The domain expert your AI agents need. One repo. Always current.
Super Skill is an open framework for creating private domain experts for you and your AI agents.
A domain is any area of knowledge you work in: your tech stack, your investment portfolio, your garden, your marketing strategy, your fitness plan. Anything you know about and want your AI to know too.
Each Super Skill is an intelligence layer: a self-growing knowledge base that your AI agents read before every task. It remembers what you have decided, what tools you use, what is waiting for your approval, and what must never be re-argued. Your AI stops starting from zero.
With NotebookLM (optional), your Super Skill also teaches you back: audio overviews, quizzes, and mind maps generated from your own domain knowledge.
Multiple agents. One source of truth. Always current.
Early-stage open-source project. No SLA, no commercial support. AI outputs are not guaranteed to be accurate or suitable for any purpose. Verify all AI-generated content before use. MIT License -- no warranty of any kind.
git clone https://github.com/Asaf-Dahan/super-skill.git super-skill-mydomain
cd super-skill-mydomain
claude .Paste the output of cat scripts/bootstrap.md into Claude Code.
Answer three questions. Your Super Skill is live in under 15 minutes.
New to Super Skill? Read on.
You need four things installed on your computer. If any of these are missing, the Bootstrap prompt below will detect it and tell you exactly what to install before continuing.
-
Claude Code -- the AI that builds your Super Skill. Install via terminal: npm install -g @anthropic-ai/claude-code Full guide: https://docs.anthropic.com/en/docs/claude-code
-
Git -- downloads the Super Skill template to your computer. Install from: https://git-scm.com/download Windows: run the installer, keep all default settings. Mac: Git is pre-installed on most Macs.
-
Python (minimum version: Python 3.10) -- runs the Super Skill automation scripts. Install from: https://python.org/downloads Windows: check the box "Add Python to PATH" before clicking Install. Mac: Python 3 is pre-installed on most Macs.
Note on
pythonvspython3: macOS and Linux usually exposepython3, while Windows installs Python aspython. Throughout this README we usepython. If you are on macOS or Linux andpythonis not found, usepython3instead. You can also run any script via the cross-platform launcher, which auto-detects your interpreter: python scripts/run.pyFor NotebookLM integration (optional): after cloning, run: pip install -r requirements.txt This is only needed if you plan to use the NotebookLM learning layer. Core activation works without it.
-
A free GitHub account -- stores your Super Skill safely online. Create one at: https://github.com You will need this if you want to save your work to the cloud.
Three steps. One prompt. Under 15 minutes (if Git, Python, and Claude Code are already installed).
Create a folder on your computer called super-skills.
This is where all your Super Skills will live.
Windows: right-click on your Desktop, select New > Folder, name it super-skills.
Mac: right-click on your Desktop, select New Folder, name it super-skills.
Open a terminal in your super-skills folder and run:
claude .
In the Claude Code terminal, run:
Mac/Linux: cat scripts/bootstrap.md
Windows (Command Prompt): type scripts\bootstrap.md
Windows (PowerShell): Get-Content scripts\bootstrap.md
Copy the full output and paste it into Claude Code.
Claude Code will check your setup, ask you about your domain, offer to read your existing files or AI context for personalization, download the template, and build your complete Super Skill automatically.
After activation, run python scripts/doctor.py to confirm setup.
