A multilingual Agent Skill Pack that turns AI agents into guided learning coaches.
Instead of generating one-off explanations or dumping a folder of Markdown files, Learn Anything Skill helps an AI agent create a structured learning repository and immediately guide the user through the first learning session.
Learn Anything Skill helps AI agents build structured learning systems for any domain.
It can create:
- domain maps
- concept breakdowns
- daily learning plans for a custom duration (30 days by default)
- exercises
- quizzes
- flashcards
- progress tracking
- source tracking
- material-grounded learning plans
- final projects
- guided Day 1 learning sessions
Current release: v0.2.4-beta. See release notes and roadmap.
Most AI learning conversations are temporary. Most generated study plans are hard to continue. Most learning repositories are useful, but beginners often do not know where to start.
This project solves that by combining:
- A structured learning repository.
- A guided learning session that starts immediately.
- Progress tracking and review loops.
- Source notes and freshness checks.
- Support for user-provided PDFs, slides, notes, and documents.
By default, after creating a learning repository, the agent does not stop at a file summary.
It immediately starts Day 1 in the chat:
- explains today's goal
- introduces the first concept in beginner-friendly language
- gives one small task
- provides a copyable answer template
- shows how the learner can tell whether the answer works
- asks the user to reply directly in the conversation
- updates progress after the user completes the task
For beginners, complete beginners, non-technical users, students, content creators, writers, marketers, teachers, and people with no coding background, Guided Learning Mode switches to an interactive beginner lesson:
- one plain-language idea for the day
- a scenario based on the learner's real goal
- I do: a fully worked example
- We do: one bad example and one better example
- You do: one tiny task, usually one workflow step
If the user only wants files, they can say:
scaffold only
or:
只创建项目，不要开始学习
User:
I want to learn harness design.
My background: beginner.
Goal: understand the basic principles in 7 days and apply them to my content creation workflow.
Daily time: 1 hour.
Agent:
Learning project created.
Freshness Notice:
This project includes evolving practice-based content. Review every 3-6 months.
Detailed log: 09_sources/freshness_log.md
You do not need to open the files first. We will understand Day 1 here.
Today, learn just one sentence first:
A harness is a simple way to run the same workflow step again and check whether the result is good enough.
Put it into your real situation:
In a content workflow, you might use a harness to test one repeatable step, such as turning a topic idea into three article titles.
I do:
Workflow step: turn one topic idea into three article titles.
Expected output: three titles with a clear reader benefit.
Simple check: at least one title makes the reader immediately understand why they should click.
Bad example:
"Make my article better."
This is too broad. It does not say which step is being tested.
Better example:
"Take one topic idea and produce three article titles for middle-aged readers."
This is easier to test because the input, output, and audience are clear.
Your task:
Pick one small step from your own content creation workflow.
Copy and fill this template:
# My Day 1 Answer
1. My workflow step:
2. What I want this step to produce:
3. One simple way I can tell whether this step worked:
4. What I am still unsure about:
Please fill in the template above and send it to me directly.
Create a learning repository and immediately start the first learning session.
Beginner-friendly sessions use an interactive lesson format: I do, We do, You do. The agent gives a worked example, compares a bad example with a better example, then asks for one small answer in chat.
Generate a structured repository with maps, concepts, exercises, quizzes, projects, reviews, and progress tracking.
Track sources, claims, freshness risk, and unverified information.
This reduces hallucination risk, but does not guarantee absolute correctness.
Build a learning plan from user-provided materials such as PDFs, PPTs, Markdown files, notes, manuals, and exported webpages.
- English: en-US
- Simplified Chinese: zh-CN
Designed for:
- Codex
- Claude Code
- Cursor
- ChatGPT
- generic file-based agents
- generic chat-only agents
- Coze / 扣子
- WorkBuddy
- Trae
- CodeBuddy
Platform support depends on each platform's file access, web access, workflow, and knowledge base capabilities.
When a learning repository is created, the agent also prints a short freshness notice in the chat. This tells the learner whether the project contains time-sensitive content, where to find the freshness log, and which claims require verification. It also shows the recommended review interval, so the learner does not need to open generated files first to know that freshness tracking exists.
