English · 简体中文
Quick start · See it work · Example repo · User guide · v0.2.4-beta
Learn Anything is an open-source Agent Skill Pack for learning any subject with an AI agent. It creates a durable learning repository, starts teaching immediately in the conversation, and keeps the next lesson tied to your real progress.
Tell the agent what you want to learn, your background, your goal, and your available time:
Use learn-anything to create a learning project for "harness design".
My background: complete beginner.
Goal: understand the basics in 7 days and apply them to my content workflow.
Daily time: 1 hour.
The agent then:
- creates a domain map, lessons, exercises, quizzes, a final project, and source records;
- writes
START_HERE.mdandTODAY.mdso the next step is obvious; - starts Day 1 in the chat with one explanation, one example, and one small task;
- checks your reply before updating
progress.mdand the next session.
See a complete Day 1 transcript or browse a generated AI Agent learning repository.
If you only want the files, add scaffold only or generate files only.
You do not need to open the generated files first. Unless you ask for scaffold-only output, Day 1 starts immediately in the chat with a plain-language idea, a worked example, a small task, a copyable answer template, and a clear way to check your answer.
git clone https://github.com/vesperchinn/learn-anything-skill.git
cd learn-anything-skillOpen this directory in Codex, Claude Code, Cursor, Trae, or another file-capable agent. If your agent supports Skills directly, import this repository or place it in its Skills directory.
Use learn-anything to create a learning project for "Python".
My background: beginner.
My goal: build a small automation in 14 days.
Daily time: 45 minutes.
Already have PDFs, slides, notes, or course material? Use:
Use learn-anything to create a learning project from my materials.
Prioritize the provided materials and mark anything that still needs verification.
Continue with learn-anything. Read my progress and run today's learning session.
The command-line scaffold is also available:
./scripts/new-domain.sh "Your Subject" en-USSee the quick-start guide for more installation paths and fallback instructions.
New learners get a guided I do → We do → You do lesson instead of a long lecture. Every session has a concrete task and a visible completion standard. Stage tests revisit weak points, while progress.md and progress-log.md keep the learning state outside a disposable chat.
| A one-off AI chat | Learn Anything |
|---|---|
| Explains a topic once | Builds a learning path you can resume |
| Gives information before practice | Teaches, demonstrates, then asks you to try |
| Forgets weak points between chats | Tracks progress, errors, and next steps in files |
| Can sound certain without evidence | Records sources, freshness, and claims to verify |
The method combines five systems: a knowledge map, glossary, exercises, a final project, and review loops. The rationale lives in learning principles.
The material-grounded path works with PDFs, slides, Markdown, notes, manuals, and webpage exports.
- Your materials remain the primary source for the learning plan.
- Outside knowledge is labeled
Supplementalinstead of being blended in silently. material_coverage_map.mdshows what is grounded, partial, or missing.- Unreadable charts, screenshots, and tables are recorded in
learning_materials/extraction_issues.mdinstead of guessed.
Keep confidential documents, personal data, paid course materials, and copyrighted books out of public learning repositories unless you have permission to store and transform them.
- Source first: no fabricated URLs, papers, dates, or benchmarks.
- No source, no claim: unsupported material is marked
[unverified]or moved to09_sources/claims_to_verify.md. - Freshness visible: every module records stability risk and a recommended review interval.
- Safe fallback: without web access, the output is labeled Unverified Draft and includes a verification checklist.
- High-stakes caution: medical, legal, financial, safety, cybersecurity, and certification topics require an educational-use notice and authoritative sources.
This reduces hallucination risk; it does not guarantee absolute correctness.
When a learning repository is created, the chat includes a short Freshness Notice before Day 1. It shows the highest freshness risk, the recommended review interval, and where to find 09_sources/freshness_log.md; fast-changing or high-risk projects also point to 09_sources/claims_to_verify.md.
| Setup | Examples | What to use |
|---|---|---|
| File-based agent / native Skill | Codex, Claude Code, Cursor, Trae | Repository root SKILL.md plus the included prompts, templates, and references |
| Platform or knowledge-base workflow | Coze, WorkBuddy, CodeBuddy | Packages and platform notes under platforms/ |
| Chat-only agent | ChatGPT or any text-only assistant | Copyable prompts and path-labeled Markdown output |
File access, web access, workflow support, and persistence differ by platform. Low-code adapters are experimental in this beta; review the capability matrix before relying on one.
English (en-US) and Simplified Chinese (zh-CN) are complete. Interface language and learning-material language can be set independently.
learn-anything-skill/
├── SKILL.md # Agent entry point and routing rules
├── core/ # Core prompts and learning protocols
├── templates/ # Complete learning-repository templates
├── examples/ # Generated repositories and session transcripts
├── prompts/ # Material-grounded learning workflows
├── references/ # Learning and reliability methods
├── adapters/ # Agent-specific setup guides
├── platforms/ # Low-code and knowledge-base adapters
├── scripts/ # Scaffolding and validation tools
├── evals/ # Behavior checks
└── harness/ # Read-only maintenance and release checks
- Quick start · User guide
- Release notes · Roadmap · Changelog
- Agent guides: Codex · Claude Code · Cursor · ChatGPT · Generic
Run the repository's read-only release checks with:
python3 harness/scripts/run_all_checks.py --root . --reportContributions are welcome: new adapters, templates, examples, tests, and prompt improvements. Read CONTRIBUTING.md before opening a change.
Inspired by @GeekCatX's article about using Codex to learn a new field quickly.
Use learn-anything to create a learning project for "the subject I want to learn".
MIT © 2026 Learn Anything Skill Pack Contributors