| When you want to... | Do this |
|---|---|
| Add a second Super Skill | Run Prompt 0 from ONBOARDING.md in a new terminal |
| Set up the global router | Run Prompt 2 from ONBOARDING.md |
| Connect NotebookLM | Follow NOTEBOOKLM_GUIDE.md |
| Set up automatic sync | Follow SYNC_SETUP.md |
| Verify your setup | Run python scripts/doctor.py |
| View your dashboard | Run /ss-dashboard |
Once your Super Skill is active, you never need to copy prompts again. Everything runs through slash commands. Type any of these into Claude Code:
| Command | What it does | Example |
|---|---|---|
| /ss-eval | Evaluate a new tool, method, or approach against your domain | /ss-eval Stripe Connect |
| /ss-graph | Generate the interactive knowledge graph | /ss-graph |
| /ss-learn | Generate a learning module on any topic in your domain | /ss-learn drip irrigation |
| /ss-pending | Show all items waiting for your approval | /ss-pending |
| /ss-query | Search wiki notes by keyword | /ss-query caching |
| /ss-sync | Pull updates from all your Super Skills and check for drift | /ss-sync |
| /ss-summary | Regenerate your SUMMARY.md from all layer files | /ss-summary |
| /ss-drift | Check your monitored sources for changes | /ss-drift weekly |
| /ss-council | Show your Expert Council debates and open challenges | /ss-council |
| /ss-expert | Run a structured expert debate session | /ss-expert DEBATE-001 |
| /ss-synthesize | Synthesize insights across multiple layers | /ss-synthesize tax efficiency |
| /ss-dashboard | Generate the interactive dashboard for one or all Super Skills | /ss-dashboard |
Claude Code Channels lets you send slash commands to a running Claude Code session from your phone via Telegram or Discord. Review PENDING items, trigger drift checks, and approve proposals -- all without opening a laptop. Your data stays local; the bot is a remote control for your existing session.
Setup requires tmux (to keep the session alive) and a Telegram bot (free, takes 5 minutes). Full guide: CHANNELS_GUIDE.md
| Script | Purpose |
|---|---|
doctor.py |
Confirms your setup is complete |
update_summary.py |
Regenerates SUMMARY.md from the layer files |
super-skill-sync.py |
Pulls all configured Super Skills and checks drift URLs |
scheduled_tasks.py |
Runs maintenance tasks weekly/monthly/quarterly |
generate_graph.py |
Generates wiki/graph.json and wiki/knowledge-graph.html |
generate_global_graph.py |
Merges graphs from all local Super Skills into a global view |
feed_notebook.py |
(NotebookLM, optional) uploads layer files as sources |
generate_learning.py |
(NotebookLM, optional) generates audio / quiz / mind map |
check_state.py |
(NotebookLM, optional) asks the notebook 3 standard questions |
generate_dashboard.py |
Generates wiki/dashboard.html with health scores, insights, and trends |
setup_channels.py |
(Channels, optional) checks prerequisites for mobile access |
run.py |
Cross-platform launcher: python scripts/run.py <name> |
/ss-graph generates two files: an interactive visual (wiki/knowledge-graph.html) and a machine-readable map (wiki/graph.json). Open the HTML file in any browser to explore how your domain's knowledge is connected. Claude Code reads graph.json to navigate your domain without loading all layer files every session.
To regenerate after any major change: /ss-graph
To generate a global graph across all your Super Skills: python scripts/generate_global_graph.py
Each Super Skill lives in its own folder and tracks one domain. To add a second domain, run Prompt 0 from ONBOARDING.md in a new terminal window. Prompt 0 creates a fresh Super Skill folder alongside your existing one -- your first Super Skill is not affected.
Once you have two or more Super Skills, you can merge them into a single knowledge map by running:
python scripts/generate_global_graph.py
The output opens at ~/.claude/global-graph.html.
To keep multiple Super Skills in sync, see SYNC_SETUP.md.
| Skill | Super Skill | |
|---|---|---|
| What it is | A file | A living knowledge architecture |
| Knows how to | Do a task | Understand a full domain |
| Self-growing | No | Yes |
| Detects drift | No | Yes |
| Teaches the agent | Yes | Yes |
| Teaches you | No | Yes -- audio, quiz, mind map |
| Works with | Claude Code | Claude Code, Cursor, Codex, Gemini, any agent |
| Time to activate | 1 minute | Under 15 minutes |
Root vs
template/-- The layer files at the repository root are your working copy: Prompt 1 fills them in during activation. Thetemplate/directory contains the same files as a clean scaffold -- use it to regenerate or compare against the originals. Do not edit anything intemplate/directly.
Every Super Skill contains the same core files regardless of domain. The structure is identical. The content is yours.
