We evaluated over 100 GitHub repositories focused on using AI for non-coding work — GTM, marketing, knowledge management, content production — and filtered it down to the best 63, ranked by use case, value, and quantitative signals like stars, forks, commit velocity, and community health.
Last updated: April 2026 · Next refresh: July 2026
The repos GTM operators actually need — not prompt libraries, not demos. Agent configurations, workflow templates, and orchestration patterns for revenue teams.
The largest single collection of Claude Code agents, organized as focused plugins with progressive disclosure (register in marketplace, load on demand). Primarily technical — 25 categories spanning development, infrastructure, security, AI/ML, and data. GTM coverage exists but is thinner than the dev side. The 16 multi-agent orchestrators and 7 team presets (review, debug, fullstack, security, migration) are where the coordination patterns shine. Three-tier model allocation (Opus/Sonnet/Haiku by task criticality) keeps costs manageable at scale.
184 agents across 78 plugins — primarily development and infrastructure, with lighter GTM coverage. The plugin architecture (load only what you need) and 7 Agent Team presets (review, debug, fullstack, security) are the real value.
A Claude Code plugin implementing BMAD (Business Model Agile Development) methodology with role-based agents for each development phase. Three workflow paths offer progressive complexity: full agile with 6 specialized agents, lightweight requirements-driven, or direct development commands. Quality gates enforce 90% thresholds with iterative refinement. Persistent specs in `.claude/specs/` survive sessions. Note: this is Claude Code-specific, not multi-tool — the Go codebase builds Claude Code plugins, not cross-platform orchestration.
BMAD methodology (Business Model Agile Development) with three workflow paths: full agile, lightweight requirements, or direct commands. Quality gates enforce 90% thresholds. Single-platform, not multi-tool.
The deepest GTM-specific agent collection: 92 agents, 263 skills across five domains (Sales, Marketing, CS, RevOps, Growth). Agent specialization goes granular — sales-prospecting splits into cold-outreach, social-selling, lead-qualification, discovery-calls, and objection-handling. 19 dedicated orchestrator plugins coordinate multi-agent sequences (campaign-launch chains strategist → designer → copywriter → email-marketer → analytics). Model allocation by role: Haiku for routine data, Sonnet for strategy. Business-first UX: describe what you need, skip the prompt engineering.
92 GTM-specific agents across 67 plugins covering Sales, Marketing, CS, RevOps, and Growth — with 19 orchestrators that coordinate multi-agent sequences. The specialization depth is real: sales-prospecting alone has 5 discrete agents.
A complete GTM workspace template with 5 mandatory context files (company profile, ICP with anti-ICP filters, signal library with decay logic and combination rules, positioning with exclusions, competitor battlecards with switch scenarios), 5 executable skills (setup auto-populates from domain research, account research, signal-to-sequence campaigns, ICP scoring, weekly maintenance), and a mature example showing what data-backed signal management looks like. The architecture cleanly separates context (knowledge), skills (execution), and workflows (human processes).
Fork this, run the setup skill with your domain, and Claude auto-populates your GTM workspace in 30 minutes — ICP, signals, positioning, competitive intel, and buyer personas. The signal library with decay logic is more sophisticated than most enterprise intent platforms.
02
Marketing & Content Production
Marketing skills that Claude actually executes — not prompt templates you paste into a chat window.
25K stars for a marketing skill library built on a foundational context layer — `product-marketing-context` loads your product, audience, and positioning before any skill executes. 7 categories: SEO (traditional + AI-SEO + programmatic), CRO (6 skills split by conversion moment), Content & Copy, Paid & Measurement, Growth & Retention (including churn prevention and referral programs), Sales & GTM, and Strategy. Skills cross-reference each other (copywriting ↔ page-cro ↔ ab-test) for natural chaining.
38+ marketing skills with a foundational context layer every skill references. The CRO coverage is the standout: 6 skills split by conversion moment (page, signup, onboarding, form, popup, paywall). AI-SEO for LLM citation is forward-looking.
A marketing automation suite built on parallel subagent execution: launch 5 agents simultaneously for content analysis, CRO, competitive positioning, SEO, and growth strategy — each returns independent scores aggregated into an overall marketing assessment. 14 sub-skills cover the full marketing workflow from email sequences (5-8 templates per flow) to 30-day social calendars and ad operations. PDF report generation via Python makes output client-ready.
15 skills with 5 parallel subagents that run concurrent marketing audits — content, CRO, SEO, competitive, growth — aggregated into a single score with client-ready PDF reports. The email sequence generator (5-8 templates per flow) is production-ready.
