With hundreds of agent skills now available on AgentSkillExchange and ClawHub, finding the ones worth your time takes effort. Most skills are niche. Some are redundant. A handful are genuinely useful across almost every workflow.

We tested dozens of skills across real development projects — building APIs, reviewing pull requests, deploying services, writing documentation — and narrowed it down to ten that earned permanent spots in our setup. Each one solves a specific problem better than Claude can on its own, which is the entire point of a good skill.

Here’s what made the list and why.

1. Self-Improving Agent

What it does: Captures errors, user corrections, and failed approaches, then stores them as learnings that persist across sessions. When Claude makes a mistake and you correct it, this skill ensures the same mistake doesn’t happen again.

Why it matters: Claude starts fresh every session. Without this skill, you’ll correct the same behavior over and over — your preferred test framework, that one API quirk in your codebase, the deployment step that always trips it up. The self-improving agent creates a feedback loop that makes Claude meaningfully better at your work over time.

Install:

clawhub install self-improving-agent

Best for: Any developer using Claude Code regularly. The longer you use it, the more value it compounds. It’s the most downloaded skill on ClawHub (244K+ downloads) for a reason — it addresses the single biggest frustration with LLM assistants: they don’t learn from their mistakes.

What to know: Learnings accumulate in a local file. Review it periodically to prune entries that no longer apply. The skill is smart about categorizing corrections (tool failures, knowledge gaps, preference mismatches), but a manual review every few weeks keeps it sharp.

2. GitHub Issues (gh-issues)

What it does: Fetches GitHub issues, spawns sub-agents to implement fixes, opens pull requests, and monitors PR review comments — all automatically. You point it at a repo and a label, and it works through the backlog.

Why it matters: Bug triage is one of the highest-leverage places to apply AI. The gh-issues skill doesn’t just read issues; it understands them, writes code to fix them, creates branches, and submits PRs with proper descriptions. For teams drowning in their issue tracker, this is transformative.

Install:

clawhub install gh-issues

Best for: Open source maintainers and teams with a backlog of well-labeled issues. Works particularly well for bug fixes, small feature requests, and documentation issues. Complex architectural work still needs human judgment, but this skill handles the rest.

What to know: Requires gh CLI authentication. Supports filtering by label, milestone, and assignee. The --watch flag lets it monitor for new issues and PR review comments continuously. Start with --dry-run to see what it would do before letting it loose.

3. GitHub CLI (github)

What it does: Wraps the gh CLI for all standard GitHub operations — checking PR status, creating issues, viewing CI run logs, listing workflows, and querying the GitHub API directly.

Why it matters: Claude can run gh commands without a skill, but it doesn’t always know the right flags, the output format to expect, or how to handle pagination. This skill teaches Claude the patterns that work: how to filter issues efficiently, how to read CI logs without dumping thousands of lines, how to create PRs with proper conventional formatting.

Install:

clawhub install github

Best for: Any developer whose workflow touches GitHub. Pairs naturally with the gh-issues skill — github handles day-to-day operations while gh-issues handles automated issue processing.

What to know: Requires gh auth login to be configured first. The skill handles both github.com and GitHub Enterprise instances. It’s particularly good at CI debugging — ask it to check why a build failed and it’ll pull the right log section instead of dumping the entire output.

4. Coding Agent

What it does: Delegates substantial coding tasks to background sub-agents (Codex, Claude Code, or Pi) so your main Claude session stays responsive. Think of it as Claude spawning workers for heavy lifting.

Why it matters: Some coding tasks — building a new feature, refactoring a module, reviewing a large PR — take time. Without this skill, your Claude session blocks while it works. The coding agent skill runs these tasks in parallel, in the background, and reports back when finished.

Install:

clawhub install coding-agent

Best for: Complex projects where you need Claude working on multiple things simultaneously. Especially useful for PR reviews (spawn in a temp directory), feature builds, and large refactoring efforts. Not needed for quick one-liner fixes — just use the edit tool directly for those.

What to know: The skill manages process lifecycle, output capture, and error handling across different coding agent backends. Claude Code uses --print --permission-mode bypassPermissions (no PTY needed). Codex and Pi need pty:true.

5. Summarize

What it does: Extracts readable text and transcripts from URLs, YouTube videos, podcasts, and local files. Ask it to summarize a 90-minute conference talk, and it pulls the transcript and gives you the key points in seconds.

