Claude Code Databricks ML Toolkit

A collection of Claude Code skills, agents, and tools for running ML experiments on Databricks. If you use Claude Code for machine learning and your compute is on Databricks, this toolkit gives you a complete workflow out of the box.

What's Inside

Status Bar (scripts/context-bar.sh)

A rich status line that replaces Claude Code's default, showing:

• Model name + effort level
• Current directory and git branch
• Uncommitted file count (or filename if just one)
• Git sync status (ahead/behind upstream, last fetch time)
• Context window usage as a color-coded progress bar
• Rate limit usage (5-hour and 7-day)
• Your last message (truncated to fit)

Add a screenshot to assets/ and uncomment the line below:

Skills (11)

Databricks ML Workflow

Skill	Trigger phrases	Description
/run-on-databricks	"run on Databricks", "train on GPU", "submit job"	Submit a training job to the Databricks GPU cluster, wait for results, pull MLflow metrics back locally
/run-training-on-databricks	"run training on databricks"	Lighter alternative: submit training and pull results
/check-results-on-databricks	"check results", "how did it do"	Fetch and analyze MLflow experiment results
/compare-runs	"compare runs", "best model so far", "what improved"	Rank experiment runs by metric, analyze what helped/hurt, recommend next steps
/init-databricks-ml	"initialize project", "add databricks"	Scaffold a complete Databricks + MLflow + Claude Code project from scratch
/iterate	"try X", "implement Y and test it", "iterate on the model"	Full cycle: implement a change, run on Databricks, compare results, suggest next step
/train-local	"test locally", "smoke test", "debug training"	Quick local CPU/MPS training for fast iteration before submitting to Databricks
/explore-data	"look at the data", "check for missing values"	Dataset EDA with distribution plots, correlation matrices, data quality checks
/research-papers	"find papers on X", "state of the art for Y"	Search recent ML papers, extract key findings, produce actionable recommendations

General Purpose

Skill	Description
/explain-code	Explain code with ASCII diagrams and analogies
/handoff	Write a context handoff document so the next agent can pick up where you left off

Agents (3)

Skills that involve heavy lifting fork work to specialized agents:

Agent	Used by	Role
experiment-runner	/run-on-databricks	Submits Databricks jobs and fetches results
data-analyst	/explore-data	Read-only EDA with pandas/matplotlib
research-agent	/research-papers	Autonomous web search for ML papers and techniques

Commands (1)

Command	Description
/commit	Create a git commit following Conventional Commits 1.0.0

How the Skills Work Together

/explore-data          Understand your dataset
       |
/train-local           Smoke test locally (CPU)
       |
/run-on-databricks     Submit to Databricks GPU cluster
       |
/check-results         Pull MLflow metrics
       |
/compare-runs          Rank runs, find what works
       |
/iterate               Implement next idea and repeat
       |
/research-papers       Search for better approaches when stuck

Installation

Quick Install

git clone https://github.com/duonginspace/claude-code-databricks-ml.git
cd claude-code-databricks-ml
bash setup.sh

The setup script copies skills, agents, commands, and the status bar script to ~/.claude/. It will prompt before overwriting existing files. Use --force to skip prompts, or --symlink to create symlinks instead of copies.

Manual Install

Copy individual files to the corresponding ~/.claude/ directories:

# Example: install just one skill
cp -r skills/run-on-databricks ~/.claude/skills/

# Example: install just the status bar
cp scripts/context-bar.sh ~/.claude/scripts/
chmod +x ~/.claude/scripts/context-bar.sh

Enable the Status Bar

Add to your ~/.claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/scripts/context-bar.sh"
  }
}

See settings/settings-example.json for a complete example.

Prerequisites

• Claude Code CLI
• A Databricks workspace with a GPU cluster (or use ephemeral job clusters)
• databricks-mcp for MCP integration
• Python 3.10+ and uv
• jq (used by the status bar script)

Starting a New Project

Run /init-databricks-ml in any directory to scaffold:

• scripts/submit_to_databricks.py and scripts/pull_results_on_databricks.py
• Makefile with make train, make pull, make results, make check
• .env.example with Databricks/MLflow config
• .mcp.json for the Databricks MCP server
• Project-level skills, agents, and permission settings
• CLAUDE.md with workflow documentation

This is the fastest way to go from zero to a working Databricks ML project with Claude Code.

License

MIT