TL;DR: "Agent Skills" became the biggest idea in AI coding in mid-2026. Superpowers (obra/superpowers) rocketed past 282,000 GitHub stars to become the most-starred agent framework, and Anthropic's official Agent Skills repo crossed 175,000 stars. Skills are not a replacement for MCP — they solve a different problem: how the agent should behave, not what it can connect to. The winning stack in 2026 is skills + MCP + a methodology like Superpowers, and it takes about 5 minutes to install.
Table of Contents
1. What Are Agent Skills?
An Agent Skill is a self-contained, reusable package of instructions and resources that you drop into an agent's project — usually a SKILL.md file plus optional scripts, templates, and examples. When the agent encounters a matching task, it reads the skill and follows its procedure.
Think of it as the difference between hiring a smart generalist and hiring a specialist with a playbook:
- Prompting = telling the agent what to do in one conversation.
- Agent Skills = giving the agent a durable, versionable playbook it can follow every time.
- MCP servers = giving the agent new tools/APIs to call (search, databases, browsers).
Skills exploded in 2026 because they fix the #1 complaint about AI coding agents: inconsistency. With a skill, the same high-quality procedure (tests first, security review, specific output format) runs every single time, on every project, across Claude Code, Cursor, Codex, Gemini CLI, and a dozen other harnesses.
2. Superpowers: The 282k-Star Methodology
obra/superpowers describes itself as "an agentic skills framework & software development methodology that works." Created in October 2025, it passed 282,000 stars by September 2026 — making it one of the fastest-growing repositories on GitHub, period.
Update (Aug 9, 2026): the chasing pack is closing in fast. Verified via the GitHub API today — mattpocock/skills ("Skills for real engineers. Straight from my .agents directory.") has crossed ~210,000 stars, and Addy Osmani's agent-skills sits at ~84,500. The ecosystem is no longer a two-horse race — the SKILL.md convention is becoming a genuine de-facto standard across the whole coding-agent industry.
Update (Sep 7, 2026): the race has widened into a fleet. Re-verified via the GitHub API today: mattpocock/skills ~254,500★, ECC (agent-harness performance work) ~251,300★, NousResearch/hermes-agent ("the agent that grows with you") ~242,500★, and the open coding agent opencode ~205,000★. Fastest riser of the month is tt-a1i/archify, an agent skill that turns ideas into architecture and flow diagrams: it went from ~17.9k★ (Aug 27) to ~50.8k★ today, a near-tripling in ten days. The skills wave is not slowing down — it is broadening into harnesses, diagrams, and research workflows.
What it actually does
Superpowers is not a single tool. It's a complete software development methodology built on top of composable skills. The core workflow:
- Brainstorming — Before writing code, the agent asks what you're really trying to do, explores alternatives, and presents a design document in digestible chunks for your sign-off.
- Git worktrees — After design approval, it creates an isolated workspace on a new branch, runs project setup, and verifies a clean test baseline.
- Subagent-driven development — It launches subagents to work through each engineering task, inspecting and reviewing their work, with true red/green TDD, YAGNI, and DRY enforced. It's not uncommon for the agent to work autonomously for a couple of hours without deviating from the approved plan.
The skills trigger automatically from a bootstrap instruction — you don't have to remember to activate them. That "just works" property is a big reason for the explosive adoption.
Works across 11+ harnesses
One of Superpowers' killer features is harness neutrality. The same skills package installs into:
| Harness | Install command |
|---|---|
| Claude Code | /plugin install superpowers@claude-plugins-official |
| Cursor | /add-plugin superpowers |
| Codex App / CLI | /plugins → search "superpowers" |
| Gemini CLI | gemini extensions install https://github.com/obra/superpowers |
| GitHub Copilot CLI | copilot plugin install superpowers@superpowers-marketplace |
| Kimi Code / Antigravity / OpenCode / Pi / Factory Droid | Marketplace or direct install |
Superpowers is even listed in Anthropic's official Claude plugin marketplace and OpenAI's official Codex plugin marketplace — a rare cross-ecosystem endorsement. (For a Chinese-language walkthrough of the framework internals, see our companion post on the main site: Superpowers Agent 框架入门.)
3. Claude Agent Skills: The Official Spec
Anthropic formalized the concept with its anthropics/skills repository — now at ~175,000 stars — which hosts official, reusable Agent Skills for Claude Code. These range from PDF/document processing to web research and code review playbooks.
The key differences from the Superpowers approach:
- Scope: Official skills are task-oriented playbooks (e.g., "analyze a PDF", "review a PR"), while Superpowers is a full development methodology that orchestrates many skills in sequence.
- Ownership: Anthropic maintains the official ones; the broader community maintains thousands of others (Superpowers alone has 700+ skill variants).
- Format: Both use the same emerging convention — a
SKILL.mdmarkdown file with frontmatter (name, description) + supporting files. This is converging into a de-facto standard.
Heads-up: The skill format is still young. If you write your own skills, pin the SKILL.md schema version you target, because both Anthropic and the community are still iterating on it (name/description frontmatter, dependency files, and execution hooks have all changed within the past year).