23 free + 10 premium marketing skills across 5 categories, each shipping as dual files (verbose + condensed for token efficiency) with three execution modes (quick, standard, deep). Built on named marketing authorities rather than generic advice. Standout skills: AI Visibility (how brands appear in AI search), voice extraction (capture authentic writing style), and de-AI-ification (restore human voice, built from analyzing 1,000+ content pieces). Cross-platform: Claude Code, OpenClaw, Cursor, Windsurf.
Marketing frameworks built on named masters (Ogilvy, Godin, Drucker) with a dual-file system: full SKILL.md for deep work, condensed SKILL-OC.md (~200 lines) for token efficiency. Three execution modes (quick/standard/deep) per skill.
03
Knowledge Management & Productivity
Repos that turn Claude Code into a knowledge operating system — personal knowledge management, research workflows, and productivity automation.
Comprehensive agent harness with 183 skills, confidence-weighted instincts that auto-extract and evolve into skills, AgentShield security (1,282 tests, 102 rules, 3-agent adversarial pipeline), and a universal AGENTS.md that serves Claude Code, Cursor, Codex, and OpenCode simultaneously. Manifest-driven selective installation lets you pick components instead of all-or-nothing. Strategic compaction at logical breakpoints (not just 95% utilization). The scope is deliberately ambitious — this is trying to be the foundation layer for everything else.
167K stars — the most-starred Claude Code repo. The instincts system (auto-extract patterns from sessions, confidence-score them, prune after 30 days, evolve into formal skills) is the standout innovation. AgentShield's 3-agent security pipeline (red team/blue team/auditor) adds enterprise-grade protection.
5-phase conversion pipeline: detect source format (llms.txt → sitemap → headless browser), scrape content, categorize by topic, enhance with AI analysis, package for distribution. Supports 18 input formats and exports to 20 platforms (12 LLM + 8 RAG/vector). The doc-code conflict detector identifies documented features missing from code, undocumented implementations, and signature mismatches. MCP server with 40+ tools lets you use the converter as a tool within Claude Code. Sync monitoring tracks when source docs change and skills need updating.
Turn any of 18 documentation formats (websites, GitHub repos, PDFs, YouTube, Confluence, Notion, Slack exports, OpenAPI specs) into Claude Code skills automatically — then export to 20 platforms in one pass. The doc-code conflict detection (finds mismatches between docs and actual implementation) is the hidden gem.
Anthropic's own plugins for knowledge workers. 11 job-function packages (Sales, Legal, Finance, PM, Marketing, CS, Data, Enterprise Search, Bio-Research, Productivity, Plugin Management) each bundling skills + SaaS connectors + slash commands + sub-agents. The Legal plugin (contract review, NDA triage, compliance, risk assessment) and Finance plugin show serious vertical depth. File-based customization — markdown and JSON only, no code required.
Official Anthropic plugins for knowledge workers — 11 plugins across Sales, Legal, Finance, PM, Marketing, CS, Data, and more. SaaS connectors (Slack, Jira, Notion, Asana, Monday, Microsoft 365) make these production-ready. Works in both Cowork and Claude Code.
Semi-automated research workflow with 40+ skills supporting the full lifecycle from ideation to publication. Integrates Zotero for literature management and Obsidian for project knowledge bases. Specialized agents — paper-miner extracts writing patterns, literature-reviewer synthesizes research directions, rebuttal-writer structures reviewer responses. Self-improvement skills iterate on skill quality. Cross-platform: Claude Code (primary), OpenCode, Codex CLI with dedicated branches.
40+ skills across a 7-stage research pipeline: ideation → literature review (Zotero) → paper notes → knowledge base (Obsidian) → experiments → analysis → publication. Human-centric: you decide what to research, AI handles the structure.
A complete starter kit connecting Obsidian's knowledge graph to Claude Code's agent capabilities. 10 skills covering daily/weekly/monthly review cadences, project management, and vault operations. 4 memory agents learn your patterns over time — the weekly-reviewer learns which reflection questions work for you, the inbox-processor learns your categorization preferences. Auto-commit on every file edit creates full version history. The /adopt skill integrates existing vaults without restructuring.
If you use Obsidian and Claude Code, this is where you start. 10 skills, 4 specialized memory agents (goal-aligner, weekly-reviewer, note-organizer, inbox-processor), auto-commit on every edit. Session startup surfaces your ONE Big Thing — forces focus before you start working.
04
Non-Technical & Specialized Domains
Proof that Claude Code works far beyond software development — academic workflows, financial analysis, data science, and more.
Anthropic's own financial services plugin ecosystem. Core plugin (Financial Analysis, 41 shared skills) is mandatory; add-ons for Investment Banking, Equity Research, Private Equity, and Wealth Management layer function-specific workflows. 11 MCP data integrations with institutional providers (Daloopa, Morningstar, S&P Global, FactSet, Moody's, MT Newswires, Aiera, LSEG, PitchBook, Chronograph, Egnyte). End-to-end pipelines: research-to-report, SEC filing to spreadsheet model, deal material drafting, portfolio-to-presentation. Slash commands: /comps, /dcf, /earnings, /one-pager, /ic-memo, /source, /client-review.
