Azure OpenAI Custom Deployment Guide
A guided Azure OpenAI deployment flow for precise version, capacity, policy, and SKU control.
The workflow includes pre-deployment user confirmation, stated role prerequisites, capacity validation, stop-on-invalid-data behavior, and a deletion command for failed deployments. It also discloses that Anthropic deployment collects tenant country code and organization name. However, it requires substantial Azure permissions, lacks complete data-flow and sensitive-data handling guidance, does not establish least-privilege boundaries, and has no integrated rollback, so points are deducted.
The documentation provides detailed phases, commands, input validation, and failure feedback, making the happy path plausible from static reading. Reliability is reduced by inconsistent phase numbering, incompletely declared bash/PowerShell/jq dependencies, and unverified Azure CLI/REST assumptions and examples. The static ceiling also applies.
The audience, customization scenarios, preset comparison, non-fit preset scenario, and model-format branches are reasonably clear. Chinese-language interaction support and reachability from mainland-China networks are not evidenced, and exact semantic trigger behavior still depends on the parent router, so points are deducted.
The skill uses progressive disclosure with referenced guides, examples, troubleshooting, MIT licensing, author metadata, and a version number. It lacks a skill-specific changelog, explicit maintenance owner, and clear update path; SKU, model, and repository-count claims may also become stale as the service changes.
The workflow covers authentication, model/version selection, SKU, capacity, RAI, advanced options, confirmation, deployment, and monitoring, so it plausibly completes the core customization task. No key-path execution output or verification is provided, and some examples and capacity calculations require manual validation; therefore the static effectiveness ceiling limits the score.
Concrete CLI/REST commands, referenced files, and repository-level CI/test materials provide some auditability. The supplied tests do not cover this deployment skill's key paths, and nothing was executed for this review, leaving limited reproducibility and no independent corroboration.
- Deployment creates external Azure resources and may incur charges; verify subscription, region, SKU, capacity, RAI, upgrade policy, and spillover target before confirmation.
- Do not treat the example models, SKUs, capacity ranges, or PTU formula as current service guarantees; validate against live API responses and official documentation.
- The Anthropic branch submits tenant country code and organization name; confirm their source, accuracy, and compliance before deployment.
- This assessment is static only; Azure CLI, REST calls, and key-path tests were not executed.
What does this skill do, and when should you use it?
This skill provides an interactive workflow for deploying Azure OpenAI models in an Azure AI Foundry project. It lets users choose the model version, deployment SKU, capacity, RAI policy, upgrade policy, and applicable advanced options. The workflow queries model availability, quota, and regional capacity, with cross-region fallback when the current region cannot satisfy the request. It is designed for controlled deployments rather than one-click selection of the best region.
It uses Azure CLI to verify authentication, parse and validate the project resource ID, list models and versions, query deployable SKUs and capacity ranges, and guide configuration choices. It supports GlobalStandard, Standard, ProvisionedManaged, and DataZoneStandard SKUs, plus dynamic quota, priority processing, spillover, and version upgrade settings where applicable. It generates and validates a deployment name, presents a configuration review, creates the deployment, polls its status, and reports the endpoint and portal link; it can also search other regions when local capacity is unavailable.
- An Azure developer needs a specific model version instead of the latest available version and wants an interactive selection flow.
- A platform engineer needs exact capacity, a production-oriented GlobalStandard deployment, or Provisioned Throughput Units and must check subscription quota first.
- An AI application team needs to choose an RAI content-filter policy or configure dynamic quota, priority processing, spillover, or upgrade behavior.
- A project cannot obtain capacity in its current region and needs the workflow to search other regions for a viable deployment location.
What are this skill's strengths and limitations?
- Provides fine-grained control over model version, SKU, capacity, RAI policy, and advanced deployment settings.
- Queries live deployability, quota, and capacity data instead of hardcoding SKU lists.
- Includes configuration review and cross-region fallback when local capacity is unavailable.
- Supports Standard, GlobalStandard, ProvisionedManaged, and DataZoneStandard deployment options.
- Requires Azure CLI authentication, a valid project resource ID, and sufficient Azure permissions.
- Requires network access to Azure APIs; deployment is blocked if capacity data cannot be obtained.
- Availability depends on the region, subscription, and model, so a successful deployment is not guaranteed.
- Full scripts, capacity formulas, and advanced-option details are delegated to reference files not included here; no independent test results for this skill are supplied.
How do you install this skill?
Run npx skills add microsoft/skills in the target project and select the required skills in the installation wizard. The skill is located at .github/plugins/azure-skills/skills/microsoft-foundry/models/deploy-model/customize/SKILL.md; the supplied material does not document how every individual Agent client exposes this nested skill.
How do you use this skill?
Prepare an Azure subscription, Cognitive Services Contributor or Owner permissions, an Azure AI Foundry project resource ID, and the Azure CLI. Run az login; optionally set PROJECT_RESOURCE_ID. Then prompt the Agent with something like: “Customize an Azure OpenAI model deployment: let me choose the version, SKU, capacity, and content-filter policy.” Before executing any phase, load references/customize-workflow.md for the full scripts and implementation details; that reference file was not included in the supplied material.
How does this skill compare with similar options?
The named alternative is preset. customize is for deliberate selection of version, SKU, capacity, and RAI policy, while preset is for quick deployment using the best available region and automatic configuration.