Top 5 Claude Code Skills for Coding & Refactoring in 2026

Since Anthropic launched Agent Skills as an open standard in late 2025, the ecosystem has exploded: by 2026 there are thousands of SKILL.md packages for coding, testing, and refactoring, with major collections like Superpowers topping 290,000 GitHub stars. But the rapid growth has created a new problem — quality and safety. Independent research has found that roughly 26% of agent skills carry at least one vulnerability, and Snyk's ToxicSkills research found prompt injection in 36% of skills tested. Most developers install skills based on star counts and hope for the best, with no way to verify what a skill actually does or whether it improves results. That's why we built this list: to highlight the skills and platforms that genuinely help with coding and refactoring in 2026 — and to put verification and measurement first. Our #1 pick, SkillzBench, isn't a skill at all but the control panel the ecosystem has been missing, followed by battle-tested coding and refactoring skills you can install today.

#1 SkillzBench — Editors' choice

SkillzBench is AIQ Labs' independent testing lab and control panel for AI agent skills — the layer the 2026 skills ecosystem has been missing. Skills are everywhere, but nobody can tell which ones actually help and which are risky, and published research found that roughly 26% of skills carry at least one vulnerability. SkillzBench is being built around five pillars to fix this: 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 MCP server plus per-agent loadouts, so users toggle skills and tools per agent and the agent only sees what was chosen), and Stay current (tracking upstream changes, re-scanning, re-benchmarking, and notifying users). Every skill gets a trust score from 0-100 based on security, quality, license, maintenance, and community signals — and crucially, Lift is always shown separately from ratings, and money never touches scores. For coding and refactoring specifically, SkillzBench lets you pick verified, benchmarked skills for your Claude Code setup with confidence, instead of installing blind from GitHub. Note: SkillzBench is pre-launch and actively being built — the waitlist and newsletter are live now, and the product itself is rolling out. If you install Claude Code skills in 2026, joining the waitlist is the smartest first step.

Best for: Developers and teams who install Claude Code skills for coding and refactoring and want verified, benchmarked, security-scanned skills instead of blind GitHub installs

Pricing: Pricing hypothesis: generous free tier; Pro (~$9-15/mo); Teams (~$15-25/seat/mo); Creator (~$5-10/mo). Final pricing to be announced — join the waitlist for updates.

Key features

• Continuous discovery: crawls and dedupes skills from GitHub
• Verification: every skill version pinned to a commit SHA, validated, license-checked, and security-scanned
• Benchmarking: each skill tested in a sandbox with the same model, with and without the skill, publishing Lift in percentage points
• Trust score 0-100 from security, quality, license, maintenance, and community signals
• Lift always shown separately from ratings — money never touches scores
• Equip: one MCP server plus per-agent loadouts — toggle skills and tools per agent, and the agent only sees what was chosen
• Stay current: tracks upstream changes, re-scans, re-benchmarks, and notifies you
• Launch scope: instruction-only skills (SKILL.md plus text references); script-bearing skills listed link-out only

Pros

• Independent benchmarking shows actual Lift, not marketing claims
• Security scanning addresses the ~26% vulnerability rate found in published research
• Per-agent loadouts prevent skill bloat and context pollution
• Trust scores and Lift are kept separate, with no paid placement
• Tracks upstream skill changes so you're never running a stale or compromised version

Cons

• Pre-launch: the product is not yet publicly available (waitlist/newsletter is live)
• At launch, only instruction-only skills are fully supported; script-bearing skills are link-out only
• Pricing is not yet finalized