Why it matters: Developers consume an enormous amount of content — blog posts, documentation, video tutorials, podcast discussions about new tools. This skill turns any media into text Claude can work with, which means you can ask follow-up questions, compare approaches across sources, or extract specific technical details without watching or reading the entire thing.

Install:

clawhub install summarize

Best for: Research-heavy work. Evaluating a new library by summarizing its documentation and intro talks. Getting up to speed on an RFC or long GitHub discussion. Extracting the three things you actually need from a 45-minute YouTube tutorial.

What to know: Handles multiple media formats — web pages, YouTube, podcast RSS feeds, local video/audio files. For YouTube, it pulls transcripts directly when available (no audio processing needed). Falls back to audio extraction for sources without text transcripts.

6. SEO Content Writer

What it does: Writes search-engine-optimized content with proper keyword integration, heading structure, meta descriptions, internal linking, and FAQ sections. Applies the CORE-EEAT benchmark for content quality.

Why it matters: “Write me a blog post” gets you generic output. This skill produces content that follows actual SEO methodology: keyword mapping, featured snippet optimization, strategic header placement, and readability scoring. For dev teams maintaining a technical blog or documentation site, it’s the difference between content that ranks and content that doesn’t.

Install:

clawhub install seo-content-writer

Best for: Developer advocates, technical writers, and anyone maintaining a blog alongside their codebase. The skill includes templates for how-to guides, comparison articles, listicles, and ultimate guides. It won’t replace a skilled content strategist, but it handles the structural SEO work that most developers skip entirely.

What to know: Works best when you provide a target keyword, secondary keywords, and competitor URLs. The skill includes reference materials for title formulas and content structure templates. Pair it with the content gap analysis skill for a full content pipeline.

7. Weather

What it does: Gets current weather and forecasts via wttr.in or Open-Meteo. No API key, no configuration, no authentication.

Why it matters: Not for the weather itself — for what it represents. The weather skill is the cleanest example of what a well-designed skill looks like: narrow scope, clear description, zero unnecessary complexity. As Thariq noted, the best skills don’t try to do everything. They do one thing, correctly, with good error handling.

Install:

clawhub install weather

Best for: Daily use (obviously), but more importantly, studying as a template. If you’re building your first skill, read the weather skill’s SKILL.md before you start. It demonstrates description optimization, scope boundaries, and fallback handling in under 50 lines. The tutorial from Day 4 covers similar principles.

What to know: Uses wttr.in by default, falls back to Open-Meteo for detailed forecasts. Works for any location worldwide. Zero configuration required — just ask about the weather.

8. Frontend Design (SuperDesign)

What it does: Guides Claude to produce modern, visually polished user interfaces with proper spacing, typography, color systems, and responsive design. Turns “make me a dashboard” from a vague request into a structured design process.

Why it matters: Claude can write HTML and CSS, but without design guidance, the output tends toward generic Bootstrap-looking layouts. This skill encodes actual design principles — visual hierarchy, whitespace ratios, component patterns, accessibility standards — so the generated UI looks like a designer built it.

Install:

clawhub install frontend-design

Best for: Developers building internal tools, landing pages, dashboards, or prototypes who don’t have a designer on the team. Also useful for hackathons and MVPs where visual quality matters but hiring a designer isn’t realistic.

What to know: Works with any CSS framework (Tailwind, plain CSS, etc.). The skill provides design guidelines and component patterns, not pre-built templates — Claude still generates the actual code, just with better design judgment. Results improve significantly when you describe the purpose of the UI, not just the layout.

9. Skill Creator

What it does: Helps you create, improve, audit, and publish agent skills. It knows the AgentSkills spec, validates your SKILL.md structure, checks description field optimization, and ensures your file layout follows best practices.

Why it matters: Building a skill is straightforward. Building a good skill that others can actually use requires knowing the spec, the conventions, and the common mistakes. This skill encodes all of that — it’s a linter and style guide for SKILL.md files.

Install:

clawhub install skill-creator

Best for: Anyone publishing skills to ClawHub or AgentSkillExchange. Run it before submission to catch problems: missing gotchas sections, overly broad descriptions, hardcoded configuration, files that should be in references/ instead of the main SKILL.md. It’s the quality gate between a skill that works for you and a skill that works for everyone.

What to know: Validates against the AgentSkills.io standard. Can audit existing skills and suggest improvements. Also handles the mechanics of publishing — directory structure, frontmatter formatting, and metadata requirements.

10. Google Workspace (gog)

What it does: Full CLI access to Gmail, Google Calendar, Drive, Contacts, Sheets, and Docs. Read emails, create calendar events, search Drive files, update spreadsheets — all through natural language via Claude.

