The 10 Best .claude/ Setups in 2026

Problem: A blank .claude/ folder is the worst place to start. Every public repo claims to be the one to install. Stars are noisy, READMEs sprawl, and half the install commands tell you to git clone into your home directory like it is 2024.

Quick Win: The 5-minute baseline. Drop one CLAUDE.md, install one official marketplace, then pick one methodology framework. Run this inside Claude Code:

# Drop the Karpathy CLAUDE.md into your repo root
curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

# Inside a Claude Code session, install Anthropic's official skills
/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills

# Pick one methodology framework
/plugin install superpowers@claude-plugins-official

That stack covers behavior, skills, and discipline. The full ranking below tells you which other layers to add when you outgrow it.

TL;DR. The 10 best .claude/ setups in May 2026

Star counts pulled live on 2026-05-15. They move daily. Karpathy's repo gained 5,828 stars in a single day after release. Treat the table as a snapshot.

A note on the ranking. Stars sort the table for fast scanning, not for quality. The 9.5k entry at the bottom does something none of the others do (covered later). Read the tier list at the end before you decide what to install.

What makes a great .claude/ setup in 2026

Most "awesome" lists treat every repo as one bucket. The setups that actually hold up split into four layers, and a strong stack has all four:

1. Behavior layer: a CLAUDE.md at the repo root that fixes the model's default bad habits. Over-engineering, ignoring patterns, surprise dependencies. The Karpathy file is the canonical example. 70 lines, zero install.
2. Skill layer: .claude/skills/<name>/SKILL.md files that load only when invoked. Reference material costs almost nothing until the model needs it. Anthropic's marketplace is the baseline.
3. Methodology layer: a workflow framework like Superpowers or ECC that forces the agent to plan, spec, test, and review before it ships. Stops impulse coding.
4. Memory layer: an MCP server like agentmemory that carries context across sessions. You stop re-explaining your project every time.

A stack with all four can run autonomously for hours without going off the rails. A stack missing any layer breaks down inside one session.

The four-layer view is also how to read the ranked list. Each entry covers one or two layers, never all four. The "what to install first" tier list at the end pairs them so you cover every layer with one install round.

The plugin marketplace install pattern

The modern install path runs inside Claude Code, not in your shell. Anthropic shipped the plugin marketplace as the canonical way to add skills, agents, and commands. It looks like this:

/plugin marketplace add <github-org>/<repo>
/plugin install <package>@<marketplace>

Old guides still tell you to git clone into ~/.claude/. That works. It is no longer the recommended pattern. The marketplace handles updates, version pinning, and the four storage locations Anthropic defined (Enterprise, Personal, Project, Plugin).

When a repo offers both, take the marketplace command. When a repo is one file (like the Karpathy CLAUDE.md), curl is still the right call.

The 10 setups, ranked

1. anthropics/skills

The official Anthropic marketplace. About 135k stars, source: github.com/anthropics/skills. It hosts skills across Creative & Design, Development & Technical, Enterprise & Communication, plus document handlers for PDF, DOCX, PPTX, and XLSX.

Install:

/plugin marketplace add anthropics/skills
/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

When to use it: always. The document skills alone pay for the install. This is your baseline before anything else.

2. obra/superpowers

Jesse Vincent's methodology framework. ~192k stars, source: github.com/obra/superpowers. Skills cover testing (test-driven-development), debugging (systematic-debugging, verification-before-completion), and collaboration (writing-plans, executing-plans, dispatching-parallel-agents, requesting-code-review, using-git-worktrees, subagent-driven-development).

The shape is a 7-phase workflow: Brainstorm, Spec, Plan, TDD, Subagent Development, Review, Finalize. The agent refuses to ship code without tests. Install:

/plugin install superpowers@claude-plugins-official

When to use it: when you want planning before code and discipline over speed. Best fit for engineers who would rather wait two minutes than ship a bad diff.

3. affaan-m/everything-claude-code

Affaan Mustafa's full harness. ~182k stars, 28k+ forks, 170+ contributors, source: github.com/affaan-m/everything-claude-code. Affaan won an Anthropic hackathon in September 2025. The repo packs 60 agents, 229 skills, 75 legacy command shims, instincts, memory optimization, continuous learning, security scanning, and research-first development. Works across Claude Code, Cursor, Codex, OpenCode, Gemini, and GitHub Copilot.

Install via marketplace:

/plugin marketplace add https://github.com/affaan-m/everything-claude-code
/plugin install ecc@ecc

Or full manual:

./install.sh --profile full

When to use it: maximum surface area, daily-driver setup. Heavy. If you cycle between harnesses, ECC is the one stack that follows you.

4. forrestchang/andrej-karpathy-skills

A single CLAUDE.md file (about 70 lines), derived by Forrest Chang from observations Andrej Karpathy posted in April 2026. ~131k stars, source: github.com/forrestchang/andrej-karpathy-skills. It hit 5,828 stars in a single day after release. Get the attribution right. Karpathy did not write the file. Forrest derived it from his post.