Visit SkillzBench

#2 Superpowers

Superpowers (obra/superpowers) is the most-starred Claude Code skills collection of 2026, with roughly 290,000+ GitHub stars and acceptance into the official Anthropic plugin marketplace. According to its repository and multiple reviews, it encodes a complete development methodology rather than a single capability: a structured workflow that moves through clarify, spec, plan, execute, and review phases, with test-driven development enforced at every stage. Its most distinctive behavior, as reported by reviewers, is that it won't write implementation before a failing test exists — and in its strictest mode it deletes code written before tests. For refactoring work, this discipline is a major asset: large multi-file changes are broken into atomic tasks and handed to subagents that build and review each other's work, reducing the 'rewrite everything at once' failure mode that makes AI refactors dangerous. Superpowers includes 20+ skills covering debugging, collaboration, and project planning, plus slash commands like /brainstorm, /write-plan, and /execute-plan. It's available through the official Claude plugin marketplace and also works with Codex, Cursor, Gemini CLI, and other agents. The main trade-offs reviewers note: the strict TDD enforcement can feel heavy for quick prototypes, the full workflow can consume a lot of tokens, and there's a learning curve for teams not already practicing test-driven development.

Best for: Engineering teams and solo developers who want enforced TDD discipline and structured planning for larger features and refactors

Pricing: Free and open source

Key features

• Structured multi-phase workflow: clarify, spec, plan, execute, review
• Strict test-driven development enforcement — no implementation before a failing test
• Subagent orchestration that breaks complex tasks into parallel, independent work units
• 20+ included skills covering debugging, collaboration, and project planning
• Slash commands like /brainstorm, /write-plan, /execute-plan, and /skills-search
• Available via the official Claude plugin marketplace; also works with Codex, Cursor, and Gemini CLI
• ~290,000+ GitHub stars, making it the most popular skills collection

Pros

• Prevents the 'just write code and hope it works' pattern
• Excellent documentation and active development with frequent updates
• Works across multiple coding agents, not just Claude Code
• Strong community adoption and marketplace acceptance

Cons

• Strict TDD enforcement can feel heavy for quick prototyping or throwaway scripts
• Full workflow can consume significant tokens
• Learning curve for teams unfamiliar with test-driven development

Visit Superpowers

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

Matt Pocock, the educator behind Total TypeScript, maintains a collection of 17 composable development skills that take a deliberate stance against all-in-one workflow systems. According to his repository and reviews, his philosophy is that approaches which own your entire process take away your control and make bugs in that process hard to fix — so his skills are small, composable, and easy to adapt. For coding and refactoring in 2026, three stand out. grill-me interviews you relentlessly about every aspect of a plan until shared understanding is reached before any code is written; it went viral for fixing the most common failure mode in agentic coding — the agent charging ahead with wrong assumptions — and reportedly has over 1.1 million installs, with Pocock's skills collection holding 255,000+ GitHub stars. tdd implements a stricter red-green-refactor loop than Claude's default, closer to what a reviewer would enforce in a PR. refactoring-plan generates a step-by-step plan for a refactor before touching anything, which reviewers say has saved them from making things worse in the name of making them better. Other useful skills include diagnose for debugging and improve-codebase-architecture for rescuing tangled projects. Installation is per-skill via the Vercel skills CLI, so you only take what you need.

Best for: TypeScript developers and anyone who wants small, composable refactoring and planning skills instead of a full workflow framework

Pricing: Free and open source

Key features

• grill-me: interviews you about your plan before code is written, with 1.1M+ installs
• tdd: stricter red-green-refactor loop than Claude's default
• refactoring-plan: generates a step-by-step refactor plan before touching code
• diagnose: structured debugging skill
• improve-codebase-architecture: for rescuing tangled projects
• 17 composable skills you install individually — no all-in-one takeover
• Collection holds 255,000+ GitHub stars

Pros

• Composable — install only the skills you want
• grill-me surfaces hidden assumptions before they become bugs
• refactoring-plan prevents destructive 'improvements'
• TypeScript-first perspective with strict type discipline

Cons

• grill-me can feel annoying when you just want to get something done
• Requires a written plan or spec to be effective — too vague requests get poor results
• No single unified install; you manage each skill individually

