TAU-HOME.COM
LOADING

Agentic Awesome Skills (AAS Core): Control Plane and Catalog of 2,950+ AI Agent Skills

An overview of Agentic Awesome Skills (AAS Core), an open-source library containing over 2,957 reusable SKILL.md playbooks for Claude Code, Codex, Cursor, and G

tau · November 16, 2026

#agentic-skills #claude-code #codex-cli #cursor #developer-tools

Agentic Awesome Skills (AAS Core): Control Plane and Catalog of 2,950+ AI Agent Skills

As AI coding assistants transition from simple prompt responders to multi-step reasoning agents, managing standardized rules and domain-specific knowledge playbooks has become a crucial workflow challenge. The open-source project Agentic Awesome Skills (formerly known as Antigravity Awesome Skills) addresses this need. As of release V20.8.0, tagged on November 12, 2026, the project offers a catalog of over 2,957 installable SKILL.md playbooks alongside AAS Core, a local control plane for catalog discovery and stack planning.

Agentic Awesome Skills logo and agent skills catalog architecture diagram

Image source: https://github.com/sickn33/agentic-awesome-skills

Rather than requiring developers to manually paste prompt templates across different tool directories, AAS Core enables connected AI agents to inspect a locally cached catalog, record chosen instructions, and preview an execution plan before altering configuration files.

Background and Role: Organizing Complex Prompt Playbooks

In agentic coding workflows, a skill represents a modular, standardized instruction set written in Markdown—defining library integration guidelines, refactoring rules, accessibility standards, or visual design checks. As workflows expand, maintaining dozens of separate prompt files inside system instructions creates bloat, driving up model token costs and degrading prompt adherence.

AAS Core provides a lightweight, agent-first architecture to solve this coordination problem. It lets agents like Claude Code or Codex query the complete local catalog and assemble a precise subset of skills tailored to a specific task.

Critically, Core acts as a neutral coordinator rather than an automated decision-maker. It does not rank, score, or recommend individual candidates. The choice of which skills to pull into a workspace remains entirely driven by the agent's contextual assessment and the human developer's explicit confirmation.

Three-Stage Inspection Pipeline and Browser Workbench

To prevent unverified configuration changes, AAS Core implements a strict read-only inspection workflow before any local files are updated:

  1. In-Memory Assembly (compose_stack): The agent validates selected skill identifiers and syntax in memory. This step operates entirely as a read-only process without touching local filesystem files. A single stack manifest (aas-stack.json) supports up to 128 skills.
  2. Manifest Verification (aas stack validate): A command-line utility checks the integrity of aas-stack.json to ensure that skill identities are valid and structured properly.
  3. Immutable Plan Preview (aas stack plan): Core writes an immutable preview artifact summarizing exactly which files and skills will be introduced or modified, allowing developers to audit changes in detail.

Additionally, the project provides a browser-based companion tool called Workbench. It lets developers inspect stack manifests and plan outputs entirely in client-side browser memory without granting web services access to local project filesystems.

Supported Target Environments and Direct GitHub CLI Installation

The AAS repository supports a wide variety of agent targets rather than locking users into a single platform. Supported systems include Anthropic's Claude Code, OpenAI's Codex CLI, Cursor AI IDE, Google's Gemini CLI, Antigravity, GitHub Copilot, Kiro CLI, and Autohand Code.

Due to a temporary maintainer account lock, publishing updates to the public npm registry is currently paused. As a result, the project officially routes current installations and updates through GitHub directly, utilizing the GitHub CLI (gh skill extension) rather than standard npm package managers.

For instance, installing a pinned skill for Codex at the user scope can be performed via the terminal:

gh skill preview sickn33/agentic-awesome-skills skills/brainstorming/[email protected]
gh skill install sickn33/agentic-awesome-skills skills/brainstorming/SKILL.md --agent codex --scope user --pin v20.8.0

For other target environments, developers specify the appropriate --agent flag (such as claude-code, cursor, or gemini-cli) or supply a custom target folder using --dir.

Trust Boundaries and Verified Caveats

Adopting AAS Core requires a clear understanding of its operational boundaries:

Structural validity does not imply functional or semantic correctness. While aas stack validate verifies that a skill ID exists and conforms to manifest schemas, it does not guarantee that the skill instructions match your project's codebase, that environmental configurations are error-free, or that executing the skill carries no unintended side effects.

Furthermore, automated workspace application (apply) and rollback recovery mechanisms remain experimental opt-in features. Because they sit outside the stable preview boundary, manual inspection of the generated plan preview remains essential before modifying production configurations.

Regarding intellectual property, original AAS code and command-line tools are released under the MIT License, with original written content licensed under CC BY 4.0. However, imported third-party skills retain their original author licenses and attribution requirements, making individual license reviews necessary when integrating specific skills into commercial enterprise codebases.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285