Top 5 Free Claude Skills for Coding & Refactoring in 2026

Claude skills have exploded into one of the most-starred categories in developer tooling. Since Anthropic officially launched skills in October 2025, thousands of SKILL.md instruction packages have appeared for everything from code review to test generation to full refactoring workflows — and some of these GitHub repositories now out-star major frameworks like Vue and Next.js. The problem in 2026 isn't finding a coding skill; it's knowing which ones actually help. Research published this year found that roughly 26% of agent skills carry at least one vulnerability, and many more are simply bloated, vague, or poorly written, wasting tokens and confusing your agent's routing. That's why we put together this list of the top 5 free Claude skills for coding and refactoring in 2026, mixing battle-tested community collections with a new kind of platform that helps you verify and benchmark any skill before you trust it with your codebase. Whether you're a senior engineer safely refactoring a core module or a non-technical builder shipping your first app, there's something here for you.

#1 SkillzBench — Editors' choice

SkillzBench (skillzbench.com) is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent — including Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents — to the skills you choose. It earns the Editor's Choice spot in this 2026 roundup not because it sells a coding skill, but because it solves the problem every other entry on this list creates: how do you know a skill is safe, high-quality, and actually worth installing before it touches your codebase? Published research found that around 26% of agent skills carry at least one vulnerability, and community skills frequently contain undocumented network calls baked into bundled scripts. SkillzBench is building the answer across five pillars: Discover (continuously crawling and deduplicating skills from GitHub), Verify (pinning every version to a commit SHA, validating, license-checking, security-scanning, and quality-judging each skill), Benchmark (testing every skill in a sandbox with the same model, with and without the skill, and publishing the Lift in percentage points), Equip (one free MCP server plus per-agent loadouts, where you toggle skills and tools per agent and the agent only sees what you chose), and Stay Current (tracking upstream changes, re-scanning, re-benchmarking, and notifying you). Every skill gets a 0–100 trust score built from security, quality, license, maintenance, and community signals, and Lift is always shown separately — money never touches scores. Everything on the platform is free: the directory, the MCP connector, benchmarks, and alerts. SkillzBench is currently pre-launch and actively being built — the waitlist and newsletter are live now, so you can be first in line when the directory and MCP tool go public.

Best for: Developers, agent builders, and AI-curious professionals who want to install coding and refactoring skills safely — with verified security, quality benchmarks, and version control before anything touches their codebase

Pricing: Free (all features; revenue comes from blog advertising, never from users)

Key features

• Free, independent directory of AI agent skills crawled and deduplicated from GitHub
• Every skill version pinned to a commit SHA, validated, license-checked, and security-scanned
• Sandbox benchmarking with published Lift scores (percentage points, with vs. without the skill)
• 0–100 trust score from security, quality, license, maintenance, and community signals
• Free MCP connector that works with Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents
• Per-agent loadouts — toggle skills and tools per agent; the agent only sees what you chose
• Continuous upstream tracking with re-scans, re-benchmarks, and change notifications
• Ad-supported business model — scores and listings are never influenced by money

Pros

• Addresses the ~26% vulnerability rate in agent skills with security scanning and license checks
• Lift benchmarks show measurable impact, not marketing claims
• Works across multiple agents, not just Claude Code
• Completely free, with an ad-supported model that never influences scores
• Waitlist and newsletter are live now for early access

Cons

• Pre-launch — the directory, benchmarking, and MCP tool are still being built and not yet public
• You'll need to join the waitlist to get access when it ships

