Version: 2.7.1 Status: Canonical Language: English Author: Asaf Dahan Date: April 2026
A Skill is a file. A Super Skill is a living knowledge architecture.
A Skill tells an AI how to perform a task. A Super Skill gives an AI a complete, current, layered understanding of a domain - including how to verify its own knowledge, how to detect when that knowledge is outdated, how to generate learning materials from it, and how to remain the most accurate and useful version of itself over time.
A Super Skill is a portable, AI-native intelligence layer that gives
any language model full context, methodology, and operational capability
over a defined domain.
A Super Skill is the new Domain Expert for you and your AI.
Every existing approach to AI personalization is static at its core.
Custom instructions are written once and forgotten. RAG retrieves documents but does not understand them. RAG (Retrieval-Augmented Generation) is a method where an AI searches a document collection before answering. It retrieves but does not understand. Agent Skills define how to perform a task but not how to understand a field. Fine-tuning burns knowledge into a model that cannot update itself. Fine-tuning permanently modifies an AI model using training data. The changes cannot be updated without retraining the entire model.
A Super Skill is none of these things.
It is a growing repository of structured knowledge, organized in layers, connected to live sources, capable of generating learning artifacts, and designed to be operated by any language model on behalf of any user who defines their domain and their context.
The model that reads a Super Skill does not just know what to do. It knows why things are done this way, what has been decided and why, what is being watched, what is uncertain, and what needs user approval before anything changes.
2022 Custom Instructions
Static. Manual. Single-session. No memory.
2023 System Prompts and RAG
Documents fed into context. Better, but not portable.
Not self-updating. Not evaluative.
2024 Model Context Protocol (MCP)
MCP is a standard that lets AI agents connect to external tools and services.
Standardized tool connectivity. The agent can act.
But it still does not know what it is.
Oct 2025 Agent Skills
SKILL.md files. Modular instructions. Portable.
Adopted by 16+ tools. A genuine open standard.
Tells the agent HOW to perform a specific task.
Dec 2025 Agent Skills Open Standard (Anthropic)
Released to the ecosystem. VS Code, Cursor, GitHub,
OpenAI, Microsoft all adopted within weeks.
Apr 2026 Super Skill
The layer above.
Tells the agent what the DOMAIN is, how it works,
what has been decided, what to watch, what to learn,
and how to keep growing.
The gap that exists today: Agent Skills define capability per task. No standard exists for domain-level intelligence across a whole field. Super Skill fills that gap.
Every Super Skill, regardless of domain, contains the same structural layers. The layers are pre-built. The content inside them is domain-specific and user-defined.
This is what makes every Super Skill both universal in structure and unique in content.
LAYER 0: IDENTITY
Who this Super Skill belongs to.
What domain it covers.
What the user is building or managing within that domain.
The operating principles that govern all decisions.
The boundaries of the domain.
LAYER 1: DOMAIN MAP
The full map of the domain and its sub-domains.
The relationships between sub-domains.
Why each sub-domain is included.
What the domain explicitly does not cover and why.
LAYER 2: CURRENT STATE
The actual state of the domain right now.
Versions, tools, methods, entities, positions - depending on type.
When each piece of information was last verified.
The sources used to verify it.
LAYER 3: EVALUATION FRAMEWORK
How to assess anything new entering the domain.
Scoring criteria: compatibility, complexity, lock-in risk, cost.
A log of everything evaluated and the outcome.
A gate requiring user approval before any change is recorded as accepted.
LAYER 4: DECISIONS AND REASONING
Every significant decision made within this domain.
The reasoning behind each decision.
What was considered and rejected, and why.
This layer prevents the model from re-litigating settled questions.
LAYER 5: MONITORING AND DRIFT DETECTION
What sources are watched for changes in the domain.
What constitutes a meaningful change requiring attention.
Where detected changes are written for user review.
Nothing is updated automatically. Everything is proposed first.
LAYER 6: LEARNING GENERATION
Structured summaries ready for NotebookLM ingestion.
Audio overview generation from domain content.
Quiz, flashcard, and mind map generation.
The Super Skill teaches the AI agent and teaches the user.
