Ecosystem picks
What is worth taking from the big catalogs, verified 2026-09-20.
Star counts in this space are noisy — several repos with six-figure counts have thin commit histories. Ordered by usefulness to this stack, with the count as context rather than as ranking.
Installed here
| Source | What you got | Why |
|---|---|---|
| obra/superpowers | 15 skills: test-driven-development, systematic-debugging, brainstorming, writing-plans, verification-before-completion, using-git-worktrees, dispatching-parallel-agents, subagent-driven-development, receiving-code-review | The most complete methodology set in the ecosystem. verification-before-completion alone — "evidence before assertions, always" — is worth the install |
| addyosmani/agent-skills | code-review-and-quality, debugging-and-error-recovery, security-and-hardening, performance-optimization, observability-and-instrumentation, incremental-implementation, deprecation-and-migration, spec-driven-development, frontend-ui-engineering, ci-cd-and-automation, git-workflow-and-versioning, doubt-driven-development, source-driven-development, context-engineering | Production engineering, broad and consistent. The closest thing to a senior-engineer checklist set |
| anthropics/skills | docx, pptx, xlsx, pdf, mcp-builder, skill-creator, canvas-design, webapp-testing, frontend-design, artifacts-builder | Official. mcp-builder and skill-creator are the meta-tools — they build the things this kit is made of |
| mattpocock/skills | tdd, pr, research, prototype, wizard, codebase-design, domain-modeling, diagnosing-bugs, resolving-merge-conflicts, setup-pre-commit, migrate-to-shoehorn | TypeScript-shaped and very active. wizard is unusual and good: it generates an interactive bash wizard for steps only a human can do |
| RobMitt/grill-me-skill | grill-me, grilling | Relentless interview until shared understanding. grill-with-docs leaves ADRs and a glossary behind |
| elayadesign/ai-design-skills | landing-page-design | Layout, conversion copy, and strict visual rules |
| dmmulroy/anti-slop | install-anti-slop | Oxlint rules that reject low-evidence code — the linter equivalent of "prove it" |
| s1dashu/ip-as-logo-skill | ip-as-logo | Character/logo generation |
| LukasNiessen/kubernetes-skill | kubernetes-skill | Prevents K8s hallucinations: insecure defaults, resource starvation, privilege sprawl, API drift. EKS/GKE/AKS/OpenShift |
| firecrawl/anydoc | convert-documents-to-markdown | Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, PDF → Markdown |
npx -y skills@latest list # what is installed, and from where
npx -y skills@latest update -y # or /skills-sync
Worth adding next
| Source | Pick | Why |
|---|---|---|
| wshobson/agents ★40k | Subagent marketplace | The gap in this kit: 3 subagents against dozens there |
| BagelHole/DevOps-Security-Agent-Skills | 10–15 of 146 | AWS/Azure/GCP, Terraform, Helm, ArgoCD GitOps, CIS benchmarks, incident response, compliance. Install selectively — 146 descriptions is a real context tax |
| a vendored third-party toolkit | ~58 not yet taken | speckit-* (12 skills), openspec-lifecycle, compass-* bootstrap tooling, agent-teacher, project-memory, long-running-task, work-splitting, architecture-advisor, gen-docs, figma-mcp-frontend |
| nizos/tdd-guard ★2.3k | Hook-enforced TDD | Blocks implementation until a failing test exists — tdd as a hook, not a suggestion |
MCP servers
From punkpeye/awesome-mcp-servers ★95k:
| Server | Why |
|---|---|
| github/github-mcp-server | Official. Issues, PRs, Actions, code search |
| microsoft/playwright-mcp | Real browser. The agent runs the test it wrote |
| headroomlabs-ai/headroom ★73k | Compresses tool output before it enters context — attacks the >150k problem directly |
| DeusData/codebase-memory-mcp ★44k | Structural code index |
| GLips/Figma-Context-MCP | Figma layout as data, not screenshots |
| grafana/mcp-grafana | Already in use — see my servers |
Catalogs themselves
| List | Use it for |
|---|---|
| hesreallyhim/awesome-claude-code ★54k | 21 categories. The best-curated overall |
| ComposioHQ/awesome-claude-skills ★75k | Skills specifically |
| VoltAgent/awesome-agent-skills ★35k | 1000+, breadth over curation |
| sickn33/agentic-awesome-skills ★47k | Local agent-first control plane |
| github/awesome-copilot ★37k | Instructions, agents, skills, configs for Copilot |
| agentation.com | Agent tooling directory, broader than Claude |
Skip
| Repo | Why |
|---|---|
| multica-ai/andrej-karpathy-skills ★199k | It is a single CLAUDE.md, no commits since April. Read it for ideas; there is nothing to install |
| Anything with six figures and three commits | The counts in this ecosystem do not survive scrutiny. Open the commit history before trusting one |
How to evaluate before installing
A skill is a prompt running with your tool permissions. Read SKILL.md first —
same reflex as reading a shell script before | bash.
allowed-tools— does it ask for more than its job needs?- Network or out-of-project writes — why?
- Commit history — real iteration, or one bulk drop?
- Description quality — a vague description means it fires at random, or never