Awesome AI Agent Skills

A community-shared collection of reusable skills for AI coding agents. Works with Claude Code, Cursor, Kilo Code, Windsurf, OpenAI Codex, and any AI tools that support skills or custom instructions.

Source of Truth

• Author skills only under skills//.
• Assign each skill to exactly one installable bundle in plugin-groups.json, which also holds marketplace ownership, plugin metadata, and routing-safe Codex prompts for multi-skill bundles.
• Change generated marketplace/manifest shapes only in scripts/lib/plugin-shape.js.
• Treat .claude-plugin/marketplace.json, .agents/plugins/marketplace.json, plugins/**, and generated README tables as sync outputs.
• Do not edit plugins//** directly; those are generated plugin packages (Claude Code + Codex) regenerated from skills/.
• Run npm run sync after changing skills or plugin grouping, then run npm run validate.

What are Skills?

Skills are self-contained instruction sets that teach AI agents specific workflows, guidelines, or capabilities. Each skill includes:

• A SKILL.md with metadata and instructions
• Optional reference documentation for detailed guidance
• Reusable across any project

Installation

Claude Code

Step 1: Add the marketplace

/plugin marketplace add thienanblog/awesome-ai-agent-skills

Step 2: Install skills

# Install a plugin (can bundle multiple skills)
/plugin install project-development-skills@awesome-ai-agent-skills

# Install Laravel guidelines
/plugin install laravel-app-skills@awesome-ai-agent-skills

# Install Docker local development skill
/plugin install devops-skills@awesome-ai-agent-skills

# Install office web UI skill
/plugin install office-web-ui-skills@awesome-ai-agent-skills

Updating the marketplace

/plugin marketplace update

Claude Desktop app

The desktop app’s plugin browser (+ → Plugins → Add plugin) only lists marketplaces you have already configured — it cannot add one. Register this marketplace first from a terminal:

claude plugin marketplace add thienanblog/awesome-ai-agent-skills

Then reopen the plugin browser. If any marketplace command fails with JSON Parse error: Unexpected EOF, your local ~/.claude/plugins/known_marketplaces.json is corrupt; delete that file and retry.

Layout

Each marketplace entry points at a self-contained plugin package under plugins//, which carries its own .claude-plugin/plugin.json and bundles only that plugin’s skills. Claude Code copies just that package into its plugin cache and pins it to the released version. See the official Claude Code docs for plugin marketplaces and plugin structure.

The manifests deliberately stay on the schema keys that every current Claude Code release accepts. Newer-only keys (top-level description/version, per-plugin displayName, $schema, renames) are rejected as unrecognized by older clients, which makes the whole marketplace fail to load rather than degrade. scripts/lib/plugin-shape.js is the single definition of every generated shape, and npm run validate regenerates and diffs each file, so any stray key fails the build.

OpenAI Codex

Add this repository as a Codex marketplace:

codex plugin marketplace add thienanblog/awesome-ai-agent-skills

Then install the plugin you want:

codex
/plugins

In the plugin browser:

1. Choose the Awesome AI Agent Skills marketplace.
2. Open one of these plugins:
   • project-development-skills
   • laravel-app-skills
   • devops-skills
   • office-web-ui-skills
3. Select Install plugin.
4. Start a new thread and ask Codex normally, or type @ to choose the plugin or one of its bundled skills explicitly.

To refresh after this repository updates:

codex plugin marketplace upgrade awesome-ai-agent-skills

Codex also supports installing from a local checkout while developing this repository:

git clone https://github.com/thienanblog/awesome-ai-agent-skills.git
cd awesome-ai-agent-skills
codex plugin marketplace add .

This repository includes a Codex-compatible marketplace at .agents/plugins/marketplace.json and plugin packages under plugins/. The layout follows OpenAI’s docs: marketplace entries point at ./plugins/, plugin manifests live in .codex-plugin/plugin.json, and bundled skills live inside the plugin root. See OpenAI’s Plugins and Build plugins docs.

The repeated skill folders under plugins//skills/ are generated package copies shared by both agents. If they differ from skills//, edit the canonical skill folder and rerun npm run sync.

Skills CLI

The open skills CLI works with Codex, Claude Code, Cursor, and many other agents.

