One-line prompts.
Shipped products.
23 skill categories. 28 agents. 176 reference docs.
The autonomous product-building OS for 32+ AI coding tools.
claude plugin install heymegabyte/agent-skills
Loads only what you need
A flat rules file dumps everything into context. Agent Skills routes your prompt to the smallest useful subset — 01 Operating System always loads first, then only the categories your task touches.
Agent Skills uses a selective router: it loads only the skill categories and reference docs relevant to your prompt instead of the whole system. A debug task might pull 3 of 23 categories; a full SaaS build pulls 8 — keeping context usage low so the AI stays sharp on your actual work.
How It Works
From idea to deployed product in three steps.
Agent Skills turns a single natural language prompt into a deployed, tested product. The system handles architecture, code generation, testing across 6 breakpoints, deployment to Cloudflare Workers, and visual verification — all without human intervention.
Install
claude plugin install heymegabyte/agent-skills
Drop into Claude Code, Cursor, Copilot, or any of 32 supported tools.
Prompt
"Build a SaaS at projectsites.dev"
One line. The system infers product type, stack, and architecture automatically.
Ship
✓ Deployed to Cloudflare Workers
Architect → build → test → deploy → verify. Fully autonomous. Zero hand-holding.
19 Skill Categories
Each category contains specialized submodules and reference docs. The router loads the smallest useful subset for every prompt.
The 23 skill categories cover the full product lifecycle: from goal definition and architecture planning through build, test, deploy, brand, design, motion, media, growth, site generation, document processing, and MCP authoring. Each category contains specialized submodules with 176 reference docs total, loaded on-demand to minimize context window usage.
Operating System
Supreme policy layer. Autonomy, speed standards, cross-skill coordination, conflict resolution.
6 reference docsGoal & Brief
Project thesis before code. Product type inference, user identification, business model.
Core skillPlanning & Research
Deep web research, competitor scanning, tech evaluation, vertical slice decomposition.
1 reference docPreference & Memory
User preferences with confidence levels. Voice of the Customer model, auto memory system.
3 reference docsArchitecture & Stack
Cloudflare-first platform selection. Workers, D1, R2, KV, Durable Objects, Queues, Vectorize.
11 reference docsBuild & Slice Loop
Vertical slices, homepage-first. Anti-placeholder rules — real content, real interactions.
23 reference docsQuality & Verification
5-level verification pyramid. Playwright E2E, AI visual QA, agentic security testing.
24 reference docsDeploy & Runtime
Deploy after every change. Typecheck → deploy → purge → E2E → visual verify → fix-forward.
9 reference docsBrand & Content
Copy system, headline rules, trust surfaces, SEO, structured data, anti-AI-slop filters.
4 reference docsExperience & Design
Anti-AI-slop design system. Dark-first, bold typography, cascade layers, container queries.
1 reference docMotion & Interaction
Meaning-first animation. Scroll-driven, View Transitions API, prefers-reduced-motion mandatory.
Core skillMedia Orchestration
Image generation, logo/icon sets, video, social previews, OG images, compression pipelines.
5 reference docsObservability & Growth
PostHog analytics, GA4 via GTM, Sentry, Stripe billing, feature flags, conversion optimization.
7 reference docsIdea Engine
Autonomous internal co-founder. Evidence-backed improvements, self-critique filter, auto-implement.
Core skillSite Generation
Research-saturated, maximalist site builds. Competitor-beating, AI-native, deploy-verified.
16 reference docsCinematic Prime Directive
One-line prompt to 100 build-breaking rules across 10 quality categories before DONE.
Core skillNon-Engineering Verticals
Enterprise credibility: C-suite personas, compliance, product management, finance.
4 reference docsDocument Processing
Worker-native document I/O. PDF, DOCX, XLSX, PPTX generated at the edge.
4 reference docsMCP Authoring
Build and harden MCP servers — the standard structured tool surface for agents.
3 reference docs28 Specialized Agents
Each agent has a defined model, permission mode, and skill set. The orchestrator spawns them in parallel.
Agent Skills uses 28 specialized agents — each assigned to an optimal model tier (Opus for architecture/security, Sonnet for implementation/testing, Haiku for content/linting). The meta-orchestrator spawns agents in parallel, coordinating architect, test-writer, and deploy-verifier simultaneously to ship faster.
Architect
OpusAnalyzes project structure, generates repo maps, designs task graphs, identifies architectural seams.
Meta Orchestrator
OpusMaster agent. Knows every tool, MCP, and skill. Plans multi-system workflows, chains MCPs, spawns agents.
Completeness Checker
OpusFeature Completeness Engine, Zero Recommendations Gate, visual verification. Nothing slips through.
Security Reviewer
OpusOWASP Top 10, secrets exposure, injection flaws, auth bypasses, CSP issues. Read-only — never edits code.
Visual QA
OpusScreenshots every breakpoint. AI vision detects layout breaks, misalignment, text overflow, broken images.