Four behavioral principles: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution. Designed to fix the worst LLM coding pitfalls at the source.

Install:

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

Or via marketplace:

/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills

When to use it: first file you drop into any new repo. Zero dependencies, immediate behavior change.

5. garrytan/gstack

Garry Tan's personal toolkit. Tan is President & CEO of Y Combinator. ~97.4k stars, 14.5k forks, source: github.com/garrytan/gstack. Tan's claim from the README: "3 production services, 40+ shipped features in 60 days, part-time, while running YC."

The shape is a virtual engineering team. 23+ slash commands across six roles:

• Planning & Strategy: /office-hours, /plan-ceo-review, /plan-eng-review, /plan-design-review, /plan-devex-review
• Design: /design-consultation, /design-shotgun, /design-html, /design-review
• Development: /review, /investigate, /autoplan, /codex
• Quality & Testing: /qa, /qa-only, /benchmark, /canary
• Release: /ship, /land-and-deploy, /document-release, /document-generate
• Infrastructure: /browse, /retro, /setup-deploy, /sync-gbrain, /pair-agent, /learn

Install:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

When to use it: solo founder running multiple products. The role-based organization is the differentiator. You think in roles, not files.

6. mattpocock/skills

Matt Pocock's personal pack. ~84.3k stars, source: github.com/mattpocock/skills. Pocock teaches TypeScript, and the skills lean that way. 20 skills total across three categories.

• Engineering (10): diagnose, grill-with-docs, triage, improve-codebase-architecture, setup-matt-pocock-skills, tdd, to-issues, to-prd, zoom-out, prototype
• Productivity (4): caveman, grill-me, handoff, write-a-skill
• Misc (6): git-guardrails-claude-code, migrate-to-shoehorn, scaffold-exercises, setup-pre-commit

Install:

npx skills@latest add mattpocock/skills

When to use it: TypeScript daily. The grill-me and caveman skills alone earn the install.

7. ruvnet/ruflo

rUv's multi-agent orchestration platform. ~51.4k stars, source: github.com/ruvnet/ruflo. 100+ specialized agents, vector memory with HNSW indexing, SONA neural patterns for self-learning, Agent Federation for cross-installation collaboration, 12 auto-triggered background workers, 32 native plugins, multi-provider support (Claude, GPT, Gemini, Cohere, Ollama).

The README claims "150x to 12,500x faster search" on the vector memory. That is a vendor benchmark. Treat it as a vendor benchmark, not an independent result.

Install (full wizard):

npx ruflo@latest init wizard

Or lightweight:

/plugin install ruflo-core@ruflo

When to use it: orchestration-heavy work where Claude Code should act like a swarm rather than one agent.

8. addyosmani/agent-skills

Addy Osmani is a Google Chrome eng lead. The pack lines up with the SDLC. ~41.9k stars, source: github.com/addyosmani/agent-skills. 23 skills across Define, Plan, Build, Verify, Review, and Ship.

• Define (3): interview-me, idea-refine, spec-driven-development
• Plan (1): planning-and-task-breakdown
• Build (6): incremental-implementation, test-driven-development, context-engineering, source-driven-development, doubt-driven-development, frontend-ui-engineering, api-and-interface-design
• Verify (2): browser-testing-with-devtools, debugging-and-error-recovery
• Review (4): code-review-and-quality, code-simplification, security-and-hardening, performance-optimization
• Ship (5): git-workflow-and-versioning, ci-cd-and-automation, deprecation-and-migration, documentation-and-adrs, shipping-and-launch

Install:

/plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills

When to use it: you want a skill for every phase of the SDLC, and the author's name does some of the trust work for you.

9. VoltAgent/awesome-claude-code-subagents

A curated subagent collection. ~19.9k stars, badge says "131+ subagents", source: github.com/VoltAgent/awesome-claude-code-subagents. Subagents land in 10 categories: Core Development, Language Specialists, Infrastructure, Quality & Security, Data & AI, Developer Experience, Specialized Domains, Business & Product, Meta & Orchestration, Research & Analysis.

Install one category at a time. Pulling all 131 will bloat your context budget for nothing:

claude plugin marketplace add VoltAgent/awesome-claude-code-subagents
claude plugin install voltagent-lang

When to use it: the bottleneck is subagents specifically, not skills.

10. rohitg00/agentmemory

Rohit Ghumare's persistent memory MCP server. ~9.5k+ stars, source: github.com/rohitg00/agentmemory. Smallest entry on the list. It does the one thing none of the others do: memory that persists across Claude Code, Cursor, Gemini CLI, and Codex sessions.

The README reports 95.2% retrieval accuracy on LongMemEval-S (R@5, 500 questions) and 92% token savings vs. pasting full context (~170K tokens/year vs. 650K+). 12 automatic hooks, 51 MCP tools, multi-agent support via MCP + REST + lease coordination, real-time viewer at port 3113, zero external dependencies (SQLite + iii-engine). Vendor-reported numbers, not independent benchmarks. They are still the most specific numbers in this space.