The community wave (Sep 2026)
Beyond the headline numbers, a second tier of skill repositories is compounding fast — all verified via the GitHub API on Sep 7, 2026: Anthropic's official claude-plugins-official marketplace repo (~36k★), scientific-agent-skills (~43k★) for research workflows, humanizer (~44k★) for stripping the "AI smell" out of generated text, and diagram-design (~32k★) for editor-grade diagrams. ruflo (~71k★) and ponytail (~129k★, an agent code-slimming skill) rounded out the trending wave. The takeaway: if you build on the SKILL.md convention in 2026, you are building on a de-facto standard — and the ecosystem is now deep enough that almost every workflow has a maintained, heavily-starred skill for it.
4. Superpowers vs Claude Agent Skills vs MCP
People keep asking "which one should I use?" — but they solve different layers. Here's the honest comparison:
| Layer | Superpowers | Claude Agent Skills | MCP |
|---|---|---|---|
| What it is | Development methodology + skill framework | Official skill package/spec for Claude | Protocol for connecting tools & data sources |
| Answers | "How should the agent work?" | "What playbook should it follow?" | "What can it reach?" |
| Install | Plugin/marketplace, per harness | /plugin install or add to project | Add server config (npx, docker, stdio) |
| Vendor | Community (obra) — neutral | Anthropic official | Open standard, many vendors |
| Stars (Sep 7, 2026) | Superpowers 282k ⭐ · mattpocock/skills 254k ⭐ · ECC 251k ⭐ | Anthropic skills 175k ⭐ · addyosmani/agent-skills 92.5k ⭐ | modelcontextprotocol ~40k+ ⭐ |
| Best for | Teams wanting a disciplined, repeatable dev process | Claude users wanting official playbooks | Giving agents real tools (DB, browser, APIs) |
They compose, they don't compete
The practical insight from 2026's best setups: run all three layers together. Superpowers provides the workflow discipline, official/community skills provide the playbooks, and MCP servers provide the tools those playbooks call. A Superpowers-driven agent that can query your database over MCP and follow an official PDF-analysis skill is dramatically more capable than any single layer alone.
5. How to Build Your Stack (Practical Guide)
Step 1: Pick your harness
If you live in one editor, install Superpowers there. If you use multiple (Claude Code + Cursor, for example), install per harness — the skills are cheap and the bootstrap instruction is small.
Step 2: Install Superpowers (5 minutes)
# Claude Code (official marketplace)
/plugin install superpowers@claude-plugins-official
# Cursor
/add-plugin superpowers
# Codex CLI
/plugins → search "superpowers" → Install Plugin
# Gemini CLI
gemini extensions install https://github.com/obra/superpowersStep 3: Add official skills for your workflow
# In Claude Code, browse Anthropic's official skills:
/plugin marketplace add anthropics/skills
/plugin install <skill-name>@anthropics-skillsStep 4: Wire up MCP for real tools
Add MCP servers for the capabilities your projects actually use — GitHub, Postgres, Playwright for browser automation, or a vector store for RAG. Keep it minimal: every MCP server adds context overhead and attack surface.
Step 5: Verify the loop
Run one real task and watch for the three Superpowers phases (brainstorm → worktree → subagent development). If your agent jumps straight to code, the bootstrap instruction didn't load — reinstall and check the plugin marketplace log.
Do this
- Start with the default skills; customize later
- Version your skills in the repo with your code
- Review the design doc phase — it catches bad requirements early
- Keep MCP servers minimal & audited
Avoid this
- Installing 50 random skills from the internet (prompt-injection risk)
- Letting the agent skip tests — the methodology is the value
- Treating skills as a replacement for MCP
- Blindly auto-merging hours-long autonomous runs without review
6. FAQ
Q: Are Agent Skills the same as MCP?
No. MCP is a protocol for connecting tools/data to agents ("what can it reach"). Skills are reusable playbooks ("how should it behave"). They're complementary — a skill can even instruct the agent to use specific MCP tools.
Q: Why did Superpowers grow so fast (282k stars)?
Three reasons: it works across 11+ harnesses including Claude Code, Cursor, and Codex; it ships a complete methodology (brainstorm → worktrees → subagent-driven development) instead of just snippets; and it's listed in both Anthropic's and OpenAI's official marketplaces, giving it credibility on both sides of the ecosystem war.
Q: Do I need Superpowers if I only use Claude Code?
Not strictly — Anthropic's official skills cover many tasks. But Superpowers adds the process discipline (design sign-off, isolated worktrees, enforced TDD) that most people find transforms agent output quality from "impressive demo" to "production code." Try both: they install side by side.
Q: Is it safe to install community skills?
Treat skills like code: review the SKILL.md before installing, prefer well-known authors, and beware of skills that ask for secrets or exfiltration endpoints. The top frameworks (Superpowers, official Anthropic skills) are widely audited; long-tail skills are not.
Q: What about OpenClaw / other agent platforms?
OpenClaw and similar platforms expose their own skill systems with the same underlying idea. The ecosystem is converging on the SKILL.md convention, so skills you write for Claude Code are increasingly portable to other harnesses with minor tweaks.