AI Claude Code skills custom commands developer tools GitHub subagents plugins

15 Claude Code Skill Candidates: Choose by Task, Evidence, and Cost

The Prompt Shelf ·

Claude Code skills package reusable instructions and optional supporting files. Claude can load a skill when relevant, or a user can invoke it directly. A skill package does not establish that its output is correct or that it improves on your existing workflow.

October 18, 2026 correction: This is an unranked candidate directory, not a tested best-of list. We could not substantiate the earlier claim of a comparative review of 200+ repositories from the evidence available for this update. We removed the star-count rankings, blanket must-install language, and unsupported success/cost guarantees. The order below does not imply measured quality. Most individual feature descriptions remain untested; check the selected maintainer documentation before relying on them.

Before installing: choose a task and check the package

October 17, 2026 correction: The earlier blanket claim that every entry is free and installs in under a minute was not established. This update corrects the generic installation advice; it does not revalidate every ranking, star count, or feature claim below. Check the selected project’s current license, prerequisites, and installation instructions before deciding.

Start with one repeatable task, such as auditing one sample page or explaining one unfamiliar module. Try your existing prompt and tools first. If they already meet your acceptance criteria, no additional skill or paid tool is needed. When comparing a skill, record setup time, review effort, output defects, and any required model/API usage separately from the repository’s license. A downloadable repository does not establish zero operating cost.

How to Install a Skill

There is no universal repository-clone or single-file-copy recipe for this list. Claude Code’s official skills documentation describes a project skill at .claude/skills/<skill-name>/SKILL.md, with optional supporting files in that skill directory. Copying only SKILL.md can omit referenced scripts or reference material; cloning a multi-skill repository into that path does not establish that its package layout is correct.

Two examples checked against their maintainers’ documentation on October 17:

CandidateInstallation distinction to check
ImpeccableIts maintainer documents an installer, plugin distribution, and tool-specific compiled files. Select the instructions for your actual tool instead of assuming the repository root is one standalone skill.
Understand-AnythingThe earlier Lum1104 URL redirects to Egonex-AI. Its README lists Claude Code plugin installation separately from other platforms. Use the current maintainer instructions for your platform.

Before changing your working environment, check which files, dependencies, hooks, or permissions that chosen method introduces. Keep required supporting files together and record the version or commit you reviewed. In an authorized disposable project, confirm that the intended command is available and compare its result with your baseline task. Merely downloading a file is not an end-to-end installation test.

This article update used documentation review only: neither package was installed or executed, and no account or permission changes were made. If the package requirements or operating costs remain unclear, postpone that installation or purchase rather than treating the list as a requirement to buy anything.

Choose the evaluation before choosing the package

Your taskCandidate with documentation checked October 18Compare with your current method
Review one interfaceImpeccable documents design review commandsUse the same page and acceptance checklist. Record actionable findings, false positives, regressions after fixes, and review time.
Resume a multi-session taskPlanning With Files documents persistent plan, findings, and progress filesCompare against your existing notes on one disposable task. Check that the next action and unresolved decisions survive a new session; record setup and maintenance effort.

These are proposed comparison procedures, not completed product tests. Documentation establishes the intended workflow, not a measured advantage. Record the selected version, model, dependencies, enabled hooks, and any billed model/API usage. Repository popularity and a permissive license cannot establish task fit or zero running cost. Keep your current method if no useful improvement is demonstrated.

Candidate directory — no quality ranking

Impeccable — Design Audit & Polish

GitHub: pbakaus/impeccable

The maintainer documents a design workflow with audit, critique, polish, and other commands. The current README uses /impeccable audit and /impeccable polish; standalone shortcuts such as /audit require the documented pinning step. Do not assume a command name from an older article is available after installation.

We reviewed the README, not a running installation. Test on an authorized sample interface before applying generated edits to a product.

Planning With Files — Persistent Task Management

GitHub: OthmanAdi/planning-with-files

The maintainer documents task_plan.md, findings.md, and progress.md as persistent working files, with integration-specific hooks and recovery behavior. The current README also says completed working plans are not automatically archived: retain important decisions in durable project documentation or version control.

That design may help a long-running task, but it is not proof that no progress can be lost. We did not install the hooks or test recovery in an agent session.

Humanizer — Remove AI Writing Patterns

GitHub: blader/humanizer

Based on Wikipedia’s “Signs of AI writing” guide, Humanizer detects and removes 20+ patterns that make text obviously AI-generated: inflated symbolism, em dash overuse, rule of three, AI vocabulary (“delve,” “tapestry,” “landscape”), and negative parallelisms.

