AI teams move fast and their infrastructure configs need to keep up. But when every engineer maintains their own YAML configs, things get messy. Shared institutional knowledge fragments, onboarding drags, and it’s not always clear what the “right” setup for a given cluster or job is. This can cost AI teams valuable time and resources.

SkyPilot Recipes let you store SkyPilot YAMLs in a shared, team-accessible registry. You can launch them directly from the CLI without any local files:

sky launch recipes:dev-cluster

Let’s take a closer look at how you can use SkyPilot Recipes to reduce DevOps overhead.

Centralized Recipes for Common Workloads

A SkyPilot Recipe is a named SkyPilot YAML configuration stored centrally and accessible to everyone on your team. Recipes can be applied to all SkyPilot resource types: Clusters, Managed Jobs, Pools, and Volumes.

With SkyPilot Recipes, you can:

  • Reuse patterns across workflows: Encode your team’s setup knowledge once and stop rebuilding it from scratch
  • Launch instantly from the CLI: No local YAML file needed, just run sky launch recipes:<recipe-name>
  • Edit in place: View and modify recipes directly in the SkyPilot dashboard, with changes immediately available to everyone
  • Share standardized configs across your org: One source of truth for infrastructure, from dev environments to production training runs

You can use SkyPilot Recipes to standardize common patterns like spinning up standard dev environments, launching tried-and-tested GPU training infrastructure with matching PyTorch versions, or to ensure cost-optimized data processing on spot instances.

Create and Launch your SkyPilot Recipes

Recipes are managed through the SkyPilot dashboard and launched from the CLI. This section will walk you through how to create, launch and manage your recipes.

Create a Recipe To create a new Recipe, run sky dashboard, navigate to the Recipes section, and click New Recipe.

Create New Recipe

Once defined and stored, your recipe is immediately available to your whole team.

Launch a Recipe Use the recipes: prefix in the CLI to launch a cluster, job, pool or volume from a predefined recipe. You do not need to know where the YAML config is or what it contains.

# Launch a cluster from a recipe
sky launch recipes:dev-cluster

# Launch with custom cluster name
sky launch -c my-dev recipes:dev-cluster

# Launch a managed job from a recipe
sky jobs launch recipes:training-job

You can override recipe fields using CLI flags. This allows canonical configs to clearly live in the recipe and still gives flexibility for one-off adjustments at launch time.

# Launch with overrides
sky launch recipes:gpu-cluster --cpus 16 --gpus H100:4 --env DATA_PATH=s3://my-data --secret HF_TOKEN

Manage a Recipe All recipes are listed in the Recipes page of the dashboard. Click any recipe to see its full YAML, launch command, and metadata. You can also edit and delete recipes here.

Recipes Dashboard

Changes made to a recipe are applied immediately and reflected for all users. Deletion is permanent and cannot be undone. Any clusters or jobs already running from a deleted recipe will continue running unaffected.

Best Practices

Here are some recommended best practices when creating your own SkyPilot Recipes:

  1. Use environment variables for flexibility, instead of hardcoding paths or other parameters into the recipe. This allows users to customize at launch time without having to edit the recipe.

  2. Use secrets for sensitive values. Secrets work just like environment variables but are redacted in the dashboard, which means they won’t show up in logs or the UI. Declare them in your YAML and then pass them at launch time.

  3. Separate your setup from run. The setup block runs once when a cluster is provisioned; the run block executes the actual workload. Keeping them clean and separate makes recipes composable and avoids slow re-installations every time you iterate on your workload logic.

  4. Recipes don’t support local file access. Instead, use Git Repositories to store code and use cloud storage buckets or SkyPilot Volumes to store data. Local paths like workdir: . or file_mounts: ~/local-dir will not work.

Get Started Now

Recipes are available now in the SkyPilot dashboard. Run sky dashboard, head to the Recipes section, and start building your team’s shared configuration library.

Check out the Recipes documentation for the full reference, and share what you’re building in the SkyPilot Slack — we’d love to see what recipes your team creates!