On macOS:

brew install skills
skills add thienanblog/awesome-ai-agent-skills --list
skills add thienanblog/awesome-ai-agent-skills --skill project-development-mindset

Without installing globally:

npx skills add thienanblog/awesome-ai-agent-skills --list
npx skills add thienanblog/awesome-ai-agent-skills --skill project-development-mindset
npx skills init my-skill

You can also copy individual skill instructions directly into your AI agent’s context or system prompt.

Usage Examples

Choose the skill that fits the task. Use the general project workflow when the work spans several responsibilities or the right starting point is unclear:

Use $project-development-mindset to implement the approved checkout redesign and verify the changed behavior.

Specialists can be used directly. Routine tests, documentation updates, framework commands, and browser checks stay within the implementation task; they do not require a routing ceremony or a fixed number of loaded skills.

Use $debugging-workflow to isolate and fix this intermittent checkout failure.
Use $ui-ux-concept-implementation to implement Concept B for this pricing page.

For a decision before implementation:

Use $brainstorm-first to compare three checkout redesign concepts, recommend one, and wait for my selection.

Brainstorming honors the requested option count and decision boundary. If you ask the agent to choose and implement, it can continue within that authorization. Reviewable delivery remains opt-in:

Use $run-reviewable-subtask-loop to deliver this migration as coherent reviewed commits and one aggregate PR.

All bundled skills run in the main conversation by default. Delegation can increase usage and requires explicit approval for the proposed agent count and scope. That approval is reused within its bounds; expanding the count or scope requires fresh approval. A request for subtasks does not authorize subagents.

Instruction design

The skills focus on task-specific decisions, evidence, and real operating constraints. They reuse existing authorization, ask only about material unresolved choices, and scale verification to changed behavior and repository requirements. They avoid fixed questionnaires, invented confidence scores, mandatory full-suite offers, and repeated workflow handoffs.

The September 2026 revision was informed by OpenAI’s GPT-6 Astra prompting guidance. The content remains tool-agnostic: it neither selects a model nor assumes that stronger models make verification or data/production boundaries unnecessary. See the audit and review notes for per-skill changes, compatibility notes, and the limits of validation.

Available Skills

Plugin Groups

Plugins bundle related skills so you can install by domain. The source of truth is plugin-groups.json.

Repository Cleanup

This repository has been narrowed to a smaller, cohesive set of skills that are intended to work together. Apologies to contributors whose community skills were removed during this cleanup; the goal is to keep this repository focused on quality-controlled project development workflows instead of hosting unrelated skill experiments.

Contributing

We welcome contributions! Here’s a quick start:

1. Fork this repository
2. Create a skill folder: skills/your-skill-name/
3. Add a SKILL.md with metadata:

   ---
   name: your-skill-name
   description: What the skill does and when to use it.
   ---
   

4. Add the skill to plugin-groups.json so it belongs to exactly one plugin.
5. Sync and validate locally before pushing:

   npm install
   npm run sync
   npm run validate
   npm run test:agent-context
   

6. Submit a pull request

See CONTRIBUTING.md for detailed guidelines, validation instructions, and troubleshooting.

Validation Workflow

• plugin-groups.json is the source of truth for plugin membership.
• npm run sync regenerates .claude-plugin/marketplace.json, .agents/plugins/marketplace.json, plugins/**, and the generated tables in README.md.
• npm run validate checks skill metadata, plugin assignments, release-version parity, generated manifests, and complete bundled-skill parity with canonical sources.
• npm run test:agent-context runs regression coverage for deterministic framework/tool discovery, precedence, symlink safety, and native fallback depth.
• Bump version in package.json before syncing a release; it is stamped into every plugin entry, and users only receive an update when it changes.
• Pull request CI reruns npm run sync and fails if generated files are out of date.

For AI Agents

See CLAUDE.md for instructions on how to work with this repository, including how to group skills into plugins and update the marketplace when new skills are added.

Compatibility

This skill format is designed to be universal and works with:

• Claude Code (Anthropic)
• OpenAI Codex
• Cursor
• Kilo Code
• GitHub Copilot
• Windsurf
• Any AI coding assistant that supports custom instructions or skills

License

MIT