Install:

npx @agentmemory/agentmemory

When to use it: stop re-explaining your project every session. One memory layer for every tool you use.

What to install first. The tier list

Star order is not install order. Pair the layers instead.

S-Tier. Install today, every project, no exceptions:

• anthropics/skills. Official baseline, free, source-available.
• forrestchang/andrej-karpathy-skills. 70 lines, immediate behavior change.

A-Tier. Pick one based on workflow:

• obra/superpowers if you want methodology and TDD discipline.
• affaan-m/everything-claude-code if you want maximum surface area.
• garrytan/gstack if you are a solo founder running multiple products.

B-Tier. Add when you hit the specific pain:

• addyosmani/agent-skills for SDLC coverage.
• mattpocock/skills for TypeScript work.
• VoltAgent/awesome-claude-code-subagents for subagent depth.

C-Tier. Specialized, install only if needed:

• ruvnet/ruflo for orchestration-heavy use.
• rohitg00/agentmemory for cross-tool memory.

S + A is the minimum viable stack. Two installs, every layer covered except memory. Add C only when you feel the absence.

What Anthropic actually recommends

Anthropic's best-practices post lays out the principles every setup above is built on. Three lines worth pulling.

On CLAUDE.md scope: "root file for the big picture, subdirectory files for local conventions." Put the stable rules at the top, the local quirks where the agent will read them.

On what belongs in CLAUDE.md: "pointers and critical gotchas only; everything else drifts into noise." The Karpathy file follows this exactly. 70 lines, no fluff.

On test discipline: "Running the full suite when Claude changed one service causes timeouts and wastes context on irrelevant output." That is the gap a methodology framework like Superpowers fills.

The official skills documentation defines the four storage locations (Enterprise, Personal, Project, Plugin) and the precedence rules between them. Read it once. Every setup on the list slots into one of those four buckets.

Honorable mentions

Three more setups worth one line each.

sabahattink/antigravity-fullstack-hq: smaller (~20 stars) but well-organized full-stack template. 10 named agents (frontend-specialist, backend-specialist, database-specialist, architect, code-reviewer, test-engineer, security-auditor, devops-engineer, performance-optimizer, documentation-writer) and 28 skills. Install:

curl -fsSL https://raw.githubusercontent.com/sabahattink/antigravity-fullstack-hq/main/install.sh | bash

rohitg00/awesome-claude-code-toolkit: 135 agents, 35 curated skills, 42 commands, 176+ plugins, 20 hooks, 15 rules, 7 templates, 14 MCP configs. Use it as a parts catalog.

quemsah/awesome-claude-plugins: a live leaderboard tracking adoption metrics across 16,920+ indexed repositories as of 2026-05-14. Useful as a reference for current install counts before you commit to anything.

Building a .claude/ that ships a SaaS

Each setup above gives you one or two layers from the four-layer model. None of them ship the database, the auth, the payments, the email, the design system, the legal pages, or the production skeleton you need to launch.

That gap is what Build This Now fills. It is a .claude/ framework that drops into a complete Next.js + Supabase production codebase. 32 specialist agents, 95+ skills, a planning team, a build pipeline, an adversarial GAN evaluator loop, and post-launch monitoring. The 10 setups above are the best way to make Claude Code smarter. Build This Now is the fastest way to ship a product with it.

FAQ

What is a .claude/ folder?

A .claude/ folder is the directory Claude Code reads on startup. It holds the CLAUDE.md, skills (.claude/skills/<name>/SKILL.md), agents, commands, and settings.json. Anthropic defines four storage locations: Enterprise, Personal, Project, and Plugin. Project (.claude/ in your repo) is where most public setups land.

What is the best Claude Code setup for solo founders?

anthropics/skills plus forrestchang/andrej-karpathy-skills plus garrytan/gstack. Three installs, full coverage, role-based commands that map to how a solo founder thinks. Add agentmemory once you have more than one project running.

Do I need to install all 10 of these?

No. The S-Tier (Anthropic + Karpathy) covers behavior and skills. One A-Tier pick (Superpowers, ECC, or gstack) covers methodology. That is your minimum viable stack. The rest are pain-driven additions.

How do .claude/ skills, agents, and plugins differ?

A skill is a file that loads on demand (.claude/skills/<name>/SKILL.md). An agent is a subagent process spawned by the main session, scoped to one task. A plugin is a marketplace package that bundles skills, agents, commands, and settings. Plugins are the modern install unit. Skills and agents are the primitives plugins ship.

Where does .claude/ live?

Four locations, in precedence order: Enterprise (org-wide), Personal (~/.claude/), Project (.claude/ in the repo), and Plugin (installed via the marketplace). Local overrides global. The official skills docs cover the rules.

How often does this list change?

Quarterly. Star counts move daily, and trending repos cool off fast. Everything Claude Code went from 32K to 163K stars in three months. We re-rank this list every quarter.

A great .claude/ setup is four layers stacked. Behavior, skills, methodology, memory. Pick one repo per layer and most of your friction disappears.