Stable foundational subjects use a brief notice. High-risk or fast-changing
subjects, such as finance, APIs, pricing, policies, platform rules, or current
tools, use a more explicit verification reminder and point to
09_sources/freshness_log.md and, when needed,
09_sources/claims_to_verify.md.
| Asking ChatGPT directly | Using Learn Anything Skill Pack |
|---|---|
| One-off conversations, knowledge doesn't stick | Everything is file-based, stored in a structured repository |
| AI tends to output explanatory prose | Built-in exercise + quiz + project systems ensure hands-on practice |
| No idea what you've learned or where you are | progress.md continuously tracks progress and weak points |
| Reinvent the method for every new domain | One Skill Pack, reusable forever |
| Wrong answer → "here's the right one" | Four-type error diagnosis → targeted remedial exercises |
| AI may sound confident without sources | Knowledge Reliability Layer tracks sources, unverified claims, and freshness risk |
The simplest path is to place this repository in a directory your agent can read:
git clone https://github.com/vesperchinn/learn-anything-skill.git
cd learn-anything-skillThen connect it based on your agent:
- Codex / Claude Code / Trae-style file agents: open this directory or add it to the agent-readable workspace.
- Agents with Skill support: import this repository as a Skill or place it in the Skills directory.
- Coze, WorkBuddy, CodeBuddy, and other Chinese agent platforms: follow the
adapter notes under
platforms/cn/. - Chat-only agents: copy the prompts from this repository when direct installation is not available.
After installation, type:
"AI Agents" is only an example. Replace it with the subject you actually want to learn, such as "Python", "nutrition", "photography", or "English writing".
Use learn-anything to create a learning project for "AI Agents".
My background: beginner.
My goal: understand the basics and build a small project in 14 days.
Daily time: 1 hour.
The agent should create the learning repository and then immediately start Day 1 in the chat. You do not need to open the generated Markdown files first.
If you already have PDFs, slides, notes, or course material, say:
Use learn-anything to create a learning project from my materials.
Prioritize the provided materials and mark anything that still needs verification.
If your agent cannot call the Skill by name but can read files, type:
Again, replace "AI Agents" with your real learning subject.
Read learn-anything-skill/core/prompts/en-US/init-repo.md.
Create a learning repository for "AI Agents".
By default, repository creation continues into a guided Day 1 session. If you
only want files, say scaffold only or generate files only.
Continue with learn-anything. Read my progress and run today's learning session.
Continue with learn-anything. Review what I learned today and update my progress.
Continue with learn-anything. Give me a stage test. Ask questions first, then grade after I answer.
./scripts/new-domain.sh "Your Subject" en-USFor example:
./scripts/new-domain.sh "AI Agent" en-US
cd learn-ai-agentSee docs/quick-start.md for the full guide.
Learn Anything now includes a Platform Adapter Layer for platforms that cannot consume the native file-based Skill workflow directly. Low-code platform support is experimental in this beta; validate each adapter in your own workspace before relying on it.
| Form | Target platforms | How it works |
|---|---|---|
| File-based Agent / native Skill | Codex, Claude Code, Cursor, Trae, and other file-based agents that can read this repo | Reads SKILL.md, core/, templates/, prompts/, and references/; Codex uses AGENTS.md, Claude Code uses CLAUDE.md; writes the learning repo and starts guided learning directly |
| Platform package | Coze, WorkBuddy, Trae, CodeBuddy, generic low-code agents | Uses platform-specific prompts, knowledge-base packages, workflows, variables, memory, and test checklists under platforms/ |
| Chat-only package | Ordinary chat agents | Copies the core protocols and outputs path-labeled Markdown blocks |
See platforms/README.md, platforms/capability-matrix.md, and dist/README.md.
