MLOps Stack

Authored by Ayush Singh | Production-grade MLOps skills for Claude Code

A collection of opinionated, human-in-the-loop Claude Code skills that guide you through building production ML systems from scratch. Not templates. Not tutorials. Co-pilots that adapt to YOUR problem, YOUR data, and YOUR constraints.

---

Skills

/mlops-tabular — Tabular Supervised Learning

The flagship skill. Takes you end-to-end from business problem to deployed, monitored ML pipeline on tabular data.

What it does:

• Phase 1: Problem Framing — Translates your business problem into an ML formulation. Defines metrics, error costs, success criteria.
• Phase 2: MLOps Architecture Design — Designs the complete MLOps stack: data plan, feature engineering plan, training plan, deployment plan, monitoring plan, versioning plan, and selects specific framework components.
• Phase 3: Implementation — Builds every pipeline step incrementally. Explains every decision in plain language with deep technical substance. Asks permission before writing every file.
• Phase 4: Ship — Final verification, documentation, GitHub push.

Key properties:

• Human-in-the-loop at every business decision
• Teaches as it builds — simple words, PhD-level depth
• Fetches current library docs before generating code (no stale API errors)
• Framework-extensible — ZenML by default, contributors can add Prefect, Airflow, etc.
• Covers classification AND regression on tabular data

Supported frameworks:

• ZenML (built-in, full support with 15+ component types)
• Others can be added via references/tooling/ — see contributor guide

/mlops-code-review — Code Review (SE + ML)

Full software engineering and ML-specific code review. Three-pass protocol: general code quality, ML-specific issues (leakage, skew, reproducibility), then severity triage.

• Reviews against Google Python Style Guide, SOLID, OWASP, and ML best practices
• Detects data leakage, training-serving skew, and pipeline smells in code
• Produces structured findings (Critical / Major / Minor) with fix proposals
• Cross-cutting: invoke at any phase to audit code quality

/mlops-system-design — System Design (General + ML)

System design covering distributed systems and ML infrastructure. Six-step protocol: requirements, estimation, high-level design, deep-dive, tradeoffs, operational concerns.

• General: API design, databases, distributed systems, scalability, reliability/SRE, microservices, messaging
• ML: model serving architecture, feature store design, ML platform design, ML infrastructure patterns
• Sources: DDIA, Google SRE Book, Chip Huyen, Stanford CS329S
• Produces system_design.md with ASCII diagrams, tradeoff tables, and cost estimates

/mlops-agent-workflow — Anti-Slop Agentic Engineering

Sets up disciplined workflows for AI coding agents. Five pillars: RPI workflow, context management, quality gates, agent isolation, anti-slop patterns.

• Research-Plan-Implement (RPI) workflow with context compaction
• Smart zone / dumb zone context management
• Quality gates: strict linting, type checking, 100% test pass rate, pre-commit hooks
• Per-agent isolation with git worktrees
• Framework-agnostic: works for any software project, not just ML

Specialized Deep-Dive Skills

Skill	Focus
/mlops-problem-framing	Business problem → ML formulation
/mlops-architecture	Full MLOps pipeline design (10 stages, 5 pipelines)
/mlops-data-and-features	Data loading, EDA, preprocessing, feature engineering
/mlops-training-eval	Experiment tracking, training, evaluation
/mlops-deploy-monitor	Deployment, monitoring, drift detection, production hardening

