Guide · Agent skill
How to use the Agent Platform Model Tuning agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Agent Platform Model Tuning skill (skill name agent-platform-tuning) gives an AI agent a step-by-step procedure for fine-tuning open models or Gemini models on Google Cloud. It is written for developers and data teams who already have a Google Cloud project and want guided help preparing data, estimating cost, launching a tuning job and monitoring it. After this guide you will know what the skill does, how to install it, how to phrase requests so it triggers, and what you need in place before it runs.
Agent Platform Model Tuninggoogle/skills · Apache-2.0 Download the skill Quick start
- 1Download and unzip the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder. Alternatively, upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
- 2Install the Python requirements. Install the packages listed in references/requirements.txt, which include the Vertex AI SDK and Cloud Storage libraries.
- 3Authenticate gcloud. Make sure the gcloud CLI is installed and signed in to the Google Cloud account that owns your project.
- 4Prepare your project and bucket. Confirm you have a Google Cloud project with billing and the tuning APIs enabled, plus an existing Cloud Storage bucket for data and outputs.
- 5Make your request. Tell the agent your project, region and whether you want an open or Gemini model, and describe the task in plain language.
- 6Review data and cost. Confirm the column mapping from the sample rows and review the cost estimate before approving the job.
- 7Monitor or cancel the job. Ask the agent to monitor the tuning job, or to cancel it using the included script if needed.
What is an agent skill and what does this one do?
An agent skill is a folder of instructions, reference files and scripts that an AI agent can load when a task matches it. This skill covers the full tuning lifecycle on Google Cloud: checking your project and region, confirming whether you want an open model or a Gemini model, preparing and validating training data, estimating cost, submitting a tuning job, monitoring it and optionally deploying the result.
- Checks your Google Cloud project and region before doing anything else
- Asks whether you want an open model or a Gemini model
- Prepares and validates JSONL training data for a Cloud Storage bucket you already own
- Estimates cost before a paid tuning job starts
- Submits, monitors and can cancel tuning jobs using included Python scripts
- Pauses to ask you before taking actions rather than acting on its own
How do you install the Agent Platform Model Tuning skill?
Download the skill ZIP from this page and unzip it. You can install it for yourself across all projects or for a single project. For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the .claude/skills/ folder inside that project.
If you use the Claude apps instead, upload the ZIP under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
What do you need before you start?
The skill expects several things to be in place. Nothing here is optional if you want the workflow to run to completion. The skill will not create a Cloud Storage bucket for you, so you need one already.
- Python with the packages listed in references/requirements.txt (Vertex AI SDK and Google Gen AI / Cloud Storage libraries)
- The gcloud CLI installed and authenticated to a Google Cloud account
- A Google Cloud project with billing and the tuning APIs enabled
- An existing Cloud Storage bucket for datasets and outputs
pip install -r references/requirements.txtHow do you trigger the skill with a natural request?
You do not need special syntax. Describe the tuning task in plain language and name your project and region. The skill is built to ask for anything missing. If you omit the project or region, it stops and asks for them before running any commands. If you have not said whether you want an open model or a Gemini model, it will ask that too.
Requests to check whether your environment is ready (APIs, IAM permissions, service agents) only need the project and region. The skill will not ask for the model category in that case.
- "I want to fine-tune an open model in my project my-project in us-central1 on my support tickets."
- "What setup do I need before I can tune a Gemini model?"
- "Estimate the cost of tuning a model on my JSONL file before I start."
- "Check whether my project is ready for model tuning."
What does a typical tuning session look like?
The agent works through a fixed decision path. It confirms project, region and model category, then checks the environment and dataset. It shows you sample rows and asks you to confirm the column mapping before continuing. For a specific purpose, it recommends three models (one open, one Gemini and one generally recommended choice) and lists pros and cons, only from its supported model catalog. After you agree to the configuration, it submits the job, monitors it and offers deployment as a final optional step.
What data format does the skill need?
Training data must be JSON Lines (.jsonl), UTF-8 encoded, and stored in a Cloud Storage bucket. The reference file describes two formats for open models. The conversational format uses a messages list with system, user and assistant roles and is recommended for chat-based models. The instruction format uses prompt and completion fields and suits base models or simple completion tasks. A separate validation file is optional.
- Conversational: one JSON object per line with a "messages" list
- Instruction: one JSON object per line with "prompt" and "completion" fields
- Example location: gs://my-bucket/train.jsonl
{"prompt": "Summarize the following text: [TEXT]", "completion": "[SUMMARY]"}What are the safety notes for scripts and network use?
This skill runs scripts and uses the network. The included scripts cover listing models, estimating cost, preparing datasets, submitting tuning jobs, monitoring them and cancelling them. Tuning jobs cost money, and the skill is designed to stop and ask you before taking actions. Review the proposed project, region, bucket and cost estimate before you approve a launch.
The skill also explicitly does not cover training a model from scratch, training outside Google's platform, or managing serving endpoints. Those belong to other separate skills.
- Confirm the project, region and bucket before approving a job
- Read the cost estimate from calculate_cost.py before launching
- Use cancel_tuning_job.py if a job should stop
What people use it for
- Fine-tuning an open model such as Gemma 3 on a set of chat transcripts in the conversational JSONL format
- Tuning a Gemini model for a coding or support chat assistant after confirming the model choice
- Estimating cost on a training file before committing to a paid tuning job
- Checking whether a project has the required APIs and permissions before a team starts work
Frequently asked questions
Is the Agent Platform Model Tuning skill free?
The skill itself is licensed under Apache-2.0. Tuning jobs run on Google Cloud and cost money, so you need billing set up on your project. The skill includes a cost estimation script so you can see an estimate before launching a job.
What do I need to install the agent-platform-tuning skill?
Download the ZIP, unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps. You also need Python with the packages in references/requirements.txt.
Does the skill create a Cloud Storage bucket for me?
No. The skill is designed to use an existing bucket that you provide and does not give setup instructions to create one.
Can the skill train a model from scratch or deploy to an endpoint?
No. The skill explicitly excludes training models from scratch or outside Google's platform. Deploying and managing serving endpoints belong to other separate skills.
Which models can the skill tune?
It works from a reference catalog of supported open models and Gemini models. The skill only recommends models listed in that catalog, and if you ask for an unlisted model it offers the closest supported alternative.
How does this compare to tuning models outside Google Cloud?
This skill is specific to Google's Agent Platform tuning service and its supported models. If you want to train or fine-tune on another platform, this skill is not the right tool, since it explicitly excludes work outside Agent Platform.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Choose this skill if you already run projects on Google Cloud, have billing and a bucket ready, and want an agent to walk you through fine-tuning an open or Gemini model with a cost check before you pay. Pick something else if you have no Google Cloud account, want to train from scratch, or need help managing deployed endpoints.
Download the skill