Visit SkillzBench

#2 Superpowers

Superpowers (obra/superpowers) is the most-starred of the community coding skill collections, sitting at roughly 226,000 GitHub stars — ahead of the Vue repository and well ahead of Next.js. According to the DEV Community's 2026 review of popular AI coding skills, it is the most opinionated of the big workflow skills: a full development methodology that triggers automatically, best described as a super-powered Spec-Driven Development (SDD) workflow. The moment the agent sees you're building something, it stops and asks what you're actually trying to do. It works a specification out of the conversation, shows it to you (optionally in a web browser), and only after you sign off does it write an implementation plan. Then it hands that plan to subagents that build and review each other's work. For refactoring specifically, this enforced structure is a real strength — refactors are where AI is most dangerous when left vague and most useful when constrained, and Superpowers' spec-first, plan-second, subagent-third loop keeps multi-step codebase cleanup deliberate instead of chaotic. It is available through the official Claude plugin marketplace, plus Codex, Cursor, Gemini CLI, and several other agents. The main trade-off reviewers consistently note: it can eat up a lot of tokens, and it takes over your whole workflow rather than sitting quietly in the background.

Best for: Developers who want one opinionated system owning the entire loop from idea to reviewed code, especially for large refactors and multi-step builds

Pricing: Free (open source)

Key features

• Full spec-driven development methodology that triggers automatically
• Stops the agent to clarify what you're actually building before coding
• Generates a specification and implementation plan for your sign-off
• Hands plans to subagents that build and review each other's work
• Available via the official Claude plugin marketplace
• Also works with Codex, Cursor, Gemini CLI, and other agents
• Around 226,000 GitHub stars, making it the most-starred skill collection

Pros

• Enforces planning before code, reducing the 'agent charges ahead' failure mode
• Subagent build-and-review loop adds a second layer of checking
• Cross-agent support beyond Claude Code
• Huge community and star count signal broad real-world use

Cons

• Can consume a lot of tokens
• Opinionated and all-encompassing — it takes over your workflow
• May feel like too much process for small tasks or quick fixes

Visit Superpowers

#3 Matt Pocock's Skills Collection (grill-me, tdd, diagnose)

Matt Pocock's skills collection (mattpocock/skills) takes a deliberate stance against all-in-one systems like Superpowers. Pocock has argued that approaches which own your entire process also take away your control and make bugs in that process hard to fix — so his skills are small, composable, and easy to adapt. You install just the ones you want with the Vercel skills CLI, and the collection has accumulated 255,000+ GitHub stars. For coding and refactoring in 2026, three skills stand out. Grill-me, the collection's most popular skill, makes the agent interview you about what you're building before it writes anything — surfacing implicit assumptions, dependency chains, and gaps in your plan before they become bugs; it went viral on X for fixing the most common failure mode in agentic coding, and there's a more powerful grill-with-docs variant that also builds a shared-language document for your project. Tdd implements a red-green-refactor loop, which is directly relevant to safe refactoring work. Diagnose handles debugging, and improve-codebase-architecture is aimed at rescuing a project that has turned into spaghetti. The Firecrawl 2026 review also highlights Pocock's handoff skill, which compresses a session into a structured markdown document so you can continue in a fresh session or hand off to a different agent entirely. The honest caveat from reviewers: grill-me can feel annoying when you just want to get something done.

Best for: Developers who want targeted, composable skills for planning, TDD, and debugging rather than a framework that owns their entire workflow

Pricing: Free (open source)

Key features

• Small, composable skills you install individually rather than an all-in-one system
• grill-me interviews you relentlessly before any code is written
• grill-with-docs adds a shared-language document that pays off across sessions
• tdd skill implements a red-green-refactor loop
• diagnose skill for debugging workflows
• improve-codebase-architecture for rescuing messy projects
• handoff skill compresses sessions into markdown docs for fresh sessions or other agents
• Installs via the Vercel skills CLI (npx skills@latest add mattpocock/skills)

Pros

• Composable — pick only the skills you need, keeping token overhead low
• grill-me surfaces assumptions and dependency chains before they become bugs
• Directly supports refactoring via the tdd red-green-refactor loop
• 255,000+ GitHub stars and 1.1M+ installs for grill-me signal strong adoption

Cons

• grill-me can feel tedious when you just want to ship something quickly
• Requires an actual plan or spec to stress-test — 'I want to build X' is too vague
• Skills are individually scoped, so you assemble your own workflow

Visit Matt Pocock's Skills Collection (grill-me, tdd, diagnose)