Anthropic's official finance plugins — 5 packages covering IB, equity research, PE, and wealth management. The 11 MCP integrations with institutional data providers (FactSet, S&P Global, Moody's, PitchBook) make this production-grade. Install the core + your function's add-on.
The densest single-vertical skill collection for trading. 80+ skills across 7 categories: Market Analysis, Economic/Earnings Calendars, Strategy/Risk Management, Earnings Momentum, Stock Screening, Market Timing, and Skill Quality automation. The standout pattern: a self-improving ecosystem where daily automated review scores one skill (dual-axis, 0-100), applies improvements if below 90, and opens a PR. Weekly, session logs are mined for new skill ideas — highest-ranked pending ideas get built by a Skill Designer agent, validated by the reviewer, and merged. Knowledge-only skills operate offline from reference materials alone. Versioned interface contracts (strategy.yaml, metadata.json) enable reproducible backtesting integration.
80+ trading skills with a self-improvement loop — daily automated quality review, weekly new skill generation from session logs. The skills breed new skills. The dual-axis reviewer (scoring < 90 triggers Claude-applied fixes + PR) is a pattern every skill library should steal.
A production-grade academic workflow template from Emory University's Econ 730 course. 14 specialized agents each review one dimension (proofreader, slide-auditor, pedagogy-reviewer, R-reviewer, domain-reviewer). The adversarial QA loop (critic + fixer, max 5 rounds) catches errors single-pass review misses. Quality scores (0-100) drive automation: 80 = auto-commit, 90 = PR, 95 = excellence. PreCompact hooks preserve context across compression. Fork it, fill the knowledge-base template with your domain, and the orchestrator/quality-gates transfer without modification. Skills: /create-lecture, /data-analysis, /review-paper --peer, /qa-quarto, /seven-pass-review.
The most sophisticated non-dev Claude Code workflow published. 28 skills, 14 single-dimension review agents, adversarial QA loops, and quality-gated auto-commits. 15+ research groups across economics, energy, and engineering have forked it successfully. The 2:1 fork ratio says it all.
A full-stack AI investment research platform built on LangGraph, not a Claude Code skill package. The standout innovation: Programmatic Tool Calling (PTC) — the agent writes Python code that executes in Daytona cloud sandboxes, processes financial data locally, and returns only final results, dramatically reducing token consumption. Persistent workspaces compound research Bayesian-style over weeks/months via agent.md cross-session memory. 23 pre-built skills covering DCF, comps, 3-statement analysis, earnings, and market intelligence. Price-triggered automations fire on real-time WebSocket price feeds with Redis deduplication. 25-layer middleware stack handles live steering, auto-compaction, and credential leak detection. Slack/Discord/Telegram/WhatsApp channel integrations. Apache 2.0.
Full-stack financial research platform, not a Claude Code skill package — LangGraph agents, FastAPI backend, React frontend. The Programmatic Tool Calling pattern (agent writes code in sandboxes, returns only results) and price-triggered automations are genuinely novel. Heavy infrastructure, but serious capability.
Institutional-grade equity research reports from a single /trading-ideas:research command. Analysis framework covers both quantitative (financial statements, valuation models, technicals, options analytics) and qualitative (competitive positioning, management quality, regulatory, ESG) dimensions. Reports include executive summaries with 12-month price targets, catalyst identification with timelines, bull/base/bear scenarios with probability allocations, and position-sizing recommendations (1-5% allocation guidance). Sector rotation intelligence and insider activity tracking round out the toolkit.
Goldman Sachs-style equity research via a single slash command. Deep on one vertical — 12-month price targets, bull/base/bear scenarios with probability weighting, catalyst timelines, options flow, ESG. Not broad, but serious depth.
The largest collection of non-development Claude Code workspace templates. 28 cluster plugins ship domain primitives (slash commands, skills, scaffolds) across legal research, health documentation, financial planning, career management, and more. Each workspace follows the same skeleton: CLAUDE.md + folder structure + MCP configs + subagent definitions. The 350+ slash command library and MCP integration layer show infrastructure maturity. Philosophy: Claude Code as general-purpose agent workspace, not coding assistant.
75+ workspace templates across 15+ non-dev domains — legal, health, finance, career, audio, sysadmin. The 28 cluster plugins that consolidate these into installable packages are more interesting than any individual template. Breadth play, not depth.
