// the take
A library of 2,743+ markdown 'SKILL.md' playbooks for AI coding agents (Claude Code, Codex, Cursor, Gemini CLI, Kiro, etc.), plus a CLI and local MCP server ('AAS Core') that lets an agent search the catalog, record a chosen set of skills, and preview an install plan before touching a project. It's for developers who want their coding agent to pull in task-specific instructions — debugging, SEO audits, browser automation, document conversion — instead of hand-rolling prompts every time.
The install flow has a real trust boundary: selection (compose_stack, read-only) → validation (aas stack validate) → an immutable plan preview (aas stack plan) → explicit opt-in apply, which is more careful than most 'curl this into your agent' skill installers. Each host gets its own install path and docs (Claude Code, Codex, Cursor, Gemini CLI, Antigravity, Kiro, Copilot...) instead of one generic script, and there's a machine-readable skills_index.json with a published schema for hosts that want to consume the catalog directly rather than through the CLI. The 128-skill manifest cap and the explicit disclaimer that structural validation isn't semantic or safety validation show someone actually thought about an agent overloading its own context or trusting a plan blindly.
A large share of the 'skills' are thin wrappers around paid third-party hosted APIs — outbound robocalling, crypto trading bots, NSFW image generation, LinkedIn/X scraping and outreach — rather than reusable instructions; installing one effectively wires your agent into another company's billed service and ToS, which isn't obvious from a headline count of '2,743 skills.' A notable chunk of those paid-service skills trace back to a handful of repeat publishers (one account alone behind roughly 20 near-identical single-skill repos), worth knowing before treating skill count as a quality signal. Apply and recovery — the part that actually changes your project — is explicitly marked experimental; the polished, documented part of the tool is selection and preview, not the part that does anything. And fundamentally this is an agent reading and then acting on 2,700+ community-contributed markdown files as instructions; review is CI lint plus maintainer audit docs, not a security review, which is a real prompt-injection/supply-chain surface for something loaded straight into a coding agent's context.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →