A collection of skills for AI coding agents that teach best practices for building with CrewAI. Skills follow the Agent Skills format.
CrewAI architecture decisions and project scaffolding. Covers choosing the right abstraction (LLM.call() vs Agent.kickoff() vs Crew.kickoff() vs Flow), CLI scaffolding, YAML configuration, wiring @CrewBase crews, writing Flows with @start/@listen, conversational Flows with handle_turn(), variable interpolation, and starting points for MCP servers and built-in tools.
Use when:
- Starting a new CrewAI project
- Choosing between abstraction levels
- Scaffolding with
crewai create flow - Setting up agents.yaml and tasks.yaml
- Wiring crew.py or main.py
- Building experimental conversational Flows
- Debugging common setup issues
CrewAI agent design and configuration. Covers how many agents to use, the Role-Goal-Backstory framework, LLM selection, tool assignment, execution limits (max_iter, max_rpm, max_execution_time) and what each really does, planning, memory and knowledge sources with embedders that work without OpenAI, agent guardrails, and YAML vs code configuration.
Use when:
- Creating or configuring CrewAI agents
- Choosing role, goal, and backstory
- Assigning tools or selecting LLMs
- Tuning agent parameters
- Setting up knowledge sources or memory
- Debugging agent behavior
CrewAI task design and configuration. Covers writing effective descriptions and expected output, task dependencies with context, structured output (output_pydantic, output_json, output_file), guardrails, human-in-the-loop review, and async execution.
Use when:
- Creating or configuring CrewAI tasks
- Writing task descriptions and expected output
- Setting up task dependencies
- Configuring structured output formats
- Adding guardrails or human review
- Debugging task execution issues
Answers CrewAI questions from the official documentation, matched to the crewai version the user runs. Covers the two docs sites (the open-source framework and the CrewAI AMP platform), their llms.txt indexes, Markdown pages and docs MCP servers, how to read the docs for a specific release, and checking a docs snippet against the installed package before trusting it.
Use when:
- A CrewAI question is not covered by the other skills
- Another skill asks to re-verify a row for a different crewai version
- Setting up the CrewAI docs MCP server in a coding agent
The current crewai 1.15.x API versus the 0.x API that coding assistants tend to remember. A "you probably wrote X, the current form is Y" table covering imports, Agent/Task/Crew parameters and defaults, kickoff variants, LLM model strings and provider extras, structured output, guardrails, unified Memory, knowledge embedders, the crewai create crew wizard, and removed features. Each row was checked against crewai 1.15.22 and 1.15.23, with a source reference for re-checking after an upgrade.
Use when:
- Writing, reviewing, or debugging any crewai code
- Before trusting remembered crewai syntax
- An import, keyword argument, or provider error appears that looks like a version mismatch
Building Flows on the current API: structured state and state.id, @start/@listen/@router wiring, or_/and_ semantics, router labels versus method names, @persist and resuming with restore_from_state_id, checkpointing, @human_feedback, plot(), and calling crews and agents from flow methods.
Use when:
- Writing or debugging a
Flowsubclass - A listener never fires, or fires twice
- Persisting, resuming, or checkpointing a flow
Giving agents tools and MCP servers: custom BaseTool and @tool, which crewai_tools names really exist, caching and usage limits, Agent(mcps=[...]) string and config forms, rewritten MCP tool names, MCPServerAdapter, timeouts, and what works once deployed.
Use when:
- Writing a custom tool or connecting an MCP server
- An agent never calls a tool, or
mcps=[...]yields no tools - An MCP server works locally but not after deploy
Deterministic, offline testing of crews and flows with pytest, using a stub BaseLLM (included) that drives text, structured output, and tool calls with no API key. Covers asserting on the prompts crewai built, testing guardrails and routing, what crewai test really does, event listeners, and the exact tracing and telemetry environment variable values.
Use when:
- Writing tests for a crewai project or running them in CI
- Stubbing the LLM or writing a custom
BaseLLM - Debugging a run with event listeners or traces
Getting a crew or flow onto CrewAI AMP: the project shape the build expects, crewai deploy validate, the three deploy paths (CLI, GitHub, ZIP upload) and what each one actually uploads, environment variables per deployment, provider keys and extras, and machine-wide org selection.
Use when:
- Running
crewai deploy create,push, orvalidate - A deployment is Online but runs old code or old environment values
- A build fails on project shape, lockfile, or entry points
Calling a deployed crew or flow over HTTP: the token and URL, GET /inputs, POST /kickoff with an {"inputs": {...}} body, which inputs are required, polling GET /status/{kickoff_id} with a deadline, terminal states, webhooks, and writing crews that stay correct under repeated and concurrent kickoffs. Includes a tested Python client and curl equivalents.
Use when:
- Writing a client, backend, or frontend that calls a deployed crew
- A kickoff returns
422 "Missing inputs: ...", or a status poll never ends or saysNOT FOUND - One run's prompt shows another run's data
Each skill loads on its own when a request matches its description. These prompts each triggered the intended skill in a fresh install:
| Skill | Try asking |
|---|---|
getting-started |
"I'm new to CrewAI. Help me set up my first project from scratch - install, create a crew, and run it locally." |
design-agent |
"I'm writing the role, goal and backstory for a CrewAI research analyst agent and it keeps going off-topic and looping. How should I write these fields and which agent settings should I set?" |
design-task |
"How should I write the description and expected_output for a CrewAI task so the output comes back as a validated Pydantic object, and should I use context to pass the previous task's result?" |
ask-docs |
"Where in the official CrewAI docs is the reference for the Crew class planning parameter, and is there an llms.txt I can point my editor at?" |
check-crewai-api |
"My crew.py fails with: ImportError: cannot import name 'BaseTool' from 'crewai_tools'. Also I set memory_config on the Crew. What is wrong?" |
build-flow |
"My Flow listener decorated @listen('summarize') on a method named summarize raises 'listen condition references the handler name'. How do I fix the router wiring?" |
connect-tools-and-mcp |
"I passed mcps=['https://mcp.example.com/mcp'] to my Agent; the stdio version worked locally but after deploy the agent never calls the MCP tool. Why?" |
test-crewai-project |
"How do I write pytest tests for my crew in CI without an OpenAI key? I keep getting ValueError: OPENAI_API_KEY is required." |
deploy-to-amp |
"crewai deploy push says Online but the deployment still runs my old code. What does push actually upload?" |
call-deployed-crew |
"I'm calling my deployed crew's /kickoff endpoint with {'topic': 'AI'} and get 422. What should the body be and how do I poll the result?" |
Every skill names the crewai release it was verified against. If crewai version reports a different major or minor version, the skill tells the assistant to re-check version-sensitive details with ask-docs.
In Claude Code, add this marketplace and install the plugin:
/plugin marketplace add crewAIInc/skills
/plugin install crewai-skills@crewai-plugins
The first command registers the marketplace from this repo's .claude-plugin/marketplace.json. The second installs the crewai-skills plugin from the crewai-plugins marketplace.
To pin to a specific branch or tag, append #<ref>:
/plugin marketplace add crewAIInc/skills#<branch-or-tag>
Each skill contains:
SKILL.md- Instructions for the agentreferences/- Supporting documentation (crew YAML configuration, tools catalog, MCP servers, structured output patterns, API contracts, etc.)
MIT