Performance Profiler
OpusRuns Lighthouse, analyzes Core Web Vitals, suggests specific fixes with file:line references.
Incident Responder
OpusSentry-triggered. Reads error events, traces to source, bisects deploys, proposes rollback or patch.
Computer Use Operator
OpusDesktop automation via Computer Use MCP. Finder, System Settings, native apps, cross-app workflows.
Deploy Verifier
SonnetPost-deploy smoke tests. Console errors, screenshots 6 breakpoints, axe-core, SEO validation.
Test Writer
SonnetTDD-first. Writes failing Playwright E2E tests emulating real users before implementation.
SEO Auditor
SonnetTitle, meta, H1, JSON-LD, OG tags, internal links, sitemap, robots.txt. Playwright-powered live audits.
Accessibility Auditor
SonnetWCAG 2.2 AA. axe-core across 6 breakpoints, keyboard + screen-reader flows, contrast, focus order.
Migration Agent
SonnetDrizzle schema migrations. Generates from diffs, validates against D1, tests rollbacks, zero-downtime.
Content Writer
SonnetMarketing copy, blog posts, SEO content. Brand voice, Flesch≥60, active voice, zero AI slop.
Media Orchestrator
SonnetCreates and optimizes media — images (AVIF/WebP), video, OG cards — and orchestrates the pipeline.
Motion Choreographer
SonnetView Transitions API, scroll-driven animation, and brand-locked micro-interactions.
Browser Operator
SonnetDrives a real browser to verify deployments and operate the product like a user would.
Resource Broker
SonnetMaintains a normalized registry of legitimately-owned account resources — credits, quotas, keys.
Code Simplifier
HaikuSimplifies code for clarity, consistency, and maintainability while preserving all functionality.
Dependency Auditor
HaikuScans for outdated deps, security advisories, license violations, unused imports. Prioritized reports.
Cost Estimator
HaikuReads wrangler.toml and projects Workers/D1/R2 spend per feature before you deploy.
Changelog Generator
HaikuConventional commits → gitmoji changelog, semantic-release version bumps, readable release notes.
Changelog Drafter
HaikuReads git log since last tag and drafts the CHANGELOG entry, grouped by conventional-commit type.
Model Router
HaikuRecommends the right model tier for each dispatched task — Opus, Sonnet, or Haiku.
Formatter
HaikuRuns prettier + oxlint --fix on modified files for consistent formatting and lint compliance.
Dead Code Remover
HaikuFinds unreachable exports, unused imports, orphaned functions, and dead variables — then prunes them.
Renamer
HaikuSemantic rename across the codebase. Greps all references, updates imports and usages safely.
Transcriber
HaikuTranscribes a plan step into code, one file at a time, straight from the spec.
Built for the Modern Stack
Opinionated defaults, escape hatches when you need them.
32 Platform Variants
One skill system. Every AI coding tool. Auto-generated convention files for each platform's format.
Modern Formats
Legacy Single-File
Named Formats
Directory Formats
How skillsl.ink compares
We benchmark the best agent-skills profiles on GitHub against six dimensions where a routed, cross-tool system is built to lead. Star counts are live from GitHub; this list keeps growing as our loop finds new champions.
Most AI-coding profiles are flat rule dumps scoped to one tool. skillsl.ink scores highest on a six-dimension champion rubric — breadth, maintenance, multi-tool portability, routing/context-efficiency, validation, and published results — because it routes 23 skill categories and 28 agents into native config for 32+ AI coding tools from a single install.
| Profile | Stars | Breadth | Maintenance | Multi-tool | Routing | Validation | Results | Champion |
|---|---|---|---|---|---|---|---|---|
| skillsl.ink · Agent Skills yourouted multi-tool OS |  |
10 | 9 | 10 | 10 | 9 | 8 | 93 |
| superpowersengineering-culture skills |  |
8 | 9 | 9 | 9 | 9 | 8 | 87 |
| anthropics/skillsofficial reference |  |
7 | 9 | 6 | 10 | 8 | 7 | 78 |
| addyosmani/agent-skillsprocess-over-prose |  |
7 | 8 | 7 | 8 | 9 | 7 | 77 |
| TerminalSkills/skillsgeneral dev library |  |
7 | 8 | 9 | 9 | 6 | 6 | 75 |
| mattpocock/skillscomposable skills |  |
6 | 8 | 9 | 7 | 8 | 6 | 73 |
| veniceai/skillsAPI-surface skills |  |
6 | 9 | 8 | 8 | 7 | 6 | 73 |
| publora/skillssocial-publishing skills |  |
5 | 9 | 8 | 7 | 8 | 7 | 73 |
| jeffallan/claude-skillsfull-stack pack |  |
9 | 7 | 6 | 8 | 6 | 5 | 68 |
| bregman-arie/devops-sre-skillsDevOps / SRE runbooks |  |
6 | 6 | 7 | 8 | 8 | 6 | 68 |
| vercel-labs/agent-skillsfrontend skills |  |
5 | 8 | 6 | 7 | 7 | 7 | 67 |
| trailofbits/skillssecurity skills |  |
5 | 7 | 7 | 6 | 9 | 6 | 67 |
| uxuiprinciples/agent-skillsUX evaluation skills |  |
4 | 7 | 7 | 6 | 6 | 6 | 60 |
| wondel/skillsUX & product skills |  |
6 | 7 | 6 | 6 | 5 | 5 | 58 |
Scored 14 profiles · champion threshold 85/100 · updated 2026-10-26. See the methodology →
Get Started
Pick your tool. Drop in the skills. Start shipping.
