SageMaker Deployment Planner
Choose a practical Amazon SageMaker deployment path for your model, traffic pattern, and latency needs.
The skill requires quota checks, IAM/context preflight, and explicit user confirmation before creating a billable endpoint, reducing unauthorized external effects. Deductions apply because it does not define least-privilege IAM, sensitive model or data handling, credential protection, deployment rollback, or complete data-flow disclosure; publisher identity is also unverified.
The phases and pathway rules are broadly self-consistent and instruct the agent to check unknown image URIs, SDK surfaces, and quotas. Deductions apply because the skill depends on several unprovided skills and includes no tests for its key paths, abnormal-input handling, diagnosable failure messages, or reproducible outcomes; static evidence caps this below 10.
Invocation conditions, audience, model categories, traffic and latency questions, and boundaries between real-time and async are fairly clear. Deductions apply because non-fit ranges remain incomplete, Chinese-language support is unspecified, and core operation depends on Amazon SageMaker, AWS CLI, and related overseas services without mainland-China reachability guidance.
The document uses a useful phased structure and provides discovery inputs, pathway criteria, instance-selection guidance, confirmation wording, Apache-2.0 licensing in repository context, and generated-artifact validation in README/CI. Deductions apply because the skill itself lacks versioning, changelog, named maintenance responsibility, update path, FAQ, installation/dependency notes, and systematic limitations or troubleshooting guidance.
The skill provides useful deployment discovery, pathway recommendation, quota-checking guidance, cost confirmation, and handoff to specialized skills, so its planning value is plausible. Deductions apply because it does not deploy directly, depends on other skills and external context, and contains no representative outputs, execution evidence, or comparative-benefit validation; static calibration caps this at 7.
The pinned revision, repository CI validation of generated artifacts, security policy, and concrete command examples provide limited auditability. Deductions apply because there are no committed tests covering this skill's key paths, third-party execution records, or cross-source corroboration, and static review cannot reproduce the conclusions; therefore the score remains at or below 5.
- This is a static source review; no commands, deployments, or tests were executed, so reliability, effectiveness, and verifiability remain uncertain.
- Before use, verify the AWS region, SageMaker quotas, IAM role permissions, sensitivity of model and inference data, and reachability of AWS/Hugging Face services in the target network environment.
- Endpoint creation can incur charges and change external resources; preserve explicit user confirmation, cost limits, and a cleanup or rollback plan for failures.
What does this skill do, and when should you use it?
This skill helps users plan deployments to Amazon SageMaker AI. It identifies the model, workload shape, latency tolerance, and relevant cost considerations before choosing among real-time endpoints, serverless inference, async inference, batch transform, and Bedrock Custom Model Import. It then coordinates follow-on skills for AWS context, Python setup, IAM, serving-image selection, and production deployment. Real-time and async inference are the two scripted deployment pathways described in the source.
Requests a Hugging Face model ID, S3 artifact path, or model name, and asks for model type, approximate call frequency, latency tolerance, or cost sensitivity when needed; selects a deployment pathway; recommends instance families after checking SageMaker endpoint quotas with the AWS CLI; and hands the work to context discovery, Python environment setup, IAM preflight, serving-image selection, and production-defaults skills after user confirmation.
- A team wants to expose a Hugging Face text-generation model through an AWS SageMaker endpoint.
- A developer needs to host an embedding or reranking model on SageMaker.
- A user needs an async deployment plan for bursty text-to-image, video-generation, or other long-running inference.
- An AWS operator wants instance guidance based on traffic, latency, model size, cost, and regional quota.
- A user only knows they want a fine-tuned model running on AWS and needs the deployment route selected.
What are this skill's strengths and limitations?
- Covers explicitly named workloads including LLMs, embeddings, rerankers, classifiers, and diffusion models.
- Selects pathways using traffic shape, latency, inference duration, and cost considerations.
- Calls for regional endpoint-quota checks before recommending an instance.
- Uses a concise discovery flow and requires confirmation before deployment.
- It primarily plans and coordinates; later deployment work depends on sibling skills in the repository.
- Serverless inference, batch transform, and Bedrock Custom Model Import are not scripted deployment paths in this workflow.
- The source provides no standalone test suite, platform test results, or current SageMaker image URI.
- AWS CLI access, a suitable AWS environment, and required permissions are needed for quota checks and deployment.
How do you install this skill?
Copy or symlink skills/hf-cloud-sagemaker-deployment-planner/ from the repository's skills/ directory into a standard Codex .agents/skills location, such as $REPO_ROOT/.agents/skills or $HOME/.agents/skills. The README does not document a standalone install command for this specific skill. For the repository collection, it documents Claude Code marketplace installation and hf skills add <skill-name> for additional skills.
How do you use this skill?
After installation, give a SageMaker deployment request such as “Deploy my fine-tuned model to SageMaker” or “Host this embedding model on AWS.” The skill normally asks first for the model and approximate call frequency, then recommends a pathway and instance choice. It waits for confirmation before resource creation.
How does this skill compare with similar options?
Real-time endpoints fit steady traffic and low latency but can waste money during idle periods; serverless fits intermittent traffic but has cold-start and memory limits that may exclude larger LLMs. Async inference fits long-running, large-payload, or bursty requests, while batch transform fits offline dataset scoring. Bedrock Custom Model Import fits supported model families and a Bedrock-compatible API, but not custom inference logic.