A product analytics platform built as Claude Code skills. 18 agents execute in a 4-phase DAG (Frame > Analyze > Story > Deck) with parallel execution where dependencies allow and checkpoint resume on interruption. Multi-source data support (CSV, DuckDB, Postgres, BigQuery, Snowflake). The self-learning .knowledge/ system captures corrections, proven SQL patterns, and business glossary to prevent methodology drift. 41 skills (17 always-active, 19 command-triggered). Charts follow Storytelling with Data methodology with WCAG-compliant palettes. Claims to automate 80% of analyst work.
18 agents in a DAG pipeline that turns business questions into validated slide decks. The 4-layer validation (including Simpson's Paradox checks) and self-learning correction loop show real analytical rigor. Smaller community but serious architecture.
05
Skills Libraries & Ecosystems
Where to find installable skills — from official Anthropic collections to community mega-repos.
Creative, technical, and enterprise workflow skills from Anthropic. THE canonical source — everything else downstream builds on this spec. The document manipulation skills (DOCX, PDF, PPTX, XLSX) are the sleeper value: most operators discover the creative skills first but come back for the document automation. Start here to understand what a well-structured skill looks like.
The official Anthropic skills repository. 124K stars. The reference implementation — but the real value is the document skills (DOCX/PDF/PPTX/XLSX) and the formal Agent Skills spec that everything else builds on.
The largest community-curated skills directory with 180+ skills across 12 functional categories. The Composio SaaS integration section (78 pre-built connections to CRM, email, DevOps, social, and more) is often overlooked. The first place most new users land when looking for skills to install.
56K stars makes this the most-starred skills list. Good categorization and the 78 Composio SaaS integrations are the hidden gem — but it's a directory, not a toolkit.
Official dev team contributions mixed with community skills. Organized by maintainer rather than function — so you discover skills by trust level (Stripe's commerce skills, Trail of Bits' security audits, Datadog's observability). The scale is real at 1,100+ and the 'hand-picked, not AI-slop' curation holds up.
1,100+ skills from official teams and community, organized by maintainer (Anthropic, Google, Stripe, Trail of Bits, Datadog). Hand-curated — the npm of Claude Code skills.
Engineering, marketing, product, compliance, C-level advisory, and regulatory skills (ISO/FDA/GDPR/CAPA) across 9 domains. The standout isn't the skill count — it's the architecture: named orchestration patterns, three persona profiles that prioritize skills by role, a security auditor that scans skills before installation, and a conversion script that ports everything to Cursor, Aider, Windsurf, and 9 other tools. Zero pip installs.
235 skills across 9 domains with 305 zero-dependency Python tools. The cross-platform conversion script (12 AI tools in 15 seconds) and named orchestration patterns (Solo Sprint, Multi-Agent Handoff, Skill Chain) make this more than a collection — it's an operating model.
Skills list specifically focused on Claude Code with a comparative decision framework explaining when to use Skills vs. MCP vs. Prompts vs. Projects vs. Subagents. The Claude-specific focus (unlike vendor-agnostic alternatives) means direct API integration paths and security guidance. Updated monthly with a detailed changelog.
Well-organized skills directory focused specifically on Claude Code. The comparative table (Skills vs. MCP vs. Prompts vs. Subagents) is what most beginners need before browsing skills.
A three-layer architecture (Skills → Commands → Plugins) that encodes real PM frameworks from specific practitioners (Torres, Cagan, Savoia, Ulwick). Commands chain skills into guided workflows with next-step suggestions — run `/discover` and it flows through assumption mapping, opportunity trees, and experiment design. The cross-functional perspective (PM + Designer + Engineer viewpoints in parallel) is the kind of structural choice most skill repos skip.
65 PM skills in 36 chained workflows across 8 plugins — from Teresa Torres' Opportunity Solution Trees to pre-mortem risk classification. Framework-native, not prompt-generic.
Agile pipeline, multi-model AI review, documentation generation, plus three custom MCP servers. hex-line-mcp hash-verifies every line edit to prevent stale-context corruption in long sessions. hex-graph-mcp builds a deterministic code knowledge graph with framework-aware overlays. The four-level orchestrator hierarchy (meta-orchestrator → coordinator → stateful coordinator → worker) with resumable checkpoints is the most sophisticated execution model in this list. Host-aware advisor routing automatically delegates reviews to a different model.
135 skills across 7 plugins with 3 bundled MCP servers. The real innovation: hash-verified editing (prevents stale-context corruption) and a four-level orchestrator hierarchy with resumable checkpoints.
06
MCP Server Collections
Model Context Protocol servers extend what Claude Code can access. These directories help you find the right one.
The reference MCP server implementation from GitHub with both remote (hosted by GitHub) and local deployment options. Covers repos, issues, PRs, Actions, code search, and code analysis. Pre-built install scripts for major IDEs mean setup takes 2 minutes.
