProvider):
    @property
    def name(self) -> str:
        return "udemy"

    @staticmethod
    def can_handle(url: str) -> bool:
        return "udemy.com/course/" in url

    async def extract(self, url: str) -> ResourceMeta:
        # fetch, parse, return ResourceMeta(...)
        ...

pip install the package — SkiLens auto-discovers it at runtime. Broken plugins fail silently with a stderr warning; they can't crash the CLI.

---

🤖 MCP server mode

Use SkiLens as a tool inside Claude Desktop, Cursor, or any MCP client:

pip install "skillens[mcp]"
skillens mcp   # speaks stdio by default

Three tools are exposed: evaluate_url(url), analyze_topic(skill), get_profile(). Drop this into your Claude Desktop config:

{
  "mcpServers": {
    "skillens": { "command": "skillens", "args": ["mcp"] }
  }
}

Now your agent can ask "is this course worth it?" mid-conversation and get a structured verdict back.

---

🌏 中文 / i18n

skillens "https://bilibili.com/video/BV1..." --lang zh

│   市场需求       ████████░░  78                             │
│   技能半衰期     ██████████  ~10y                           │
│   信息密度       ██████░░░░  62                             │
│                                                              │
│   ⚡ 结论: 值得学                                            │

Currently en and zh. Auto-detects Chinese from .bilibili.com / .zhihu.com / .cn URLs when --lang auto is set.

---

🛠️ Architecture

skillens/
├── cli.py               Typer CLI with bare-URL injection
├── core/
│   ├── evaluator.py     Pipeline orchestrator
│   ├── scorer.py        Rule-based + LLM scoring engine
│   ├── dataset.py       Bundled skill-demand dataset
│   ├── config.py        ~/.skillens/config.toml
│   └── models.py        Pydantic data models
├── providers/           Coursera · YouTube · GitHub · arXiv · Webpage + plugins
├── llm/                 OpenAI · Anthropic · Ollama backends + factory
├── discovery/           ★ Alternative-resource discovery via DDG
├── market/              Skill/topic trend analysis
├── profile/             Profile storage + token-overlap matcher
├── display/             Rich panels · JSON · compare table · i18n
├── mcp_server.py        FastMCP server exposing evaluate_url / analyze_topic
└── data/
    └── skill_demand.json  2026.04 curated keyword → demand map

---

📦 Install options

pip install skillens                    # base — zero config, works immediately
pip install "skillens[youtube]"         # adds yt-dlp for rich YouTube metadata
pip install "skillens[llm]"             # adds openai + anthropic
pip install "skillens[discover]"        # adds DuckDuckGo search for alternatives
pip install "skillens[mcp]"             # adds MCP server support
pip install "skillens[pdf]"             # adds PyMuPDF for local PDFs
pip install "skillens[all]"             # everything

---

🧪 Development

git clone https://github.com/YixiaJack/skillens
cd skillens
python -m venv .venv && . .venv/Scripts/activate   # Windows
pip install -e ".[dev]"
pytest                                               # 67 tests, ~0.4s

Pass rate: 67 / 67 ✅ (1 MCP test skips when the optional mcp extra isn't installed, as designed).

---

🤝 Contributing

The easiest contribution is a new provider. The Plugin ecosystem section shows the full template — you don't even need to fork SkiLens, just ship your own package with the right entry point.

Other welcome contributions:

• New language packs in display/i18n.py
• Skill-demand dataset updates in data/skill_demand.json
• New LLM backends in llm/

See CONTRIBUTING.md for full details.

---

🗺️ Roadmap

• Core CLI + rule-based scoring engine
• YouTube · Coursera · GitHub · arXiv · Webpage providers
• Rich terminal output with score bars
• JSON output mode
• Alternative-resource discovery (the ★ killer feature)
• OpenAI / Anthropic / Ollama deep mode
• Profile management + personal match scoring
• --compare side-by-side mode
• Bundled skill-demand dataset
• MCP server mode
• Chinese output (--lang zh)
• Provider plugin system via entry points
• VS Code extension wrapper
• Browser extension (evaluate as you browse)
• Obsidian plugin integration
• Auto-updating dataset via scheduled GitHub Action

---

🧭 Philosophy

In a world with infinite content, curation is the bottleneck.

SkiLens shifts evaluation from "after 40 hours" to "before the first hour."

Three invariants the project will never break:

1. 5-second first result. Rule-based scoring always beats the LLM to the punch.
2. Zero config for basic use. pip install && skillens URL must just work.
3. The terminal is the product. No web dashboard. No SaaS. Forever.

---

📄 License

MIT — do whatever you want with it.

---

If SkiLens saved you from a bad course, give it a ⭐

Built with 🔥 by @YixiaJack