Why it matters: Most developers live in two worlds: their IDE and their browser (usually with Gmail and Calendar open). This skill bridges them. Ask Claude to check your calendar before scheduling a deployment. Have it draft an email summarizing today’s PR activity. Search Drive for that architecture doc someone shared three months ago. It removes the context-switching tax.

Install:

clawhub install gog

Best for: Developers who use Google Workspace and want their AI assistant to integrate with it. Particularly powerful for team leads who need to coordinate across code work and communication — checking emails, scheduling meetings, and updating shared docs without leaving the terminal.

What to know: Requires one-time OAuth setup with a Google Cloud project. Supports multiple Google accounts. The CLI covers Gmail (read, send, search, labels), Calendar (events, scheduling), Drive (search, upload, download), Sheets (read, write, formulas), and Docs (read, create). Setup takes about 10 minutes but only needs to happen once.

Honorable Mentions

Ten is arbitrary. Here are four more that nearly made the cut:

• nano-banana-pro — Image generation and editing via Gemini. Useful for creating diagrams, social images, or quick visual assets without leaving your workflow.
• tmux — Remote-controls tmux sessions for interactive CLIs. Essential if you run long-lived terminal processes that need monitoring.
• healthcheck — Security hardening and risk assessment for your deployment host. Checks firewalls, SSH config, updates, and exposure. Good for VPS hygiene.
• openai-whisper-api — Audio transcription via OpenAI’s Whisper API. When the summarize skill’s transcript extraction doesn’t cover your audio format, this is the fallback.

How to Choose Skills for Your Setup

Don’t install all ten. Start with the ones that match your actual workflow:

• Every developer: self-improving-agent (always), github (if you use GitHub), coding-agent (for complex projects)
• Open source maintainers: Add gh-issues for automated issue processing
• Content creators / DevRel: Add seo-content-writer and summarize
• Full-stack developers: Add frontend-design for UI work
• Skill authors: Add skill-creator before publishing anything
• Productivity-focused: Add gog for Google Workspace integration

The beauty of agent skills is that they’re modular. Install what you need, remove what you don’t. Claude only loads a skill when its description matches your request, so installing extra skills has minimal overhead — they don’t compete for context until they’re activated.

What’s Missing from the Ecosystem

After testing this many skills, some gaps stand out. The ecosystem still lacks strong options for:

• Database schema management — skills that understand your schema and help with migrations, query optimization, and data modeling
• Monitoring and alerting — connecting Claude to Datadog, Grafana, or PagerDuty for intelligent incident triage
• Infrastructure as Code — Terraform and Pulumi skills that go beyond basic template generation
• API testing — automated test generation and execution for REST and GraphQL endpoints

If you’re looking for a skill to build, these are the areas where the community needs help most. Check the skill categories taxonomy for more context on where each of these would fit.

Frequently Asked Questions

How do I install agent skills?

Use the ClawHub CLI: clawhub install [skill-name]. Skills install to your workspace’s skills directory and are available immediately. No restart required. You can also install from a GitHub URL or a local directory.

Do installed skills slow down Claude?

No. Claude reads skill descriptions at the start of each turn (a few lines each), but only loads the full SKILL.md when a skill’s description matches your request. Ten installed skills add negligible overhead. A hundred might start to matter for description scanning, but most setups stay well under that.

Can I use these skills with other AI coding tools?

Skills that follow the AgentSkills.io standard are designed to be portable. Claude Code has the deepest integration, but Cursor, Windsurf, and other tools are adding skill support. The SKILL.md format is plain markdown — any tool that reads markdown files can use them, even if the integration isn’t as seamless.

How do I know if a skill is well-made?

Check for three things: a gotchas section with specific failure points (not generic advice), a description field that includes trigger phrases and exclusions, and a file structure that uses progressive disclosure (references/ and scripts/ directories rather than one massive file). Skills on AgentSkillExchange are curated for these standards.

Wrapping Up

Agent skills are the difference between Claude as a general-purpose assistant and Claude as a specialized tool that understands your stack, your workflow, and your preferences. The ten skills listed here cover the broadest range of developer needs — from code review to content creation, from GitHub automation to Google Workspace integration.

Start with self-improving-agent. Add the ones that match your daily work. Build your own for the gaps. The ecosystem is still early enough that a well-built skill in an underserved category can become the default choice for thousands of developers.

Browse all available skills on AgentSkillExchange, or check the 15-minute tutorial if you’re ready to build your own.