GitHub's official MCP server. 29K stars. Remote (hosted) and local modes, OAuth + PAT auth, pre-built install for VS Code, Claude Desktop, Cursor, Windsurf, Codex. If you're using Claude Code with GitHub, this is table stakes.
Comprehensive categorized list of Model Context Protocol servers. THE starting point for MCP discovery. Slightly aging but comprehensive enough to remain the reference.
200+ MCP servers catalogued across 30 categories — the broadest directory available. Slightly aging (last commit Sept 2025) but comprehensive enough to remain the starting point for MCP discovery.
Curated, opinionated list of high-quality MCP servers accessible over the internet. The remote-only focus matters — these are production-ready services you don't need to run locally.
80+ production-ready remote MCP servers with quality gate (official support, active maintenance required). Auth method documented per server (OAuth 2.1, API key). Supabase, HubSpot, Stripe, Notion, Linear, Vercel, Sentry all present. The remote-only focus means these are copy-paste integrations — no hosting required.
07
Agent Frameworks & Multi-Agent Patterns
Orchestration systems for running multiple Claude Code agents in coordination. From parallel agent farms to self-improving swarms.
Enterprise AI orchestration with 100+ agent types, queen-led swarms (Raft/BFT/Gossip consensus), self-learning Q-Learning router, and a RuVector intelligence layer with HNSW search and LoRA adaptation. Supports Claude, GPT, Gemini, Cohere, and local models. The anti-drift swarm config (hierarchical topology, 6-8 agents max, verification gates) is the most practical pattern for operators who just want multi-agent work to stay on track. Formerly claude-flow.
The kitchen-sink orchestration platform. 34K stars and 6,000+ commits. If you want swarm consensus, self-learning routing, and WASM-accelerated transforms in one package, this is it — but the complexity ceiling is real.
A web UI and desktop app (Windows/macOS/Linux) that turns Claude Code into a visual multi-agent workspace. @mention agents by name for direct routing, or describe a task and let the orchestrator decompose it across agents. Agents can run locally or on remote machines (Mac Mini, cloud instances). The key differentiator: no API keys required — everything runs through your existing Claude subscription via CLI auth. Currently single-room; multi-room support planned. Forked from sugyan/claude-code-webui.
The visual layer for multi-agent Claude Code. 856 stars. @mention agents in a chat UI, mix local and remote instances, no API keys — just your Claude subscription.
Orchestrates 20-50 parallel Claude Code sessions via tmux for systematic codebase improvement. Three workflows: bug fixing (lint/typecheck problems), best practices implementation (guided by 31 markdown guides across 34 tech stacks), and cooperating agents (distributed coordination via prompt engineering alone — no infrastructure code). The standout insight: the coordination protocol is implemented entirely through prompts. Agents autonomously create lock files, claim work, check completed logs, and handle stale locks. Real-time dashboard monitors context usage and heartbeat. Auto-recovery with adaptive idle timeout.
Brute-force parallel agent farming. 793 stars. Run 20-50 Claude Code agents simultaneously in tmux for mechanical codebase improvements — bug fixing, best practices sweeps, coordinated refactoring.
A complete software development lifecycle orchestrator with 18 specialized agents (Researcher, Architect, Coder, Security Auditor, PR Shepherd, etc.) and a 9-phase workflow from issue to merged PR. The standout patterns: a parallel 5-agent design review gate (PM, Architect, Designer, Security, CTO) with a 3-iteration cap, a 4-phase execution loop where the orchestrator never trusts subagent self-reports, and cross-model adversarial review (writer always reviewed by a different model). Self-reflection after every merge extracts patterns from code review feedback and user corrections. Uses BEADS for git-native task tracking. Works with Claude Code, Gemini CLI, and Codex CLI.
Production-extracted SDLC orchestration from a Technorati founder. 227 stars belies the depth — 18 agent personas, 9-phase workflow, adversarial cross-model review, and a self-reflection loop that learns from your corrections.
08
Learning, Reference & Research
From beginner guides to systematic academic analyses. Where to go when you want to understand Claude Code deeply, not just use it.
Four-principle CLAUDE.md derived from Karpathy's viral post on LLM coding pitfalls. Encodes Think Before Coding (surface assumptions, push back on complexity), Simplicity First (no features beyond what was asked, no speculative abstractions), Surgical Changes (touch only what the request demands), and Goal-Driven Execution (transform imperative tasks into declarative goals with verification loops). Also available as a Cursor-compatible .mdc file and a Claude Code plugin via marketplace. 90K stars make it the community's canonical starting point for CLAUDE.md authoring.
