The definitive guide to production-grade agent skills — the reusable capability layer that turns AI coding agents into reliable engineering teammates.
This comparison covers 10 tools spanning the full production agent skills lifecycle: skill frameworks (vercel-labs/skills, obra/superpowers), production skill packages (addyosmani/agent-skills), skill routers (reverse-skill), skill recording (microsoft/skill-recorder), skill generation (book-to-skill), and quality layers (ratchet, ponytail-improved, i-have-adhd). Combined, the category exceeds 404,000 GitHub stars.
| # ▲ | Tool ↕ | ⭐ Stars ↕ | Growth ↕ | Type ↕ | Score ↕ | Best For |
|---|---|---|---|---|---|---|
| 1 | addyosmani/agent-skills | 81,976 | +226/day | Production Skill Package | 96 | Category Defining |
| 2 | obra/superpowers | 266,458 | — | Skills Framework & Methodology | 94 | Framework Leader |
| 3 | zhaoxuya520/reverse-skill | 19,064 | +9,904/wk | Skill Router Pack | 85 | RE & Security |
| 4 | virgiliojr94/book-to-skill | 16,990 | +4,596/wk | Skill Generation | 84 | Books → Skills |
| 5 | ayghri/i-have-adhd | 17,278 | +3,874/wk | Output Quality Layer | 83 | Agent UX |
| 6 | microsoft/skill-recorder | 969 | — | Skill Recording Platform | 78 | Enterprise Capture |
| 7 | vercel-labs/skills | — | — | Skills CLI Framework | 74 | npx skills add |
| 8 | ponytail-improved | 580 | — | Agent Prompt Optimization | 62 | YAGNI Discipline |
| 9 | ratchet | 412 | — | Agent Compliance Checking | 60 | Policy Compliance |
| 10 | codex-as-mcp | 169 | — | MCP Bridge | 55 | Codex + MCP |
Click column headers to sort. Use filter buttons above to narrow by tool type. Scores are composite: 40% production readiness, 25% category impact, 20% community adoption, 15% ecosystem influence.
The reference implementation of production-grade agent skills. Where other repos ship experimental skills, agent-skills ships a complete engineering SDLC: start with /spec to write requirements, /plan to architect, /build to implement, /test to verify, /review to critique, /webperf to optimize performance, /code-simplify to reduce complexity, and /ship to release. Each skill is written with the same care as Chrome DevTools documentation — precise, actionable, and deterministic.
The project that invented the agent skills paradigm in its modern form. Superpowers treats software development as a subagent-driven skill system — every phase of the SDLC becomes a composable, reusable skill managed by specialized subagents. With 266K+ stars it remains the biggest repo in the ecosystem, and its "skills + subagents + methodology" model is the blueprint that addyosmani/agent-skills productized for production teams.
The fastest-growing security skill pack in the ecosystem, growing nearly 10K stars a week. Reverse-skill is an AI-powered skill router for reverse engineering, penetration testing, and security research — it routes tasks to the right toolchain, bootstraps tools on demand, and maintains a self-evolving knowledge base. Supports Claude Code, Kiro, Cursor, Cline, and more.
The most creative skill-generation tool in the ecosystem: it turns entire technical books into executable Claude Code skills. Feed it a technical book, and it extracts the methodologies, patterns, and workflows into structured skills the agent can actually apply. Growing at nearly 4,600 stars a week, it proves that skill authoring doesn't have to start from a blank page — the world's technical literature is a skill library waiting to be compiled.
A skill that reworks how agents present information for ADHD-friendly consumption — short, scannable, action-oriented output instead of walls of text. It's the clearest evidence yet that the skills paradigm covers human factors, not just engineering. Nearly 17.3K stars and growing ~3,900/week shows how underserved the "agent UX accessibility" layer was before skills made it codifiable.
Microsoft's enterprise answer to skill authoring: record your on-screen workflow, and the GitHub Copilot CLI reconstructs it into a reusable Skill or Automation for Microsoft Scout, Copilot Cowork, or Copilot Studio. This is the "teach once, automate forever" pattern — no hand-writing skill files. Lower star count than the consumer hits, but the highest enterprise credibility in the category.
Vercel's skills CLI is the package manager moment for agent skills: npx skills add installs a skill into your agent's skills directory from a registry. If addyosmani/agent-skills is the content and obra/superpowers is the methodology, vercel-labs/skills is the plumbing — a standardized install/update/remove interface that makes skills as easy to distribute as npm packages. This is the layer that turns a grassroots movement into a platform.
An enhancement layer for the ponytail skill system: optimized skill injection, sharper reasoning chains, and stricter YAGNI enforcement. Where the 95K⭐ parent project defines the "lazy senior dev" philosophy, ponytail-improved is the tuning fork — better prompts, better tool use, lower token consumption. A small but well-loved quality layer in the stack.
The compliance guardrail for the skills era. Ratchet verifies that agents follow organizational policy — coding standards, security rules, approved toolchains — before actions are taken or code is shipped. As skills make agents more capable, ratchet makes them more trustworthy: a policy-checking layer that enterprises will bolt onto any skills framework.
