Agentic Awesome Skills (AAS)

sickn33‘s catalog of 2,003 SKILL.md playbooks for coding agents — MIT (with a separate LICENSE-CONTENT for the skill payloads), ~44.5k★ / 6.5k forks / 2,464 commits, release v15.9.0, installable with npx agentic-awesome-skills and targeting Claude Code, Cursor, Codex CLI, Gemini CLI, Antigravity, Kiro, OpenCode and Copilot. Independent; the README states plainly it is not affiliated with Google.

The interesting thing is not the catalog. It is that the catalog got big enough to break, and the project’s response was to build a control plane on top of it.

From a list to a selection machine

v15.9.0 ships “AAS Core”: a local, read-only MCP server exposing search_skills, get_skill, compose_stack, inspect_stack, diff_stack and evidence export. The documented flow is that Codex or Claude inspects your project and picks the skill IDs — Core does not rank, recommend, exclude or hide anything. The selection is persisted as aas-stack.json with a aas-selection-evidence.json sidecar, validated by aas stack validate, previewed as an immutable per-target plan by aas stack plan, and reviewed by a human in a browser-local “Workbench” before anything is written.

That is agentic-resource-discovery with the retrieval problem admitted rather than assumed. The skills model works because a skill’s frontmatter is cheap to scan and its body loads on demand — but that scan is still linear in the number of skills, and two thousand of them does not fit anywhere useful. So the project inverted the arrangement: the agent queries a catalog instead of a directory being handed to it, and selection becomes the product.

The scale failure, stated by the project itself

The README documents its own breakage, which is worth more than the catalog:

Antigravity watches ~/.agents/skills and may load enough installed instructions to exhaust its context, slow startup, trigger truncation errors, or enter a crash loop.

For that target the installer therefore refuses to run bare — you must pass --skills, a metadata filter, or an explicit --all. A manifest is capped at 128 skills for technical reasons.

This is the clearest evidence the spoke holds that progressive disclosure has a ceiling, and that the ceiling is reached by the index, not by the bodies. It belongs beside the token-budget corner (headroom, safe-prompt-pruning, pxpipe, md-this-page, pullmd), but it is a different lever from all five: those reduce what a loop carries, this reduces what a loop can even be offered. And it is the counter-example to the “more skills is more capability” reading that the star count invites.

What it explicitly does not claim

Unusually for a project at this star count, the disclaimers are specific and repeated:

• Structural validity is not semantic fit. Core validates IDs and structure — “does not certify semantic fit, compatibility, setup correctness, operational safety, or safety to apply.”
• Apply and recovery are experimental, outside the supported preview path.
• Metadata is informational, not an eligibility gate.
• A full install “includes critical and authorized-use-only offensive instructions”; installing copies files without executing them, “but an agent may act on an installed skill later.”

That last one is the agent-guardrails problem in a supply-chain shape, and it is the reason this page is cross-linked to the defensive-security reading of ../defensive-security-wiki: a catalog that installs two thousand instruction files into an autonomous agent’s working set is a distribution channel, and the trust boundary is the selection, which by the project’s own statement nobody certifies.

Tier and what to distrust

T3 — a first-party repository README. Everything above is the project describing itself. Three things are not established by it and should not be inferred:

• Quality of the 2,003 skills. Nothing is tested, benchmarked or ranked; the project says so. The credits list a long tail of community and “official” sources, so an unknown share is aggregated from elsewhere rather than authored.
• What the star count means. ~44.5k★ against 2,464 commits and a v15.9.0 version number is a velocity that measures attention, not use — openclaw is the spoke’s standing reminder, and jetbrains-ponytail-tested is what happens when someone measures instead.
• Whether agent-owned selection works. The entire premise is that an agent picks better than a ranking would. No evaluation of that is offered.

Related

sickn33 · agentskills-spec · agentic-resource-discovery · agent-skills · model-context-protocol · agent-guardrails · claude-code · superpowers · awesome-agent-evolution · openclaw · jetbrains-ponytail-tested · synthesis