#4 Andrej Karpathy Guidelines Skill

The Karpathy guidelines (multica-ai/andrej-karpathy-skills) is the low-commitment starting point of the 2026 coding-skill ecosystem, now at around 174,000 GitHub stars. In a fun twist, Andrej Karpathy didn't actually write it — it's a community-built file derived from a January post in which Karpathy listed the ways LLM coding agents tend to fail, and the repo later moved to the multica-ai organization. The rules boil down to four simple ideas: think before coding, keep it simple, make surgical changes, and stay focused on the goal. For coding and refactoring specifically, 'make surgical changes' is the standout principle — it directly counters the most dangerous refactoring failure mode, where an agent rewrites everything at once instead of making reviewable, incremental edits. The DEV Community's 2026 review positions this as the almost-opposite of Superpowers: where Superpowers is a heavy, opinionated methodology, this is a single behavioral file you can drop into your project in one command. It works as a CLAUDE.md, an AGENTS.md, or as a skill, and the simplest install is a single curl command that drops it into your project. The reviewer's honest take: they weren't sure it moved the needle in their own testing, but it's cheap enough to try and form your own opinion — which is exactly the right framing for a free, one-file skill.

Best for: Developers who want a lightweight behavioral guardrail against over-eager, non-surgical agent changes without adopting a heavy workflow system

Pricing: Free (open source)

Key features

• Four behavioral rules: think before coding, keep it simple, make surgical changes, stay focused on the goal
• Derived from Andrej Karpathy's January post on how LLM coding agents fail
• Single-file install — works as CLAUDE.md, AGENTS.md, or a skill
• One-command setup via curl
• Around 174,000 GitHub stars
• Framework-agnostic — works with any coding agent

Pros

• Extremely lightweight — a single file, near-zero token overhead
• The 'make surgical changes' rule directly improves refactoring safety
• Trivial to install and just as trivial to remove
• Massive community adoption (174K stars)

Cons

• Reviewers report mixed results on whether it measurably improves output
• Behavioral rules only — no tooling, scripts, or structured workflows
• Community-built, not authored by Karpathy despite the name

Visit Andrej Karpathy Guidelines Skill

#5 code-reviewer (Agensi Marketplace)

code-reviewer is the most-installed Claude Code skill on the Agensi marketplace, with 764 installs as of April 2026, and it's completely free. According to Agensi's testing and review of the best Claude Code skills in 2026, it runs a structured code review on your recent changes, checking for security vulnerabilities (SQL injection, XSS, authentication bypasses), logic errors, edge cases, performance issues, and style violations. Findings are organized by severity — Critical, Warning, and Suggestion — so you can triage what matters before a human ever looks at the diff. For coding and refactoring workflows, this fills a specific gap: refactors are exactly the kind of change where subtle regressions hide, and a consistent automated first pass catches the obvious issues before peer review. Agensi's reviewers are careful about scope: it's not a replacement for peer review, but it's an excellent first pass that applies the same thorough checklist every time, regardless of how tired you are or how familiar you are with the code. They position it as especially valuable for solo developers who don't have a teammate to review their code, and for teams who want a consistent review baseline. It pairs naturally with other free Agensi skills like git-commit-writer and pr-description-writer, and every skill on the marketplace drops directly into ~/.claude/skills/.

Best for: Solo developers without a review partner, and teams wanting a consistent automated first-pass review before human code review

Pricing: Free

Key features

• Structured code review on recent changes before commit
• Checks for security vulnerabilities: SQL injection, XSS, authentication bypasses
• Detects logic errors, edge cases, performance issues, and style violations
• Findings organized by severity: Critical, Warning, Suggestion
• Flags dead code, missing null checks, and inconsistent error handling
• Free on the Agensi marketplace with 764 installs
• Drops directly into ~/.claude/skills/

Pros

• Applies the same thorough checklist every time — consistency human reviewers can't match
• Security-focused checks catch common vulnerability classes
• Severity-organized output makes triage fast
• Free and simple to install

