Wisdom
A Claude Code plugin that turns Claude into a tutor for any skill, starting from wherever you are, even zero.
Wisdom finds out where you are on your journey, then teaches you fully: the why, the building blocks, and how each new idea connects to what you already know. You practice, and the questions get harder as you get better. Alongside, you work through a great source: a book, a video course, or a web course, chosen for you. Claude reads every chapter, or watches every video through its transcript and frames, before it assigns you a page or a timestamp, and lessons quote it directly. You build a real project, and Claude writes the plumbing and leaves the interesting core parts for you. Everything you learn comes back just before you'd forget it, and your progress, notes, and streak live in a .wisdom/ folder in your project.
Get started
claude plugin marketplace add suncloudsmoon/plugins
claude plugin install wisdom@suncloudsmoon
Restart Claude Code after installing. To try it from a local checkout instead, run claude --plugin-dir /path/to/wisdom.
Then, in the repo where you're building something:
/wisdom:tutor Claude wrote most of this and I want to actually understand it. Teach me Python properly while we keep building.
Wisdom reads your code to see where you are, asks what it can't tell from the code, and takes the app you're building as your project. It picks the best main source for your level and stack (a book, a video course, or a web course), then teaches your first concept in the same sitting, starting from a line in your own code. For a book, it creates .wisdom/<source>/ and asks you to put your copy there, in any format: PDF, EPUB, photos of the pages, or nearly anything else. It downloads videos and saves web pages there itself. .wisdom/ keeps itself out of git: wisdom writes a .gitignore inside it, so your excerpts, videos and progress are never committed by accident, and your own .gitignore stays untouched.
You can also start a skill with no app at all, like /wisdom:tutor I want to learn Spanish, and wisdom proposes a project for you.
After that, just open Claude Code in the project and say hi. The start-of-session hook tells Claude where you are ("6 reviews due, chapter 5 pages 95–105 assigned, streak 4 days"), and a session picks up from there.
Learning while Claude builds
Most days you won't ask for a lesson. You'll ask Claude to build things: "add streaming, a --model flag, and a cost tracker." In a project with .wisdom/, Claude builds them and you learn along the way. It explains new ideas where they show up in the work, then hands you the one part of each request that holds the most learning, as a TODO(human) with a test that fails until your code works. When you're stuck, you get one hint at a time, never the answer unless you ask for it, because the struggle is what makes it stick. Say "just build it" when you only want it built.
A session
- Recap: you recall what you did last time before Claude reminds you.
- Warm-up: the reviews that are due today, mixed across topics.
- One new concept, taught fully: hook, building blocks, the idea, the why, a worked example, then your turn.
- Practice that climbs from recognizing to recalling to applying to combining ideas.
- Reading: pages Claude has already read, with questions to read for.
- Project: the next core part, marked
TODO(human) in your code, with a test that tells you when it works.
- Celebration and a journal entry: XP, streak, level-ups, mastered chapters, and fun facts along the way.
Start each study session in a fresh conversation (/clear). Everything wisdom needs is saved in .wisdom/, and the start-of-session hook loads it, so you lose nothing, and a fresh conversation runs faster and costs less than one long thread.
How it gets harder
- When: every concept has its own memory model (FSRS-6, the scheduler Anki uses), so it comes back right before you'd forget it.
- How hard: each concept climbs four rungs: recognize, recall, apply, transfer. Two confident right answers move it up, and a miss moves it down one.
- How fast: Wisdom aims for about 85% of your answers being right. When you're above 90% it speeds up, and when you're below 70% it slows down and rebuilds the basics.
- Honest grading: you say how sure you are with every answer. A lucky guess counts as not-yet-known. A confident mistake gets special attention, because those are the errors that stick best once they're corrected.
- Project assignments grow from 5–10 lines, to whole functions, to modules, to whole features as you prove each size.
Commands
| Command |
What it does |
/wisdom:tutor |
Start or continue a session. Add review, read, project, or progress to jump straight to that part. |
/wisdom:explain |
Re-explain the last concept simply, with an analogy |
/wisdom:quiz |
2–3 questions on what was just built, recorded in your reviews |
/wisdom:map |
ASCII diagram of how the pieces connect, in the code or in what you've learned |
/wisdom:why |
Deep dive on the last decision and its alternatives |
/wisdom:teach |
Maximum explanation mode |
/wisdom:speed |
Minimal explanation mode |
Works with the Learning output style
Wisdom uses the same conventions as Claude Code's built-in Learning output style: TODO(human) markers, "● Learn by Doing" requests, and ★ Insight blocks. When Learning is on (/output-style learning), wisdom follows it and adds the curriculum: which parts to hand you, and what to review. When Learning is off, wisdom supplies Anthropic's Learning instructions itself, so the experience is the same.
Outside projects with a .wisdom/ folder, Wisdom adds nothing to your sessions.
What's in .wisdom/
.wisdom/
├── learner.md your goal, starting point, daily time, interests
├── journal.md one entry per session
├── progress.json reviews, XP, streak, assignments, project stage
├── lessons/ lesson files (they quote your sources, so they stay here)
└── <source>/
├── <your book file, or the downloaded video and transcript, or saved pages>
└── notes/ Claude's notes on each part, written before assigning it
Requirements
Claude Code and Python 3.9 or later. Optional tools make sources cheaper and richer. pdftotext (from poppler) lets wisdom read text PDFs as text instead of page images. yt-dlp (or uv, which runs it without installing anything) and ffmpeg handle video courses. Wisdom asks before anything gets installed.
Recommended: humanizer
Wisdom has its own writing guide, so lessons read like a person wrote them even when it runs alone. With the humanizer plugin installed, the tutor also runs every lesson through humanizer's rewrite pass, which catches the habits that make AI prose easy to spot. Wisdom finds it automatically.
claude plugin marketplace add blader/humanizer
claude plugin install humanizer@humanizer
Credits
Wisdom builds on two plugins, copied in full and extended:
- learn by varma (MIT): the teaching layer and the explain, quiz, map, why, teach, and speed commands.
- learning-output-style by Anthropic (Apache-2.0): the Learning and Explanatory instructions.
The tutor's writing rules draw on Wikipedia's Signs of AI writing and humanizer by Siqi Chen (MIT). It summarizes them in its own words and copies no text from either.
The learning science behind each design choice, and what it borrows from Duolingo, Anki, Khan Academy, CodeCrafters, and quizzer, is in docs/learning-science.md. See LEGAL.txt and NOTICE.txt for licensing details.
License
MIT. hooks/learning-style.md is Apache-2.0.