Skip to main content

Claude Agents Library — Quick Reference Guide

A concise reference for claude-agents-library — a curated collection of 34 role-based personas you can drop into any Claude Code project to make Claude consistently behave as a domain specialist.

For the hands-on walkthrough, see claude-agents-library-tutorial.md. For how this compares to agent-skills and Agent OS, see comparing-claude-agents-library-agent-skills-agent-os.md.


What It Is (in one paragraph)

Claude Agents Library is a persona library — 34 markdown files, each a complete job description that Claude can step into. Copy one into your project's .claude/agents/ directory, reference it in CLAUDE.md, and Claude will consistently answer as a Backend Architect, a Growth Hacker, or a Performance Benchmarker without you having to prompt-engineer the role every time. It's not a workflow framework (there are no verification gates), not stateful (nothing persists between sessions), and not a plugin (no manifest, no slash commands). It's a lightweight costume rack for Claude, MIT-licensed, zero install.


The 34 Agents at a Glance

design/ (5) ui-designer, ux-researcher, brand-guardian,
visual-storyteller, whimsy-injector

engineering/ (6) frontend-developer, backend-architect,
mobile-app-builder, ai-engineer,
devops-automator, rapid-prototyper

marketing/ (7) tiktok-strategist, instagram-curator,
x-twitter-strategist, reddit-community-builder,
app-store-optimizer, content-creator, growth-hacker

product/ (3) trend-researcher, feedback-synthesizer,
sprint-prioritizer

project-management/ (3) experiment-tracker, project-shipper,
studio-producer

studio-operations/ (5) support-responder, analytics-reporter,
infrastructure-maintainer, legal-compliance-checker,
finance-tracker

testing/ (5) tool-evaluator, api-tester, workflow-optimizer,
performance-benchmarker, test-results-analyzer

Total: 34 agents across 7 categories. All files follow the same anatomy, all are independent, all are MIT licensed.


Anatomy of an Agent File

Every agent is a markdown file with YAML frontmatter and six required sections:

---
name: Content Creator
category: marketing
version: 1.0
---
  1. 🎯 Purpose — 2-3 sentences in second person ("You are...") that establish the persona and its core value
  2. 📋 Core Responsibilities — 3-5 subsections, each a bullet list of specific tasks
  3. 🛠️ Key Skills — tools, technologies, methodologies the agent uses
  4. 💬 Communication Style — tone, approach, what to emphasize
  5. 💡 Example Prompts — 4-6 ready-to-use prompts that invoke the agent
  6. 🔗 Related Agents — 3-5 cross-references forming a collaboration graph

That's it. No frontmatter beyond name, category, version. No process steps, no verification gates, no rationalization rebuttals. The power comes from specificity — each agent lists real tools ("Figma, Sketch, Adobe XD"), granular responsibilities (not "manage projects" but "identify at-risk items early, facilitate status updates, escalate effectively"), and realistic example prompts you can copy verbatim.


Installation

There is no install. It's copy-paste.

Pattern 1: Project integration (most common)

# Clone once
git clone https://github.com/vibecoding/claude-agents-library.git ~/src/claude-agents-library

# In your project
mkdir -p .claude/agents
cp -r ~/src/claude-agents-library/engineering .claude/agents/
cp -r ~/src/claude-agents-library/testing .claude/agents/

Then reference the ones you want in CLAUDE.md:

## Active Agents

- [Backend Architect](.claude/agents/engineering/backend-architect.md)
- [API Tester](.claude/agents/testing/api-tester.md)
- [DevOps Automator](.claude/agents/engineering/devops-automator.md)

Claude Code picks up CLAUDE.md automatically. The linked files become part of its context whenever it reads the CLAUDE.md.

Pattern 2: Direct prompt reference

Skip the file copying entirely. Paste or reference the agent in the prompt:

"Acting as the Backend Architect agent, design a database schema
for a multi-tenant SaaS app with per-tenant encryption keys."

Works in Claude Chat, Claude Code, or any other Claude surface.

Pattern 3: Multi-agent composition

Stack personas for work that spans specialties:

"Using both the Growth Hacker and Analytics Reporter agents,
analyze the funnel data in reports/funnel-q1.csv and propose
three experiments ranked by expected impact."

Claude blends the two perspectives — Growth Hacker contributes experiment design, Analytics Reporter contributes measurement rigor.


When Each Agent Earns Its Keep

You're about to…Use
Sketch a database schema or service boundaryengineering/backend-architect
Implement a React/Vue componentengineering/frontend-developer
Design a CI/CD pipeline or Dockerize somethingengineering/devops-automator
Spike a concept in a day or twoengineering/rapid-prototyper
Wire up an LLM or embedding pipelineengineering/ai-engineer
Draft a landing-page headline or email sequencemarketing/content-creator
Plan a growth experimentmarketing/growth-hacker
Review an API contracttesting/api-tester
Benchmark performance and set SLOstesting/performance-benchmarker
Write user research interview questionsdesign/ux-researcher
Audit a UI for brand consistencydesign/brand-guardian
Triage a pile of user feedbackproduct/feedback-synthesizer
Decide what ships this sprintproduct/sprint-prioritizer
Draft a launch checklistproject-management/project-shipper

Strengths

  • Zero install, zero config. A file is all it takes.
  • Clarity of voice. Each agent is a complete job description — no ambiguity about who Claude is in this context.
  • Composition graph. The "Related Agents" sections connect the library into a workable team chart.
  • Platform-agnostic. Works in Claude Code, Claude Chat, Cursor, anywhere Claude runs.
  • Easy to fork. They're just markdown; change the voice, add responsibilities, tune to your project.

Limits (and when to reach for something else)

  • Not a workflow. An agent tells Claude who to be, not what steps to take. A "Frontend Developer" persona won't stop Claude from skipping tests or shipping unreviewed code. For that, use agent-skills.
  • Not stateful. Nothing persists between sessions. For roadmap tracking, decision logs, and multi-session memory, use Agent OS.
  • Not domain-tuned. The library targets consumer/SaaS work: marketing, product, growth, web/mobile engineering. There is nothing for scientific computing, HPC, MPI, GPUs, reproducibility, or research workflows. See the HPC RSE guide for how to fix that.
  • No verification. Agents don't enforce quality gates. They can still produce shallow, unverified output — the persona only shapes voice, not discipline.

How it compares to its siblings

claude-agents-libraryagent-skillsAgent OS
UnitPersonaWorkflow (SKILL.md)Product plan + spec + tasks
Answers"Who should Claude act as?""How should Claude do this?""What are we building, in what order?"
StateStatelessStatelessStateful (writes .agent-os/)
DisciplineNoneAnti-rationalization tables, verification gatesHuman-review gates between phases
Best forVoice / roleTask disciplineLong-running projects

Full comparison: comparing-claude-agents-library-agent-skills-agent-os.md.


Filename conventions

  • File: kebab-case.md (frontend-developer.md, growth-hacker.md)
  • Title: [Emoji] [Title Case] Agent (🎨 Frontend Developer Agent)
  • Name: descriptive and role-based — "Frontend Developer", not "WebUI Specialist"

If you write your own, match these conventions so the file sits cleanly alongside the upstream library.