The minimum viable CLAUDE.md. 90K stars for four principles from Karpathy's viral post on LLM coding pitfalls — Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution. If you do nothing else, encode these four rules.
Comprehensive progression from basic Claude Code usage to advanced agentic engineering patterns with 82+ practical tips. Not just principles — includes real workflow implementations from 10+ production projects and community-contributed orchestration patterns. The title ('Vibe Coding to Agentic Engineering') captures the journey most operators take, and the depth backs it up.
The best-practices repo the community adopted. 48K stars. 82+ practical tips, real-world workflow implementations from 10+ projects, and community-contributed orchestration patterns. Covers the research-plan-execute-review-ship architecture that actually works.
From basic concepts to advanced agents with copy-paste templates that bring immediate value. The visual-first approach is rare and effective — you see what things look like before you build them.
Structured 10-module tutorial from basics to agent orchestration. 29K stars. Copy-paste templates that work, Mermaid diagrams that explain, and a progressive 'combine features into workflows' approach that's rare in the ecosystem.
Ready-to-use agent skills spanning research, science, engineering, analysis, finance, and writing. The cross-domain breadth makes it useful beyond pure academia.
133 ready-to-use scientific research tools spanning biology, chemistry, medicine, genomics, and materials science. The cross-domain integration (all skills bundle together for interdisciplinary workflows) is what 19K stars validates.
Complete archive of Claude Code's 110+ system prompts extracted directly from source, updated within minutes of each release. Covers agent prompts, tool descriptions, system reminders, memory system internals, hook configuration, and plan mode variants. The essential reference for understanding the prompt substrate your CLAUDE.md instructions coexist with. Maintained by Piebald AI with a CHANGELOG spanning 162 versions. Also ships tweakcc tool for patching local Claude Code installations.
The Rosetta Stone for Claude Code power users. 110+ system prompts extracted directly from source, updated within minutes of each release. Before you write a CLAUDE.md, understand the 110+ system prompts your instructions coexist with.
ARIS (Auto-Research-In-Sleep): the standout pattern is adversarial cross-model collaboration where Claude executes and GPT reviews, catching the blind spots same-model review misses. Chains idea discovery → experiment design → execution → review into an overnight loop. Lightweight markdown-only skills, no framework, no lock-in. 7.5K stars validates the concept.
ARIS: autonomous ML research that runs overnight. Claude Code executes experiments while GPT-5.4 adversarially reviews them — no self-review blind spots. Skills stored as plain Markdown, zero dependencies. Real papers accepted at AAAI 2026 (7-8/10 scores). The 'run it before bed, review results in the morning' workflow actually delivers.
Covers Claude Code from beginner to power user with production-ready templates, workflow guides, learning materials, and quizzes. Organized as a course with progression, not a flat reference — that's why it works for beginners. Unique among learning resources for its security threat intelligence: tracks 24 CVEs and 655 known malicious skills, plus methodology workflows (TDD/SDD/BDD) and production hardening strategies.
The most thorough Claude Code guide. 24K+ lines, 247 production templates, 271 quiz questions, 41 visual diagrams. The security section alone (24 CVEs tracked, 655 malicious skills catalogued) makes this unique. Course-structured, not flat reference.
Complete research-to-publication skill chain that mirrors the actual academic workflow. Each stage is a proper Claude Code skill, not a prompt template.
Full research-to-publication pipeline with anti-sycophancy safeguards, fabricated-citation detection, and post-publication audits. The human-in-the-loop philosophy (AI as copilot, not pilot) is the real differentiator.
Complete hooks reference implementation covering all 13 Claude Code lifecycle events with working Python scripts. Demonstrates exit code flow control (exit 0 success, exit 2 blocking error), PreToolUse security blocking, Stop hook forced continuation, builder/validator agent teams (Builder has all tools, Validator read-only), UV single-file script architecture (zero dependency management), meta-agents that generate new agents, 9 progressive status lines, and 8 output styles. The definitive hands-on guide for deterministic control over Claude Code behavior.
The definitive hooks deep-dive. Every lifecycle event, every exit code, every flow control mechanism demonstrated with working code. The builder/validator team pattern and UV single-file architecture are the two patterns most worth highlighting.
Research paper quality analysis of Claude Code architecture for designing AI agent systems. Unusual depth for a GitHub learning resource — built for people who want to understand the system, not just use it.
Academic-grade reverse engineering of Claude Code v2.1.88 (~512K lines of TypeScript). The insight: '98.4% infrastructure, 1.6% AI.' Traces five human values through thirteen design principles to implementation choices. Theory-to-practice bridge for anyone designing agent systems.
Strategic research thinking agents for idea evaluation, project triage, and structured brainstorming. Positioned upstream of writing tools — it helps you decide what's worth pursuing.
