How to use the Agent Platform Tuning Management agent skill

What does the skill make the agent do?

Agent skills are instruction packages that teach an AI agent how to perform a specific task. This skill covers one area: managing GenAI tuning jobs, which are ongoing runs that customize an AI model with extra training data. It uses the Agent Platform Python SDK to do three things: list tuning jobs, get details for a specific job, and cancel a job that is running too long.

Listing and checking happen immediately. Cancelling is treated differently and requires a confirmation step first.

• List tuning jobs in a project and region, showing each job's name, base model and state.
• Get details for one job by its tuning job ID, including its tuned model display name.
• Cancel a running tuning job, but only after you explicitly confirm.

Install the skill from the ZIP file

Download the ZIP file for this skill from this page and unzip it. The skill is a folder that contains a file named SKILL.md. Place that folder in one of two locations, depending on how you want to use it.

For a personal setup available across your projects, put it in ~/.claude/skills/. For a single project, put it in the .claude/skills/ folder inside that project.

If you use the Claude apps instead, you can upload the skill under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/agent-platform-tuning-management/
• Project only: .claude/skills/agent-platform-tuning-management/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the skill

Set up Google Cloud and Python first

Before the agent runs any snippet, the environment must be ready. The skill requires a Google Cloud account, the gcloud CLI installed and authenticated, and Python 3. You also need your Google Cloud project ID, your region, and permission to view or cancel tuning jobs.

Authenticate with both commands below. The first signs in your account; the second sets up Application Default Credentials, which the Python code uses.

The skill checks for the google-cloud-aiplatform library and installs it only if it is missing. It advises against creating a virtual environment, so it runs with your plain python3.

• Google Cloud account with the gcloud CLI installed
• Python 3
• Google Cloud project ID and region
• Permission to view tuning jobs (and to cancel them, if you plan to)

Trigger the skill with natural requests

You do not need special syntax. Describe what you want in plain language and the agent decides whether this skill applies. The skill is meant for requests about checking the status of tuning runs, finding an active tuning job, or stopping a job that is running too long.

If the agent does not know your project ID or region, it will ask you for them. It checks your gcloud configuration for this information before asking, and it will not search other regions on its own.

• "What tuning jobs do I have running in my project?"
• "What is the status of tuning job 1234567890123456789?" (replace with your own job ID)
• "Cancel tuning job 1234567890123456789 in europe-west4." (the agent will ask for confirmation first)

Understand the cancellation confirmation

Cancelling is a destructive, interruptive action. Before it runs, the agent shows a dry-run confirmation card listing the target resource, the exact command or Python code it would run, and the expected effect. The expected effect is that the ongoing tuning job stops and any in-progress training is halted and cannot be resumed.

You must then reply with explicit typed confirmation, such as "I confirm" or "Yes, cancel it". The agent will not cancel in the same turn as showing the card, even if you already typed a confirmation in your first message. If you correct the job ID, it shows a new card for that job and waits again.

• Target resource is shown in full, for example projects/<PROJECT_ID>/locations/<REGION>/tuningJobs/<JOB_ID>.
• The exact command or code is shown before anything runs.
• Cancellation runs only after your confirmation in a separate turn.

Handle errors and missing resources

If a snippet fails with an error such as 403 Permission Denied, 404 Not Found, INVALID_ARGUMENT, or a missing project or job, the agent will tell you that the project or job does not exist or cannot be accessed. It will ask for a valid Project ID or Job ID and stop running scripts until you reply. It does not retry in a loop.

Check that your project ID, region and job ID are correct, and that your account has the needed permissions.