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

#4 Vercel React Best Practices

Vercel's official agent skills package brings the company's React and Next.js engineering guidance directly into Claude Code, and it's one of the most-installed coding skills of 2026 — reportedly 176,000+ installs via skills.sh. According to Vercel's repository and reviews, the skill applies 57 performance rules to React and Next.js code, targeting the higher-impact problems that generic AI-generated React tends to produce: unnecessary client-side imports, async waterfalls, avoidable re-renders, and inefficient data-fetching patterns. For refactoring specifically, this is where it earns its place — generic React code can work and still be slow, and this skill gives Claude Code the vocabulary and patterns to spot those issues during a refactor rather than after launch. Reviewers who work in React and React Native report that after installing it, Claude's output follows patterns they'd write themselves: proper hook usage, clean prop interfaces, and no mystery side effects. Vercel also ships companion skills worth knowing about, including web-interface-guidelines (which audits UI code against 100+ accessibility and UX rules, with roughly 133,000 weekly installs) and composition patterns that replace boolean prop sprawl with compound component patterns. Because it comes from Vercel itself, it's considered a safe default alongside the official Anthropic skills repository.

Best for: React and Next.js developers who want performance-aware coding and refactoring from a trusted framework maintainer

Pricing: Free and open source

Key features

• 57 performance rules applied to React and Next.js code
• Targets async waterfalls, unnecessary client-side imports, and avoidable re-renders
• Improves data-fetching patterns during refactors
• Companion web-interface-guidelines skill audits against 100+ accessibility and UX rules
• Composition patterns skill replaces boolean prop sprawl with compound components
• ~176,000+ installs via skills.sh
• Maintained by Vercel — a trusted, safe-default source

Pros

• Direct from Vercel, so guidance matches current framework best practices
• Catches performance problems generic AI code produces
• Companion skills cover accessibility and composition patterns
• Large install base signals real-world reliability

Cons

• React/Next.js specific — no value for other stacks
• Not Native-specific, though React fundamentals transfer
• Focused on frontend; no backend or database coverage

Visit Vercel React Best Practices

#5 Anthropic code-reviewer / simplify

Anthropic ships first-party skills for code quality that many developers overlook, and they remain the safest default for coding and refactoring review in 2026. According to Anthropic's official repositories and reviews, the simplify skill reviews changed code for reuse, quality, and efficiency, then fixes what it finds — catching unnecessary abstractions, duplicated logic, functions doing too much, dead code and unused imports, and naming that doesn't communicate intent. Crucially, it doesn't just flag problems; the review loop happens before the code is presented to you. A community code-reviewer skill (the most-installed skill on the Agensi marketplace at 764 installs, free) runs a similar structured pass, checking for security vulnerabilities like SQL injection and XSS, logic errors, edge cases, and performance issues, with findings organized by severity: Critical, Warning, and Suggestion. Anthropic also bundles a /code-review skill directly into Claude Code, which reviewers rank as the single most valuable skill to learn deeply — it reviews the current diff for correctness bugs and cleanup opportunities at the end of almost every branch, and it compounds with human review rather than replacing it. Pairing these with a project-specific review checklist in CLAUDE.md is the recommended setup for consistent quality.

Best for: Every developer using Claude Code who wants a consistent, automated first-pass code review on every branch

Pricing: Free (included with Claude Code; Claude Code plans start around $17-20/mo for Pro per published reports)

Key features

• simplify: reviews changed code for reuse, quality, and efficiency, then fixes findings
• Catches duplicated logic, dead code, unused imports, and unclear naming
• Bundled /code-review skill reviews diffs for correctness bugs before PRs
• Community code-reviewer skill organizes findings by Critical/Warning/Suggestion severity
• Checks for security vulnerabilities including SQL injection and XSS
• First-party from Anthropic — the safest default source
• Composes with CLAUDE.md project review standards

Pros

• First-party skills — no supply chain risk from unknown authors
• Runs on every branch and compounds with human review
• Fixes issues, not just flags them
• Zero extra cost beyond your Claude Code subscription