Strategic research triage — tells you when an idea isn't worth pursuing and identifies the single riskiest assumption to test first. Positioned upstream of writing tools. The 'taste and strategic problem selection' framing is what every research lab needs but nobody builds.
Ready-to-use hook collection focused on safety (dangerous command blocking, secret protection), automation (auto-staging), and notifications. Three configurable safety levels with 262 tests validating behavior. The complement to hooks-mastery — grab a working hook and adapt it.
Pre-built safety hooks with configurable levels (critical/high/strict) and 262 passing tests. Blocks dangerous commands, protects secrets, auto-stages files. Drop into any Claude Code setup. Less educational than disler's mastery guide, but more immediately practical.
09
Meta: Awesome Lists & Indices
The directories that index the directories. Start here if you want breadth, not depth.
Skills, hooks, slash commands, agent orchestrators, applications, and plugins — all in one curated list. The first place most people discover the Claude Code ecosystem.
41K stars. THE community index. If you're only going to bookmark one meta-list, it's this one.
A complete, working example of Claude Code project configuration. Hooks, skills, agents, commands, and GitHub Actions workflows — all wired up. The reference architecture for Claude Code projects.
The best 'show me a complete project' reference. Hooks, skills, agents, commands, GitHub Actions — all configured and working together. The intelligent skill evaluation (auto-suggests skills based on prompt analysis) is an underappreciated feature.
Plugins organized by capability with adoption signals. The quantitative metrics add editorial judgment that pure directories lack.
~40+ production-ready plugins with adoption metrics, plus the 'connect-apps' plugin that wires Claude to Gmail, Slack, GitHub, Notion, and 500+ services. The action-execution angle (not just code generation) differentiates it from pure listing repos.
The most comprehensive single index of the Claude Code ecosystem with verifiable counts and Apache 2.0 licensing. Was #1 trending on GitHub in February 2026. The taxonomy structure (agents, plugins, skills, commands, hooks organized by category) is itself a useful reference for how to organize a skill ecosystem.
135 agents, 35 curated skills, 42 commands, 176+ plugins across 850+ files. Was #1 trending on GitHub. The quantified claims hold up, and the domain-specific agent coverage (blockchain, SRE, security) goes deeper than pure aggregate lists.
Automated collection of Claude Code plugin adoption metrics using n8n workflows. The only automated quantitative tracking of the Claude Code ecosystem.
Quantitative star-ranked index of 14,023 Claude Code repos (top 100 displayed). Automated via n8n workflows. The only source for ecosystem market sizing — if you want to know where gravity is concentrating, this is it.
10
Developer Tooling & Utilities
Developer-focused tools, CI/CD integration, session management, context optimization, and linting.
Curated collection of 131+ Claude Code subagents with a standardized architecture: YAML frontmatter, role-based tool permissions, and model routing by cognitive load. The vertical depth is impressive — healthcare alone has 51 sub-agents. The real value for operators is the permission model (which tools each role archetype gets) and the deployment pattern (project `.claude/agents/` overrides global `~/.claude/agents/`).
The definitive subagent collection. 18K stars. 131+ agents across 10 categories — but the real contribution is the permission model (read-only vs. research vs. code-creator tool grants) and the model routing strategy (Opus for deep reasoning, Haiku for quick tasks). This is how agent architecture patterns are consolidating.
Anthropic's official GitHub integration for Claude Code. @claude mentions in PRs and issues trigger a full agent that reads context, implements changes, and outputs structured JSON for downstream workflow steps. Runs on your runners with Anthropic, Bedrock, Vertex, or Foundry auth. The path-specific triggering and checkbox progress tracking are the details that separate this from 'just another code review bot.'
Official GitHub Actions integration. 7.3K stars. The bridge between Claude Code and CI/CD — but the real power is validated JSON outputs that let you chain Claude into multi-step workflows. Not just code review; it's an agent that participates in your pipeline.
End-to-end scientific writing pipeline: literature review, outline generation, draft writing, citation management, and formatting for specific journals. Handles papers, conference posters, grant proposals, and technical reports. The structured output matches what reviewers expect.
Deep research + scientific writing. Papers, reports, posters, grant proposals — publication-ready output.
Exports Claude Code session context into a portable format you can import into Cursor, Windsurf, or other AI coding tools. Zero-install via npx. Useful when you start exploration in Claude Code and want to continue in an IDE-native tool, or vice versa.
Resume any AI session in another tool — 182 cross-tool handoff paths across 14 tools (Claude Code, Codex, Copilot, Gemini, Cursor, Amp, and 8 more). Reads native session formats (JSONL, JSON, SQLite, YAML). Zero install via `npx continues`.