| Platform | Adapter | Recommended form | File writing | Main limitation |
|---|---|---|---|---|
| Coze | platforms/cn/coze/ | Bot + knowledge base + workflow + variables + memory | Usually no local file writing | Do not assume it can read SKILL.md; split into prompt, KB, workflow |
| WorkBuddy | platforms/cn/workbuddy/ | Office task Skill + report output | Depends on task environment | Best for reports, task sheets, and material processing |
| Trae | platforms/cn/trae/ | File-based engineering Agent | Yes | Can preserve direct repo reading |
| CodeBuddy | platforms/cn/codebuddy/ | Code/document Agent + knowledge base | Yes when repo-connected | Package references, templates, and prompts into a KB |
| Generic low-code Agent | platforms/cn/generic-lowcode-agent/ | System prompt + workflow + KB + state | Usually no | Needs explicit fallback for no file read, no web, or no workflow |
Capabilities differ by platform, product version, workspace policy, and enabled connectors. File-based agents can create and maintain a learning repository. Low-code platforms usually approximate the workflow through knowledge bases, workflows, variables, memory, and prompts. Chat-only agents can only output copyable Markdown and compact state summaries.
If you already have course PDFs, slide decks, notes, documentation exports, or webpage exports, use Material-Grounded Learning Mode:
- Put original files in
learning_materials/raw/, or tell the agent where the files are. - Run
prompts/{locale}/material-intake.mdto register and extract the materials. - Run
prompts/{locale}/material-grounded-learning-repo.mdto build the knowledge map, plan, concepts, quizzes, reviews, and progress tracking from those materials. - Use
material_coverage_map.mdto see which modules are grounded in the materials, partially grounded, or supplemental.
In this mode, user-provided materials are the primary source. Outside knowledge
must be labeled Supplemental. If a PDF/PPT chart, screenshot, table, or
flowchart cannot be extracted, the issue is recorded in
learning_materials/extraction_issues.md rather than guessed.
If the agent cannot read files, paste the text, provide OCR, convert the files to Markdown/TXT, export slides as text plus images, or ask for a material processing checklist.
Do not put confidential company documents, contracts, private health or financial records, paid course materials, unpublished manuscripts, or copyrighted books into a public learning repository. Keep sensitive materials in a private repo, remove personal data before extraction, and make sure you have the right to store and transform the material.
learn-anything-skill/
├── SKILL.md # Skill entry point (routing file for agents)
├── README.md # English homepage (you are here)
├── README.zh-CN.md # Chinese homepage
├── core/ # Core prompts (agent-agnostic)
│ ├── prompts/
│ │ ├── en-US/ # 13 English prompt modules
│ │ └── zh-CN/ # 13 Chinese prompt modules
│ ├── *-protocol.*.md # Platform-neutral protocols
│ └── principles.md # Learning principles
├── templates/ # Learning repo templates
│ ├── en-US/ # English (default)
│ └── zh-CN/ # Chinese
├── references/ # Methodology references
│ ├── en-US/ # English
│ └── zh-CN/ # Chinese
├── examples/ # Complete example repositories
│ ├── en-US/learn-ai-agent/ # English example
│ └── zh-CN/learn-ai-agent/ # Chinese example
├── adapters/ # Cross-agent adaptation guides
├── platforms/ # Platform adapters (Coze / WorkBuddy / Trae / CodeBuddy, etc.)
├── dist/ # Distribution manifests and build notes
├── prompts/ # Material-grounded learning prompts
├── skills/codex/ # Legacy wrapper / compatibility files
├── scripts/ # Automation scripts
├── evals/ # Test suites
│ ├── en-US/ # English eval cases
│ └── zh-CN/ # Chinese eval cases
└── docs/ # User documentation
We provide a suite of Python tools in the scripts/ directory to enhance your learning experience:
init_learning_repo.py: Cross-platform repository scaffolding (Windows/Mac/Linux).generate_index.py: Dynamically generates a Table of Contents (index.md) for your learning repository.export_flashcards.py: Extracts flashcards into an Anki-compatible CSV file.validate_locale.py: Detects "language bleed" (e.g. Chinese text in an English repo) using character heuristics.check_unverified_claims.py: Finds[unverified]and unverified-draft markers that still need review.check_stale_modules.py: Checks09_sources/freshness_log.mdfor modules past their review date.check_source_notes.py: Ensures learning modules include Source Notes, Freshness Risk, Claims to Verify, Last Verified, and Recommended Review Interval.