A small but strategically important bridge: it exposes OpenAI Codex as a Model Context Protocol server, letting MCP-based tools and skill systems drive Codex directly. As the skills ecosystem standardizes on MCP for tool connectivity, bridges like this determine which agents can consume which skills. 169 stars today; outsized ecosystem influence per star.
The agent skills ecosystem is exploding for five structural reasons, and Run #53's data makes each one measurable:
81,976⭐ in weeks
Until now, the skills movement was led by indie hackers and framework pioneers. A Chrome DevTools engineering leader shipping addyosmani/agent-skills changed the audience: enterprise engineers who trust browser-tooling discipline now have a reference implementation they can defend in architecture reviews. The "production-grade" label went from marketing to reality.
Deterministic • Reviewable • Composables
A skill is version-controlled, diffable, tested, and composable — everything a prompt is not. Teams can put skills through code review, measure their effect, and roll them back. That's the difference between "agent vibes" and "agent engineering," and it's why enterprises are standardizing on skills.
npx skills add
vercel-labs/skills gave the ecosystem its npm moment: a one-command installer backed by a registry. Combined with Claude Code's native skills directory and Cursor's rules support, installing a skill pack is now a 10-second operation. Distribution bottlenecks were the last thing holding the category back — they're gone.
Record it • Compile it • Ship it
Three parallel authoring revolutions: microsoft/skill-recorder (record your screen, get a skill), book-to-skill (compile a book into skills), and the sheer volume of open skill packs (reverse-skill, i-have-adhd). When creating a skill is easier than writing a blog post about it, the ecosystem compounds.
ratchet • ponytail-improved • i-have-adhd
Quality and compliance layers matured in parallel: ratchet for policy compliance, ponytail-improved for output discipline, i-have-adhd for human-readable output. Enterprises won't deploy capable agents without guardrails — the fact that the guardrails now exist as skills themselves is the ecosystem maturing in real time.
In Q1 2026 the "agent skills" category barely existed. By August 2026 it exceeds 404K stars across just the 10 tools on this page, with three repos growing at 3,800–9,900 stars per week. The comparison to npm (2015), Kubernetes (2016), and LangChain (2023) is apt — this is category-creation velocity, and the window to adopt (and build on) skills is now.
The ecosystem isn't a grab bag of repos — it's a coherent four-layer stack. Understanding the layers explains why the pieces fit together and where the gaps (and opportunities) are.
The foundation: how skills are defined, packaged, installed, and executed.
The content: production skill collections that encode real expertise.
The intelligence: deciding which skill applies to which task, and loading the right toolchain.
The flywheel: creating new skills from demonstrations, documents, and books without hand-authoring.
| Use Case | Best Tool | Why |
|---|---|---|
| Production engineering teams | addyosmani/agent-skills | 24 battle-tested skills + full /spec→/ship lifecycle from Chrome DevTools credibility |
| Full SDLC methodology | obra/superpowers | 266K⭐ subagent-driven framework — the deepest methodology in the category |
| Security & reverse engineering | zhaoxuya520/reverse-skill | 19K⭐ AI-routed skill pack with self-evolving knowledge base, +9.9K stars/week |
| Training agents from existing knowledge | virgiliojr94/book-to-skill | Compile books and docs into Claude Code skills at +4.6K stars/week |
| Enterprise skill capture | microsoft/skill-recorder | Record workflows → publish skills for Copilot ecosystems, Microsoft-backed |
| Agent output readability | ayghri/i-have-adhd | 17.3K⭐ proof that scannable, action-first agent output is a feature, not a nicety |
| Distributing skills at scale | vercel-labs/skills | npx skills add — the package-manager moment for the whole ecosystem |
| Governance & compliance | ratchet | Policy checking and pre-action gating for regulated environments |
| YAGNI discipline | ponytail-improved | Optimized prompt injection and stricter anti-over-engineering guardrails |
| Cross-runtime interop | codex-as-mcp | Expose Codex to MCP-native skill routers and clients |
Start with addyosmani/agent-skills
For most teams in August 2026, the fastest path to production-grade agent skills is: install addyosmani/agent-skills as your baseline skill package, layer vercel-labs/skills (or your agent's native skills dir) as the installer, add ratchet or equivalent compliance gating if you're regulated, and study obra/superpowers for the methodology. Then start feeding your own knowledge in via book-to-skill-style pipelines. The stack is ready — and unlike most AI hype, it's versionable, reviewable, and shippable.
Skill frameworks, routers, and recording pipelines need infrastructure. Self-host a skill router, run a skills registry, or deploy agent endpoints:
Skills are authored and consumed inside AI-native IDEs. These are the best platforms for running the skills stack:
Run the skills stack at scale — skill verification pipelines, compliance checking (ratchet), and enterprise recording:
💡 Cursor recommendation: Ai-first code editor with agentic autocomplete. Try Cursor →
Affiliate disclosure: we may earn a commission if you sign up via this link, at no extra cost to you.
💡 Cloudflare recommendation: Cdn, dns, ddos protection, and edge compute for any site. Try Cloudflare →
Affiliate disclosure: we may earn a commission if you sign up via this link, at no extra cost to you.