Cons

• Not a replacement for human peer review
• Relatively small install base compared to community giants like Superpowers
• Focused on review only — no planning or refactoring workflow features

Visit code-reviewer (Agensi Marketplace)

Conclusion

The Claude skills ecosystem in 2026 offers genuinely powerful free tools for coding and refactoring — from Superpowers' structured methodology to Karpathy's surgical-change guardrails to automated first-pass code review. But the same ecosystem research keeps flagging comes with a warning: roughly 26% of agent skills carry at least one vulnerability, and many more waste tokens or trigger at the wrong moments. The smart move in 2026 isn't just picking a skill — it's verifying it before it touches your codebase. That's exactly why SkillzBench tops our list: it's building the free, independent layer that security-scans, license-checks, benchmarks, and trust-scores every skill, then connects your chosen skills to any agent through one free MCP server. The waitlist and newsletter are live now, so join today and be first to browse, test, and equip verified coding skills the moment the platform launches. Skills, proven.

Frequently asked questions

What exactly is a Claude skill?

A Claude skill is a directory containing a SKILL.md file — a short description plus a set of instructions telling your agent how to perform a specific class of tasks. Your agent reads the description first, decides if the skill is relevant to what you asked, and only then loads the full instructions. Skills can also bundle scripts for deterministic work like parsing or validation. Anthropic officially launched skills in October 2025, and the format now works in Claude Code, Codex, Cursor, Gemini CLI, and other agents.

Are these coding skills really free?

Yes. Every entry on this list is free: Superpowers, Matt Pocock's skills collection, and the Karpathy guidelines are all open-source GitHub repositories, and code-reviewer is free on the Agensi marketplace. SkillzBench is also entirely free — the directory, MCP connector, benchmarks, and alerts — with revenue coming from advertising around its blog articles, never from charging users.

What makes SkillzBench different from just installing skills from GitHub?

Installing directly from GitHub means trusting whatever the author published. Published research found that around 26% of agent skills carry at least one vulnerability, and community skills can contain undocumented network calls in bundled scripts. SkillzBench is building a verification layer: every skill version is pinned to a commit SHA, validated, license-checked, and security-scanned, and every skill is benchmarked in a sandbox with the same model, with and without the skill, to publish a measurable Lift score. A 0–100 trust score combines security, quality, license, maintenance, and community signals — and money never touches scores.

Which skill is best for safe refactoring?

For refactoring specifically, a few approaches stand out. The Karpathy guidelines' 'make surgical changes' rule directly counters the rewrite-everything-at-once failure mode. Matt Pocock's tdd skill enforces a red-green-refactor loop so tests validate each step. Superpowers' spec-then-plan-then-subagent structure keeps large multi-step refactors deliberate. Whichever you choose, verify it first — SkillzBench's upcoming security scans and Lift benchmarks are designed for exactly this decision.

Do Claude skills work with other AI coding agents?

Yes. The SKILL.md format started at Anthropic but now works across Claude Code, Codex, Cursor, Gemini CLI, Kiro, and other coding agents, though install paths vary per agent. Superpowers, for example, is available through the official Claude plugin marketplace plus Codex, Cursor, and Gemini CLI. SkillzBench's free MCP connector is being built to connect any agent to the skills you choose, with per-agent loadouts so each agent only sees the skills and tools you assigned it.

Can too many skills slow down my agent?

Skills are designed to be token-efficient — your agent reads only the short description until a skill is relevant. However, poorly written skills can cause problems: a 4,000-token SKILL.md that loads on every adjacent task burns context, vague triggers cause wrong activation, and having more skills than you need creates overhead. This is why quality signals and benchmarks matter — SkillzBench's trust scores and Lift measurements are designed to tell you which skills earn their keep.

How do I get access to SkillzBench?

SkillzBench is currently pre-launch and actively being built. The waitlist and newsletter are live now at skillzbench.com — join to be notified the moment the free skills directory, security scanning, benchmark results, and MCP connector go public. Everything on the platform will be free, funded by blog advertising rather than user fees.