Skip to module content
๐Ÿง AI Academy
0/91
Sign in
Module 04 ยท ~7 min

The Skill Ecosystem

The repos worth installing, and hygiene rules for third-party skills.

Reading progress
0/8 ยท 0%

The big idea

๐Ÿ’กKey idea
After the Dec 2025 open-standardization, the skills ecosystem exploded โ€” repos clearing 22k+ stars, an official Anthropic marketplace, hundreds of thousands of community-authored skills. The scarce resource is no longer skills. It's **curation**: knowing which skills are actually production-quality, safe to install, and worth the context budget. The ecosystem organizes into four tiers worth tracking: **Foundational** (Anthropic's official repos and engineering posts), **Directories** (the index repos you check monthly), **Frameworks** (full harnesses with opinionated workflows), and **Marketing-adjacent** (the vertical tools you'll actually use on client work โ€” SEO suites, browser relay, data viz, outbound orchestration).
Quick check
1 question ยท instant feedback
0/1
  1. The #1 supply-chain risk with third-party skills is:

Numbers that matter

22k+
Star repos in the skills ecosystem after Dec 2025 open-standardization.
135 / 35 / 150+
Agents / skills / plugins bundled in awesome-claude-code-toolkit.
4
Hygiene rules non-negotiable for third-party skills: read, fork, benchmark, version-per-client.
Quick check
1 question ยท instant feedback
0/1
  1. Why fork instead of hotlink?

Deep dive

4/4 open

**`anthropics/skills`** is the canonical reference implementation. It ships skills for reading docx, PDF, PowerPoint, and Excel files; generating canvas designs; creating algorithmic art; building MCP servers (`mcp-builder`); and evaluating other skills (`skill-creator`). This is where you learn the house style.

**`claude-code-action`** is the official GitHub Actions integration โ€” the mechanism for running Claude Code headlessly in CI pipelines.

The **Official Plugin Directory** (in-product) is where installable plugins appear as Anthropic certifies them.

**Anthropic's engineering posts** โ€” especially the multi-agent research post โ€” are the canonical design docs for orchestrator/subagent architecture. Read them before building anything complex.

**`hesreallyhim/awesome-claude-code`** is the de-facto ecosystem index. Check it monthly. It catalogs skills, hooks, slash commands, frameworks, and plugins across the entire community.

**`ComposioHQ/awesome-claude-skills`** has the clearest explanation of the skill format and cross-platform install instructions (`npx skills add โ€ฆ`). Start here if you're onboarding someone new.

**`travisvn/awesome-claude-skills`** includes skill templates and a sharp framing of when to use Skills vs Subagents vs Projects โ€” the three mechanisms people most often confuse.

**`VoltAgent/awesome-claude-code-subagents`** catalogs 100+ production-oriented role agents. Mine this for your Strategist, Analyst, and Reviewer subagent definitions.

**`obra/superpowers`** (~94k stars) is the 7-phase TDD framework referenced throughout this Field Manual. It's now in the official Anthropic marketplace. Cherry-pick the phase-gate pattern; don't wholesale-install.

**'Everything Claude Code'** (~100k-star class) is a full harness reference. Same rule: cherry-pick, don't wholesale.

**Garry Tan's public setup** is the canonical role-based orchestration template โ€” six subagents (CEO, Designer, EM, Release, Doc, QA) translated to any knowledge-work domain.

**`awesome-claude-code-toolkit`** bundles 135 agents, 35 skills, and 150+ plugins with SkillKit access โ€” the densest starting point for a GMS OS package.

**`agents-best-practices`** (also called DenisSergeevitch) is the single best resource on production harness engineering: risk taxonomy, permission matrices, security evals. Its language maps directly to client SOWs and ยง18 risk management.

Third-party skills are code. Treat them that way.

**Read every SKILL.md and every bundled script before installing.** Skills can contain prompt-injection attempts or supply-chain risks, especially any skill that requests write-capable tools.

**Fork, don't hotlink.** If you reference a skill directly from someone else's GitHub repo, an upstream change can silently alter your agent's behavior mid-engagement. Vendor into your own repo and pin the version.

**Benchmark before adopting.** Use `skill-creator`'s eval harness to compare with-skill vs no-skill on real tasks. If the skill doesn't measurably beat the baseline, it's context pollution โ€” its always-visible description costs tokens for nothing.

**Version per client.** Client-specific skills live in the client repo. General-purpose skills live globally. Never let client A's conventions bleed into client B's sessions.

Quick check
1 question ยท instant feedback
0/1
  1. Which repo is the de-facto index of the ecosystem?

How to run it

  1. Fork before use
    Vendor into your own repo; pin the version.
  2. Read scripts
    Every bundled script gets a code review before install.
  3. Benchmark vs baseline
    Use skill-creator's eval harness โ€” must measurably win.
  4. Per-client vs global split
    Never let client-A skills leak into client-B sessions.
Quick check
1 question ยท instant feedback
0/1
  1. The DenisSergeevitch/agents-best-practices repo is uniquely valuable for:

In the field

๐Ÿ”ฌWorked example
Skill_Seekers converts a documentation website into a skill โ€” point it at n8n docs, Clay docs, or the Meta API docs and your agent instantly 'knows' the product. Bedrock (Obsidian second brain) is 8 skills building an entity-typed knowledge vault โ€” the pattern for a GMS internal knowledge OS.
๐ŸšซWhen not to reach for it
Never install a third-party skill blindly. Skills are prompt-injection and supply-chain surface. Anything with bundled scripts requires code review; anything requesting write-capable tools gets extra scrutiny.
Quick check
1 question ยท instant feedback
0/1
  1. A skill that fails to beat the no-skill baseline in skill-creator's harness is:

Pitfalls & takeaways

Failure modes

  • Hotlinking upstream instead of forking โ€” upstream changes silently alter agent behavior mid-engagement.
  • Skipping benchmarks โ€” an unmeasured skill can be pure context pollution.
  • Letting client-A skills leak into client-B sessions.

Durable takeaways

  • Curation is the scarce resource โ€” start with the directories monthly.
  • Fork every adopted skill; upstream must never silently change client behavior.
  • Benchmark or don't install โ€” context pollution is a real cost.
  • The agents-best-practices risk taxonomy is your SOW language.

Do the work

๐Ÿ‹๏ธProve you learned it

Adopt one marketing-adjacent skill (e.g., Ay-Skills SEO audit). Fork it into your own repo, read every SKILL.md + script, run skill-creator's eval harness against a baseline, and write a one-paragraph adoption note (what it does, what tools it can write to, benchmark delta).

0 chars
๐Ÿ“ฆArtifact to produce
Vendored skills repo with per-client versioning + a written adoption/benchmark log.

Sources

  • ยท GMS Field Manual ยง4 (Skill Ecosystem)