Does a second audit pass specifically looking for “obviously AI-generated” markers. Evaluate whether its edits preserve meaning and suit your intended readers.

Understand-Anything — Codebase Knowledge Graphs

GitHub: Lum1104/Understand-Anything

Turns any codebase into an interactive knowledge graph. A multi-agent pipeline scans your project, extracts every file, function, class, and dependency, builds a knowledge graph, and opens an interactive web dashboard with color-coded architectural layers.

Key commands:

  • /understand-diff — Visualize what changed
  • /understand-explain — Get architectural explanations
  • /understand-chat — Ask questions about the codebase

Perfect for onboarding onto unfamiliar codebases.

Dev Browser — Give Claude a Web Browser

GitHub: SawyerHood/dev-browser

Gives Claude the ability to use a real web browser with cookies, localStorage, and page state that persist between executions. Uses ARIA-based element discovery with semantic snapshots.

Claude autonomously decides when to use the browser — for testing deployed features, verifying API responses, or checking how a page renders. Turns Claude from a code-only assistant into one that can see and interact with your running application.

Trail of Bits Security Skills

GitHub: trailofbits/skills

From one of the most respected security research firms. Includes CodeQL and Semgrep static analysis, Semgrep rule creation, smart contract entry-point analysis, variant analysis, audit context building, and security-focused differential review.

If you write code that handles money, auth, or sensitive data, these skills turn Claude into a security auditor that catches vulnerabilities before they ship. One common finding: real API keys or tokens hardcoded in a config file the audit skill flags. 1Password service accounts fix that pattern at the source — secrets get pulled at runtime instead of sitting in a file for a scan (or a leaked laptop) to find.

Playwright Skill — Browser Automation

GitHub: lackeyjb/playwright-skill

Claude autonomously writes and executes Playwright tests and automation. v4.0 includes auto-detection that scans common dev server ports (3000, 5173, 8080) and smart test management.

Different from Dev Browser in that this focuses on automated testing workflows rather than interactive browsing.

Godogen — Build Godot Games From Text

GitHub: htdt/godogen

A candidate aimed at Godot 4 game generation from text descriptions. The AI pipeline handles architecture planning, 2D/3D asset generation (via Gemini and Tripo3D), GDScript writing, and visual QA by capturing screenshots from the running engine.

Includes a custom language spec and quirks database for GDScript, which is critical because GDScript has enough differences from Python that generic AI coding produces broken code.

Context Engineering Kit

GitHub: NeoLabHQ/context-engineering-kit

Hand-crafted context engineering techniques with minimal token footprint. Includes Spec-Driven Development (turns prompts into production-ready implementations through structured planning), domain-driven design patterns, and subagent-driven development.

We have not verified a benchmark or production success rate for this entry. Treat its techniques as candidates to evaluate against your own test cases, not a working-code guarantee.

Learning Opportunities — Deliberate Skill Development

GitHub: DrCatHicks/learning-opportunities

The most unusual skill on this list. After you complete architectural work (new files, schema changes, refactors), Claude offers optional 10-15 minute learning exercises based on cognitive psychology research: prediction, generation, retrieval practice, and spaced repetition.

The idea is that AI-assisted coding can make you a worse programmer if you never engage with the “why.” This skill fights that by turning every coding session into a learning session.

Sahil Lavingia’s Entrepreneur Skills

GitHub: slavingia/skills

From the CEO of Gumroad. Claude Code skills based on “The Minimalist Entrepreneur” methodology. Includes /validate-idea, /find-community, /first-customers, and /company-values.

Not a coding skill — it turns Claude into a business advisor. Interesting proof that Skills are not limited to technical workflows.

Claude Sessions — Session Tracking

GitHub: iannuttall/claude-sessions

Custom slash commands for comprehensive development session tracking and documentation. Solves the problem of losing context between Claude Code sessions by maintaining structured logs of what was done, what was decided, and what needs to happen next. For a deeper look at memory and session persistence strategies, see Claude Code Memory Persistence Guide 2026.

Second Brain Skills

GitHub: coleam00/second-brain-skills

Turns Claude Code into a “Second Brain” with three core capabilities: Brand System (define your voice and brand once, reuse everywhere), MCP Client Integration (progressive disclosure of tool capabilities), and Remotion skill for programmatic video creation.

