Use W&B Skills

Install W&B Skills to teach your coding agent how to train models, build agents, and analyze experiments using W&B’s AI development platform.

W&B Skills are reusable instruction sets that teach coding agents how to use W&B effectively. Instead of manually guiding your agent through W&B APIs and best practices, install Skills so that the agent can work with experiment tracking, tracing, evaluations, and monitoring on its own.

Skills work with several major coding agents, including:

• Claude Code
• Codex
• Cursor
• GitHub Copilot
• Gemini CLI

For a full list of supported agents, see the W&B Skills CLI documentation.

W&B Skills capabilities

Skills covers both the W&B Models SDK (training runs, metrics, artifacts, sweeps) and the Weave SDK (traces, evaluations, scorers). Includes helper libraries, reference docs, and data analysis patterns.

Workflow	Capabilities
Model training	Log metrics and rich media during training and fine-tuning. Track and compare experiments. Analyze runs and results, such as loss curves and accuracy scores. Tune hyperparameters.
Agent building	Trace agentic AI applications. Analyze traces and classify failure modes. Evaluate models and agents with labeled datasets. Run online evaluations for production monitoring.

Prerequisites

Skills requires the following:

• Node.js (for the npx command).
• A W&B API key. Create one at wandb.ai/authorize and then set it as an environment variable:
  export WANDB_API_KEY=<your-api-key>
  
• (Optional) Set your W&B project name as a WANDB_PROJECT environment variable. This allows your agent to target the correct W&B project without you specifying it each time.

Install W&B Skills

To install W&B Skills globally, run the following command with the --global flag:

npx skills add wandb/skills --skill '*' --yes --global

To install Skills for a specific project, run the following command from your project directory:

npx skills add wandb/skills --skill '*' --yes

You can also install Skills for specific agents using the --agent flag:

npx skills add wandb/skills --agent claude-code --skill '*' --yes --global

For list of --agent and --skill options, see the skills CLI documentation.

Use W&B Skills

Once installed, you can ask the agent to perform W&B-related tasks for your project. The following example prompts demonstrate some of the tasks your agent can do with W&B Skills:

• “Log training metrics for my PyTorch model to W&B.”
• “Analyze the loss curves for my last 10 runs and identify the best performing configuration.”
• “Trace my LangChain agent and log the results to Weave.”
• “Run an evaluation on my agent using the test dataset and summarize the results.”
• “Find the failure modes in my last evaluation and classify them.”
• “Compare the configs of run A and run B and show me the differences.”

Usage tips

Skills performs better when you use more specific queries versus broader open-ended questions. The following table provides some recommended example prompts versus prompts that are too vague.

Recommended	Not recommended
”What is the final validation loss for my last 5 runs?"	"How is my model doing?"
"Summarize the token usage across my last 10 traces."	"Show me all my traces."
"Compare the configs of run A and run B."	"What are my best runs?"
"What eval had the highest F1 score?"	"How are my evaluations going?”