SUMMARY.md session entry point -- 80 lines, always current
SKILL.md agent skills format entry point
CLAUDE.md Claude Code operating instructions
AGENTS.md instructions for any other agent
CONTEXT.md layer 0: identity, principles, scope
DOMAIN_MAP.md layer 1: domain structure and sub-domain relationships
CURRENT_STATE.md layer 2: verified current state
EVALUATION.md layer 3: framework for evaluating anything new
DECISIONS.md layer 4: decisions log and reasoning
MONITORING.md layer 5: drift detection sources
LEARNING.md layer 6: NotebookLM notebook plan
PENDING.md layer 7: approval queue
LOG.md change log and audit trail
experts/COUNCIL.md layer X: expert council and debates
experts/[name].md individual expert profiles
notebooks/ NotebookLM-ready learning files
(gitignored by design -- user-generated, not part of template)
wiki/ quick-reference notes, searchable via /ss-query
wiki/graph.json machine-readable domain map (generated by /ss-graph)
wiki/knowledge-graph.html interactive visual (open in browser)
template/SCHEDULE.md scheduled maintenance tasks
scripts/ automation: feed, generate, check, sync, update_summary, scheduled_tasks
template/.claude/commands/ slash commands for Claude Code
template/experts/debates/ parallel expert debate sessions
The model proposes. The user decides. The Super Skill records. The system executes.
No layer file changes without your approval. Every proposed change goes to PENDING.md first. You review, you decide, then and only then it is recorded.
NotebookLM is optional but transforms what a Super Skill can do.
Without NotebookLM: your agents read Markdown files from the repository. Full context, zero re-explanation, every session.
With NotebookLM: the knowledge base grows beyond static files. You and your agents can add research, documents, and new sources. Claude proposes additions with your approval. You add directly in the notebook. The notebook generates audio overviews, quizzes, and mind maps from your own domain knowledge. Token usage drops significantly because Claude queries the notebook instead of loading full files.
Connect it via the NotebookLM section in ONBOARDING.md. Skip it and add it later if you prefer.
During activation, Claude Code researches your domain and proposes 10 real, leading experts with verifiable published work. You choose who joins your council. Expert profiles are structured representations based on public information. They are not endorsed by the experts named and do not represent their official positions. Each expert is assigned to the layer they contribute most to and given a full profile with methodology, frameworks, red lines, and five questions they would ask you.
When experts disagree, the disagreement is recorded as a debate in experts/COUNCIL.md. Claude Code surfaces expert perspectives only when genuine friction exists -- not on routine tasks.
If an expert's position conflicts with a decision in DECISIONS.md, it goes to PENDING.md as an expert challenge for your review.
Super Skill works for any domain where knowledge matters and decisions accumulate.
Technical: software infrastructure, database architecture, security frameworks, DevOps pipelines, product methodology, AI agent systems.
Professional: investment portfolio, legal review, medical documentation, marketing strategy, real estate, financial modeling, operations.
Personal: home garden, fitness, nutrition, home renovation, language learning, travel planning, personal finance.
Craft: photography, woodworking, wine, cooking techniques, academic research.
If you can describe it, a Super Skill can master it.
Super Skill v2.7.1
MIT License -- Fork it. Build your own.
Super Skill is an open-source public template. No affiliation or ownership is implied beyond authorship of the original template.
Fork it. Build your own. Publish your Super Skills.
notebooklm-py is an unofficial library that uses undocumented Google APIs. It may change without notice. Suitable for personal and internal use. Not recommended for production systems serving external users.
Keep domain-specific Super Skills in private repositories. Never commit API keys or tokens to any layer file. Use .env for all sensitive values.
Do not put real people's names, email addresses, phone numbers, or financial account details in any layer file. Layer files committed to a repository -- even a private one -- should contain no personally identifiable information.
Claude Code sessions may be logged by Anthropic per their usage policy. Do not enter secrets, credentials, or sensitive personal data into Claude Code prompts.
All agent actions follow one rule: The model proposes. The user decides. The Super Skill records. The system executes. Nothing changes without your approval.
Super Skill was built by Asaf Dahan, AI Solutions Architect at Gitit Inc.
Company: Gitit Inc GitHub: https://github.com/Asaf-Dahan