Both scaffolding scripts support --dry-run and refuse to overwrite an existing
learn-{domain-slug} directory.
The read-only maintenance guard layer lives in harness/. It is not a new learning feature; it helps maintainers catch structure drift, locale mismatch, platform adapter gaps, material-grounding gaps, and reliability-rule gaps before release.
Maintenance Loop is for maintainers only; it does not change ordinary learner sessions or add default learner loops.
Run all checks:
python3 harness/scripts/run_all_checks.py --root . --reportReports are written to harness/reports/ with timestamped filenames and never overwrite older reports. PASS means OK, WARN means human review is needed, and FAIL means the issue should be resolved before release.
Before changing SKILL.md, review change-impact-matrix.md and run check_skill_manifest.py, check_docs_consistency.py, and check_eval_coverage.py. Before adding a platform adapter, use platform-adapter-checklist.md. Before release, use release-checklist.md and release-gates.md.
Learn Anything includes a Knowledge Reliability Layer for generated learning repositories:
- Source-first policy: claims should be backed by primary or authoritative sources. The agent must not fabricate URLs, papers, publication dates, official documents, or benchmark results.
- No source, no claim: unsupported claims are marked
[unverified]or moved to09_sources/claims_to_verify.md. - Freshness risk: each module is tagged as 🟢 Stable, 🟡 Evolving, or 🔴 Volatile, with a recommended review interval.
- No-web fallback: if the agent cannot browse or search, generated material is labeled Unverified Draft and a verification checklist is produced.
- High-stakes domains: medical, legal, financial, safety-critical, cybersecurity, and certification content requires an educational-use-only notice and authoritative sources first.
- Private or copyrighted materials: keep them out of public repositories unless you have permission and have removed sensitive data.
| Agent | Support Level | Adapter |
|---|---|---|
| Codex | Full (native Skill) | codex.md |
| Claude Code | Documented workflow (CLAUDE.md) | claude-code.md |
| Cursor | Documented workflow (.cursorrules) | cursor.md |
| ChatGPT | Copy-paste prompts | chatgpt.md |
| Generic Agent | Manual prompt copy | generic-agent.md |
| Capability | Codex / Trae / file agents | Coze / WorkBuddy / CodeBuddy KB mode | Chat-only agents |
|---|---|---|---|
| Read repository files | Yes | Usually no, unless uploaded to KB | No |
| Write learning repo files | Yes | Usually report or platform output only | No |
| Learn from materials | Direct file reading | Upload or knowledge base | Pasted text/OCR |
| Source records | 09_sources/ files |
Reports, variables, or memory | Conversation summary |
| Workflow | Agent execution | Platform workflow | Manual multi-turn chat |
| Fallback | Path-labeled blocks when files are unavailable | KB/report mode when plugins are unavailable | learning_state plus Markdown blocks |
This Skill Pack is built on five core systems:
- Knowledge Map — solves "I don't know what's in this field"
- Glossary — solves "I don't understand the terminology"
- Exercise System — solves "I thought I understood but I didn't"
- Project System — solves "I learned a lot but can't apply it"
- Review System — solves "I forgot it all and never fixed my mistakes"
See references/en-US/learning-principles.md.
| Locale | Interface | Materials | Status |
|---|---|---|---|
en-US |
English | English | ✅ Complete |
zh-CN |
中文 | 中文 | ✅ Complete |
The {interface_language} and {learning_language} can be set independently.
For example: "Chat in Chinese but build the learning repo in English."
See SKILL.md § Language and Locale Policy.
Contributions welcome — new adapters, templates, examples, or prompt improvements. See CONTRIBUTING.md or CONTRIBUTING.zh-CN.md.
Inspired by @GeekCatX's article on using Codex to rapidly learn any field.
MIT © 2026 Learn Anything Skill Pack Contributors