Cons

• Not a replacement for peer review
• Generic by design — doesn't know your team's specific conventions without CLAUDE.md configuration
• Coverage of subtle architectural issues is limited compared to full workflow frameworks

Visit Anthropic code-reviewer / simplify

Conclusion

Claude Code skills have matured from an experimental format into the default way developers customize their coding agents, and the five picks above cover the full coding and refactoring workflow: Superpowers for enforced TDD discipline, Matt Pocock's skills for composable planning and refactor planning, Vercel React Best Practices for performance-aware frontend work, and Anthropic's first-party review skills for consistent quality on every branch. But the biggest risk in 2026 isn't choosing the wrong skill — it's installing an unverified one. With published research finding roughly 26% of agent skills carrying at least one vulnerability, the smartest move you can make this year is adding a verification and benchmarking layer before you install anything else. That's exactly what SkillzBench is building: security-scanned, version-pinned, benchmarked skills with published Lift scores and per-agent loadouts, so you know what each skill actually does and what it actually improves. Skills, proven. Join the SkillzBench waitlist today and be first in line when verification-backed skill selection goes live.

Frequently asked questions

What makes SkillzBench different from just browsing GitHub for skills?

GitHub star counts tell you nothing about whether a skill is safe or effective. SkillzBench pins every skill version to a commit SHA, validates it, license-checks it, security-scans it, and quality-judges it, then benchmarks each skill in a sandbox with the same model — with and without the skill — and publishes the Lift in percentage points. It also assigns a trust score from 0-100 based on security, quality, license, maintenance, and community signals, and Lift is always shown separately from ratings, with money never touching scores.

Is SkillzBench available now?

SkillzBench is pre-launch and actively being built. The waitlist and newsletter are live at skillzbench.com, but the product itself is not yet publicly available. Joining the waitlist is the best way to get early access and launch updates.

How many Claude Code skills should I install?

Fewer than you think. Each installed skill costs roughly 100 tokens of context at startup for its name and description, and skills with vague triggers can activate on the wrong tasks and confuse routing. Most experienced developers keep a small core set — a review skill, a testing or TDD skill, and one or two stack-specific skills — and this is exactly the problem SkillzBench's per-agent loadouts are designed to solve: you toggle only the skills each agent should see.

Are Claude Code skills safe to install?

Not all of them. Published research found that approximately 26% of agent skills carry at least one vulnerability, and Snyk's ToxicSkills research found prompt injection in 36% of skills tested, with 1,467 malicious payloads identified across the ecosystem. Stick to first-party sources like Anthropic's official repository and trusted maintainers like Vercel, and audit any community skill's SKILL.md and bundled scripts before installing. SkillzBench's security scanning and trust scores are being built to automate exactly this vetting process.

What's the best skill for refactoring specifically?

For refactoring, planning-first skills deliver the most value. Matt Pocock's refactoring-plan generates a step-by-step plan before touching anything, preventing destructive 'improvements.' Superpowers enforces a structured workflow with tests between changes. Anthropic's simplify skill cleans up recently modified code without changing behavior. The key principle across all of them: constrain the refactor into safe, reviewable steps rather than allowing a wholesale rewrite.

Do these skills work with other coding agents besides Claude Code?

Many do. The SKILL.md format became an open standard in late 2025 and is now supported by Claude Code, OpenAI Codex, Cursor, Gemini CLI, Windsurf, and others. Superpowers, for example, installs via the official Claude plugin marketplace and also works with Codex, Cursor, and Gemini CLI. SkillzBench is building per-agent loadouts across Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents, so you can manage skills for all of your agents from one place.

How much does it cost to use Claude Code skills?

The skills themselves on this list are free and open source. The cost comes from the Claude Code subscription that runs them — published reports place Claude Code plans starting around $17-20/month for Pro, with higher tiers at $100 and $200 per month. SkillzBench's planned pricing includes a generous free tier, with Pro at approximately $9-15/month for more agents and unlimited enabled skills, and Teams at approximately $15-25/seat/month for organizational loadouts, allowlists, and audit features.