LAYER 7: PENDING APPROVALS
Nothing changes in a Super Skill without user approval.
All proposed changes, evaluations, and updates live here.
The user reviews and decides.
The model executes only after explicit approval.
LAYER X: EXPERT COUNCIL
Real domain experts selected by the user during activation.
Each expert assigned to a primary layer they contribute most to.
Expert profiles contain methodology, frameworks, red lines,
and domain-specific questions.
Disagreements between experts surface as debates in COUNCIL.md.
Expert positions that conflict with DECISIONS.md go to PENDING.md
for user review.
Claude Code surfaces expert perspectives only when genuine
friction exists, not on routine tasks.
NotebookLM sources per expert load into the domain notebook.
This is the architectural insight that separates Super Skill from every existing approach.
Every domain has sub-domains. Every sub-domain has its own context, cost of change, rate of drift, and relationship to every other sub-domain.
In a technical stack domain, changing the database layer affects authentication, API design, edge behavior, and monthly cost. A model without this map gives advice that is locally correct and systemically wrong.
In an investment portfolio domain, a change in one asset class affects liquidity, tax position, risk exposure, and rebalancing requirements across the whole portfolio. A model that sees each holding in isolation will always miss the full picture.
In a medical documentation domain, a change in regulatory language in one sub-domain affects every template, every workflow, and every compliance checkpoint connected to it.
The layer architecture forces the model to understand these connections before it acts.
When a user defines their domain and activates their Super Skill, they give the model:
- A structured map of how everything in the domain relates
- Which choices have been made at each layer
- What the real cost of changing any piece actually is
This is why a Super Skill produces advice that a Skill cannot.
A Skill knows how. A Super Skill knows why, what else changes if you do, and whether you should.
A Super Skill does not only serve the AI agent. It serves the user.
Every layer of a Super Skill can be converted into structured learning content and pushed into NotebookLM as a source. The user receives:
- Audio overviews of their own domain knowledge
- Quizzes generated from their own decisions log
- Mind maps of their own domain structure
- Flashcards generated from their own current state
This creates a feedback loop no existing system offers.
The Super Skill teaches the AI agent.
The Super Skill also teaches the user.
As the user learns, they make better decisions.
Better decisions improve the Super Skill.
A better Super Skill makes the AI agent more accurate.
A more accurate agent produces better outputs.
Better outputs are recorded in the decisions log.
The decisions log becomes a richer learning source.
The loop continues.
Knowledge does not sit in one place. It circulates.
Clone. Paste one prompt. Answer three questions. Your Super Skill is live.
The full activation sequence, including environment checks, personalization, and expert council setup, is in README.md (Bootstrap prompt) and ONBOARDING.md (Prompt 1). The MANIFESTO describes the framework -- README.md and ONBOARDING.md run it.
The model proposes.
The user decides.
The Super Skill records.
The system executes.
This principle is structural, not optional. It is encoded into Layer 7 of every Super Skill. No change to any layer is accepted without passing through the user approval gate.
This is what makes a Super Skill safe to run continuously. The model monitors, evaluates, detects, and proposes at any frequency. The user controls what actually changes.
A user who activates and maintains a Super Skill will have an AI agent that:
- Knows every element of their defined domain and how those elements relate
- Knows what has been decided, why, and what was rejected
- Knows the current verified state and when it was last confirmed
- Detects drift and surfaces it for user review before it causes problems
- Evaluates new entrants against established criteria
- Generates learning content that keeps the user growing alongside the model
- Never acts on a proposed change without explicit user approval
There is no existing system that produces all of these properties together.
This is the Domain Expert the next generation of AI users needs: Not a model that knows everything about everything. A model that knows everything about your specific domain, your specific context, your specific decisions, and your specific goals.
Super Skill works for any domain where knowledge accumulates and decisions matter: technical, professional, personal, or craft.
See README.md for the full list of domain types with examples.
Every Super Skill contains the same structural layers regardless of domain. The full file list with descriptions is in README.md.
Super Skill v2.7.1 -- April 2026 Authored by Asaf Dahan MIT License - Fork it. Build your own.