GitHub Skills
gh skill install heymegabyte/agent-skillsClaude Code Plugin
claude plugin install heymegabyte/agent-skillsnpm
npm i @heymegabyte/agent-skillsJSR
jsr add @heymegabyte/agent-skillsOpenAI Codex
git clone https://github.com/heymegabyte/agent-skills \
~/.codex/skills/claude-skillsManual (any tool)
git clone https://github.com/heymegabyte/agent-skills
# Copy the convention file for your toolBefore & After
What changes when you add 14 years of engineering judgment to your AI.
Without skills, AI coding tools produce generic boilerplate with no tests, no deploy pipeline, and placeholder content. With Agent Skills, the same prompt produces a Cloudflare-deployed product with Playwright E2E tests across 6 breakpoints, Lighthouse 95+ accessibility, real content, and zero manual fixups.
Without Skills
With Agent Skills
How Skills Flow
The router loads the smallest useful subset. No wasted context.
FAQ
Does this work with Cursor, Copilot, and Windsurf?
Yes. Agent Skills generates 32 platform-specific convention files. When you install it, your tool automatically picks up the right format — MDC rules for Cursor, copilot-instructions.md for Copilot, .windsurfrules for Windsurf, and 29 other variants.
Is this different from .cursorrules files?
Fundamentally. A .cursorrules file is a flat text dump. Agent Skills is a 23-category system with 176 reference docs, intelligent routing (load only what you need), and 28 specialized agents. It generates .cursorrules as one of 32 outputs.
What if I don't use the default stack?
The system has escape hatches for every choice. Don't use Angular? The frontend skill adapts. Don't use Cloudflare? The deploy skill handles other platforms. The architecture decisions are opinionated defaults, not locked dependencies.
How do the 28 agents work?
Each agent has a defined model (Opus, Sonnet, or Haiku), permission mode, and skill set. The meta-orchestrator spawns them in parallel — architect designs the structure, test-writer creates failing tests, deploy-verifier checks production. They coordinate automatically.
Is this free?
Yes. Open source, free forever. The skills are published on npm, JSR, and GitHub. You bring your own AI tool subscription — Agent Skills just makes it dramatically more effective.
How do I create custom skills?
Follow the SKILL.md format with frontmatter (name, description, submodules). Place it in the numbered directory structure. The router will pick it up automatically. See the CONVENTIONS.md for the full spec.
How does the skill router minimize context usage?
The router analyzes your prompt and loads only the relevant skill categories and their submodules. A billing feature loads skills 05 (Architecture) and 13 (Observability) but skips 11 (Motion) and 12 (Media). This keeps context usage under 20% of the available window for most prompts.
Does Agent Skills work with OpenAI Codex CLI?
Yes. Agent Skills generates a CODEX.md file and supports the ~/.codex/skills/ directory format. Install with: git clone https://github.com/heymegabyte/agent-skills ~/.codex/skills/claude-skills. The same skill content works across both Claude and OpenAI tools.
What verification does the system run after deployment?
Five layers: TypeScript compilation check, Playwright E2E tests across 6 breakpoints (375px to 1920px), axe-core accessibility audit (zero violations required), AI visual QA via GPT-4o screenshot analysis, and Lighthouse scoring (accessibility ≥95, performance ≥75). Deploy fails trigger automatic fix-forward cycles.
Can I use this for existing projects, not just new ones?
Absolutely. Drop the skills into any repo and the system adapts to your existing stack, patterns, and conventions. It reads your codebase, respects your architecture, and applies its quality standards to new features while preserving what's already there.
How is this different from a SaaS template or boilerplate?
Templates give you starting code. Agent Skills gives you an engineering brain. It makes real-time decisions about architecture, generates tests before code (TDD), deploys and verifies automatically, writes production copy (not "Lorem ipsum"), and iterates until zero recommendations remain. It's a system, not a scaffold.
What security checks are built in?
The Security Reviewer agent scans for OWASP Top 10 vulnerabilities, hardcoded secrets, injection flaws, auth bypasses, and CSP issues. Every form gets Turnstile protection. Every database query uses parameterized Drizzle ORM. detect-secrets scan runs on every commit via pre-commit hooks.