Rust-based context compression that sits between your AI tools and your codebase. The 10-mode reading system (from full file to aggressive compression) lets you tune the cost/context tradeoff per project. AST-aware parsing means it compresses intelligently — keeping function signatures and types while stripping implementation detail. Works with Claude Code, Cursor, and Copilot simultaneously.
Token cost reduction via intelligent context compression — 70-90% real savings. Hybrid shell-hook + MCP architecture with 10 file reading modes (full, map, signatures, diff, aggressive, entropy, task, reference, lines, auto). Rust single binary with tree-sitter AST parsing for 18 languages. Cross-session memory means it gets smarter over time.
Native macOS app that aggregates sessions from Claude Code, Cursor, and other AI coding tools into one dashboard. Tracks token usage, rate limits, session history, and cost per project. Useful for anyone managing AI spend across multiple tools.
Session browser + analytics + rate limit tracking across 7+ AI coding tools. Native macOS app with 'Agent Cockpit' live HUD. Local-only indexing, no telemetry.
Ships DESIGN.md templates that define aesthetic families, component styles, and remix recipes Claude Code can follow during frontend work. Fills the gap between "make it look good" and actually giving the agent design constraints to work within. Includes comparison matrix vs Figma Make, Lovable, and v0.
DESIGN.md templates with 9 aesthetic families and 30+ real-brand examples (Linear, Stripe, Figma, Arc). Anti-slop guidance and token-budget recipes. The only design-focused skill repo that fills the gap between 'make it look good' and giving agents actual constraints.
Plugin-based workflow system with 16 recipe skills and 11 knowledge skills. Six workflows: backend, frontend, fullstack, quick fix, diagnosis, and reverse engineering. The architecture isolates each phase in a fresh agent context — preventing the quality degradation that plagues long-running agent sessions. The design-sync agent (cross-document consistency verification) and the requirement-analyzer (automatic complexity routing) are the standout patterns. Install via plugin marketplace.
Dev workflow system. 320 stars. Six workflows from backend to reverse engineering. The key insight: each phase runs in a fresh agent context so quality doesn't degrade as complexity grows. The vertical slicing pattern (deliver working slices, not layers) and the UI Specification bridge are the opinionated choices that matter.
Knowledge graph-based recommendation system that watches what you develop, walks 963K edges connecting skills, agents, and MCP servers, and loads the right bundle on the fly. The hook integration (PostToolUse and Stop events) means it runs passively — no manual skill discovery. Session-aware memory means recommendations improve over time. The ctx-monitor dashboard gives you a live view of what's loaded, graph relationships, and quality scores. For operators running 50+ skills, this solves the 'token budget vs. capability breadth' tradeoff.
Skill recommendation engine. 241 stars. Walks a 13K-node knowledge graph (1,968 skills, 464 agents, 10,786 MCP servers) to load the right 10-15 tools per session and unload stale ones. Hooks into PostToolUse and Stop events for automatic context management. This is the missing layer between 'I have a skill library' and 'the right skills are loaded right now.'
Rust-based linter with 414 rules that validates AI agent configs across Claude Code, Cursor, Copilot, Kiro, Gemini CLI, MCP, and Cline. The real value: confidence-tiered auto-fix (safe/medium/unsafe), an LSP for real-time editor feedback, and an MCP server mode so agents can lint their own config. The GitHub Action preset (`--target claude-code`) makes CI integration trivial. If you maintain more than a handful of SKILL.md files, this is infrastructure.
AI config linter. 210 stars. 414 rules across 9 platforms. Vercel research showed skills invoke at 0% with bad syntax, and 66% of developers cite 'almost right' configs as their top AI frustration. This catches that. The confidence-tiered auto-fix (`--fix-safe` vs. `--fix-unsafe`) is the detail that matters for CI.
Lightweight CLI that manages multiple project contexts — switch between workstreams, bind conversation transcripts to projects, and keep CLAUDE.md files organized per-workspace. Exact transcript binding (not summaries), safe workstream branching, local-first SQLite storage. Small but solves a real friction point for anyone running Claude Code across several codebases.
Local context manager with workstreams and transcript binding. Solves the 'I have 5 projects' problem.
Claude Code skill that automatically selects the right scraping strategy (static HTML, headless browser, API extraction) based on the target page. Smaller project, but the auto-detection approach means you don't need to figure out which scraping method to use for each site.
Intelligent scraping with quality-gated recon (5-phase architecture) and API-first approach (10-100x faster than HTML). Auto-detects the right strategy per site. Progressive disclosure reduces token usage 70-80%. Small project, clever architecture.
Know a repo we missed?
This page refreshes quarterly. Numbers update automatically; editorial verdicts get a human review. Submit a repo for evaluation — we apply the same Tier 1 + Tier 2 process to every submission.