Prompt Master — AI Prompt Writing

GitHub: nidhinjs/prompt-master

Meta-skill: Claude writes optimized prompts for other AI tools. Token or credit savings have not been measured for this article. Useful when you are building AI-powered features and need to craft prompts for production use.

Anything to NotebookLM

GitHub: joeseesun/anything-to-notebooklm

Multi-source content processor. Converts web pages, YouTube videos, PDFs, Markdown, and search queries into Podcasts, presentations, mind maps, and quizzes via NotebookLM integration.

Awesome Lists for Discovery

If you want to explore beyond this list:

Packaging claims are not a selection criterion

The earlier ecosystem section made unverified claims about release dates, dominant architectures, and the superiority of plugin packaging. Those claims are withdrawn in this update. A plugin or a standalone skill still needs a supported installation path, a known version, and a task-level comparison. Use the official skills documentation for current behavior rather than inferring capability from a package label.

What Makes a Good Skill

Use these evaluation criteria; this article does not establish a comparative benchmark:

  1. Focused scope. One skill does one thing well. Impeccable does design. Trail of Bits does security. Mixing concerns produces mediocre results.
  2. Runnable examples. Skills that include example commands and expected outputs get adopted faster.
  3. Manage reference length. The official documentation recommends keeping SKILL.md under 500 lines and moving detail into supporting files. That is authoring guidance, not proof that longer skills are automatically deprioritized or worse.
  4. Domain knowledge baked in. The value of a skill is the expertise it encodes. Godogen’s GDScript quirks database is more valuable than the automation around it.

Frequently Asked Questions

What is a Claude Code Skill?

A Claude Code Skill is a reusable instruction module stored in ~/.claude/skills/{name}/SKILL.md (or .claude/skills/ per project). It encodes domain expertise — security review, performance audit, design system rules, code generators, and more — that Claude can invoke on demand without re-explaining the context every session.

How are Skills different from CLAUDE.md or AGENTS.md?

CLAUDE.md and AGENTS.md are always loaded into the session context. Skills are invoked on demand — Claude reads the SKILL.md only when the task description matches. This keeps the context window lean while preserving access to deep expertise. Skills also support helper files (scripts, templates, references) that load with the skill.

Where do I find official Anthropic Skills?

The anthropics/skills repository contains Anthropic’s official reference implementations plus the Agent Skills spec. For community discovery, ComposioHQ/awesome-claude-skills is one index; it is not a benchmark.

Can I use Claude Code Skills with Codex, Cursor, or Gemini CLI?

Skills are designed for Claude Code, but the SKILL.md format is text-based markdown. Many skills work across tools when the underlying concepts overlap. VoltAgent/awesome-agent-skills curates 1,000+ skills explicitly compatible with multiple agents including Codex, Gemini CLI, and Cursor.

What makes a good Claude Code Skill?

Look for a defined task, runnable examples, useful reference material, and results that meet your acceptance criteria. Length alone does not establish quality; follow current authoring guidance and measure results on your task.

How do I create my own Claude Code Skill?

Create ~/.claude/skills/{skill-name}/SKILL.md with YAML frontmatter (name, description, optionally allowed-tools) followed by markdown instructions. Add helper scripts or templates in the same directory if needed. The skill is auto-discovered on the next Claude Code session.

How do Claude Code Skills relate to Plugins?

Claude Code plugins can package skills with other components. Standalone skills are another distribution option. Check the current documentation and the selected package; this article does not establish a release date, market share, or packaging advantage.

Can I share Skills across my team?

Yes. Skills are plain markdown files, so version control is the standard sharing mechanism. Project-level skills go in .claude/skills/ inside the repository (committed and shared). Personal skills live in ~/.claude/skills/ (per-user). Teams often maintain an internal skill repository and onboard new engineers by cloning it into their global skills directory. The community standard for team sharing is now: ship as a Plugin (private GitHub repo) and add it to your team’s CLAUDE.md as a required install.

Are Claude Code Skills available on claude.ai (web or mobile)?

Skills are a Claude Code feature (the local CLI). They are not directly available on the claude.ai web interface or the Claude mobile apps. If you need a workflow that runs on both the desktop CLI and the web, build it as a system prompt or use Anthropic’s Computer Use / Workspaces features. That said, the SKILL.md format is portable text — you can paste a skill into the web UI as a prompt prefix to get similar behavior, just without auto-discovery.

Related Articles

Explore the collection

Browse all AI coding rules — CLAUDE.md, .cursorrules, AGENTS.md, and more.

Browse Rules