How to use the SageMaker Deployment Planner agent skill

What the skill makes the agent do

An agent skill is a folder with a SKILL.md file that gives an AI agent specific instructions for one kind of task. This skill covers the planning part of a SageMaker deployment: it asks a few short questions about the model, expected traffic and latency needs, recommends a hosting pathway, and checks whether your account has enough GPU quota for the instance it suggests.

The skill lists real-time endpoints, real-time with scale to zero, serverless, async inference, batch transform and Bedrock Custom Model Import as pathways. It states that real-time, real-time scale-to-zero and async are the three that are scripted. Serverless, batch transform and Bedrock Custom Model Import get only a brief hand-off explanation.

• Asks the minimum needed: which model, and roughly how often it will be called
• Picks a pathway and says why, in one sentence per option
• Runs a read-only quota check with aws service-quotas before recommending an instance
• Hands setup to companion skills for AWS context, Python environment, IAM, image selection and deployment defaults

Install the skill from the ZIP

Download the ZIP for this skill from this page and unzip it. To make it available across all your projects, place the unzipped folder in ~/.claude/skills/. To limit it to one project, place it in that project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills

Ask for a deployment in plain language

You do not need special syntax. The skill's description lists phrases it responds to, such as deploy a model, host this LLM on AWS, serve this embedding model, deploy a reranker, host this for async inference, create an endpoint, or serve my fine-tuned model. It also responds to vague requests like wanting a model running on AWS with the agent choosing the details.

Give the agent the model name or an S3 path to your artifacts if you have them. The model name usually tells the agent whether it is an LLM or an embedding model.

• Good: "Deploy Qwen3-0.6B on SageMaker. I'll call it a few times a day."
• Good: "Host this fine-tuned model from s3://my-bucket/model for async inference."
• Vague but supported: "I just want this running on AWS, you figure it out."

What the agent will ask and how it decides

The skill tells the agent not to front-load questions. A typical minimal set is what model you want and roughly how often it will be called. If you already answered something, the agent should not ask again. Region comes from the companion context-discovery skill, so you will not be asked for it unless you mention it.

Before creating anything, the agent states a recommendation and waits for your confirmation, so you know what you are about to spend money on.

What the skill needs

You need an AWS account with SageMaker access, and the AWS CLI available to the agent, because the quota check runs aws service-quotas commands. The skill also coordinates five companion skills: hf-cloud-aws-context-discovery, hf-cloud-python-env-setup, hf-cloud-sagemaker-iam-preflight, hf-cloud-serving-image-selection and hf-cloud-sagemaker-production-defaults. The deployment scripts deploy.py, deploy_ic.py and deploy_async.py are referenced from the production-defaults skill and are not included in this file, so you will need that companion skill to run them.

• AWS account with SageMaker access and permission to create SageMaker resources
• AWS CLI installed and usable by the agent
• Companion skills listed above
• Note: the skill's own file list is SKILL.md only

Safety notes: scripts and network access

The skill is marked as running scripts and using the network. Its quota check is read-only and runs through the AWS CLI. The deployment itself is carried out by scripts in the companion production-defaults skill, so review those before running them.

The skill says a failed quota check should not stop the workflow, and it should not ask you to change IAM for a preflight check. The real limit will appear as ResourceLimitExceeded during deployment. Do not confirm a deployment for an instance type you have not agreed to, because the skill says the user should know what they are spending before anything is created.

• Quota check is a preflight optimization, not a gate
• Deployments create billable SageMaker resources only after you confirm
• Scripted deployment lives in the companion skill, not in this file

Tips for better results

Tell the agent how often the model will be called and how quickly responses must arrive. Those two answers drive most of the pathway choice. If you need an interactive app, say so, because the skill warns that scale-to-zero real-time endpoints can fail every request during the wake-up period with a 400 error.

For GPU instances, the skill notes that ml.g5 and ml.g6 are both 24 GB tier options and that ml.g6e suits models that do not fit in 24 GB. For small embedding models, the skill suggests asking for CPU variants first.

• Name the exact model ID or S3 path
• State your latency need: interactive, near-real-time or async
• For embedding models under 1B parameters at moderate traffic, expect a CPU recommendation
• If the quota shows zero for a GPU type, ask for a different instance or request an increase before deploying