---

Installation

# Clone the repo
git clone https://github.com/ayush488-glitch/mlops-stack.git

# Symlink skills into Claude Code
ln -sfn $(pwd)/mlops-stack/skills/mlops-tabular ~/.claude/skills/mlops-tabular
ln -sfn $(pwd)/mlops-stack/skills/mlops-code-review ~/.claude/skills/mlops-code-review
ln -sfn $(pwd)/mlops-stack/skills/mlops-system-design ~/.claude/skills/mlops-system-design
ln -sfn $(pwd)/mlops-stack/skills/mlops-agent-workflow ~/.claude/skills/mlops-agent-workflow
ln -sfn $(pwd)/mlops-stack/skills/mlops-problem-framing ~/.claude/skills/mlops-problem-framing
ln -sfn $(pwd)/mlops-stack/skills/mlops-architecture ~/.claude/skills/mlops-architecture
ln -sfn $(pwd)/mlops-stack/skills/mlops-data-and-features ~/.claude/skills/mlops-data-and-features
ln -sfn $(pwd)/mlops-stack/skills/mlops-training-eval ~/.claude/skills/mlops-training-eval
ln -sfn $(pwd)/mlops-stack/skills/mlops-deploy-monitor ~/.claude/skills/mlops-deploy-monitor

Then invoke with /mlops-tabular (recommended entry point) or any skill directly.

---

Context7 MCP Setup (Required for Live Docs)

The skills use Context7 to fetch live, up-to-date documentation for libraries like ZenML, scikit-learn, MLflow, pandas, etc. Without it, the agent relies on its training data which may have stale API information.

Warning: If Context7 is not configured, the skills will still work but may generate code against outdated APIs. You'll see a warning at the start of each session.

Step 1: Get a Context7 API Key

1. Go to context7.com and sign up
2. Copy your API key (starts with ctx7sk-...)

Step 2: Set the API Key as an Environment Variable

macOS / Linux (zsh):

echo 'export CONTEXT7_API_KEY="your-api-key-here"' >> ~/.zshrc
source ~/.zshrc

macOS / Linux (bash):

echo 'export CONTEXT7_API_KEY="your-api-key-here"' >> ~/.bashrc
source ~/.bashrc

Windows (PowerShell — permanent):

[System.Environment]::SetEnvironmentVariable("CONTEXT7_API_KEY", "your-api-key-here", "User")

Then restart your terminal.

Windows (Command Prompt — permanent):

setx CONTEXT7_API_KEY "your-api-key-here"

Then restart your terminal.

Step 3: Add the MCP Server to Claude Code

Create or edit ~/.claude/.mcp.json (global — works across all projects):

macOS / Linux:

cat > ~/.claude/.mcp.json << 'EOF'
{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"],
      "env": {
        "DEFAULT_MINIMUM_TOKENS": "10000",
        "CONTEXT7_API_KEY": "${CONTEXT7_API_KEY}"
      }
    }
  }
}
EOF

Windows (PowerShell):

$mcpConfig = @'
{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp@latest"],
      "env": {
        "DEFAULT_MINIMUM_TOKENS": "10000",
        "CONTEXT7_API_KEY": "${CONTEXT7_API_KEY}"
      }
    }
  }
}
'@
$mcpConfig | Out-File -FilePath "$env:USERPROFILE\.claude\.mcp.json" -Encoding utf8

Step 4: Restart Claude Code

Restart Claude Code to load the MCP server. You should see context7 in the available tools.

Verifying It Works

In a Claude Code session, you can test it:

Use Context7 to fetch the latest ZenML pipeline documentation

If Context7 is working, it will fetch live docs. If not, you'll get a tool-not-found error — check your .mcp.json path and API key.

---

Project Structure

mlops-stack/
└── skills/
    ├── mlops-tabular/                 # Flagship orchestrator
    │   ├── SKILL.md                   # Skill definition
    │   ├── SKILL.md.tmpl              # Template source (edit this)
    │   ├── gen-skill.sh               # Template generator
    │   └── references/
    │       ├── capabilities/          # 19 MLOps concept guides
    │       ├── tooling/zenml/         # 6 ZenML implementation guides
    │       └── examples/              # Reference timelines & case studies
    ├── mlops-code-review/             # Code review (SE + ML)
    │   ├── SKILL.md
    │   └── references/capabilities/   # 10 review reference guides
    ├── mlops-system-design/           # System design (general + ML)
    │   ├── SKILL.md
    │   └── references/capabilities/   # 10 system design reference guides
    ├── mlops-agent-workflow/          # Anti-slop agentic engineering
    │   ├── SKILL.md
    │   └── references/capabilities/   # 5 workflow reference guides
    ├── mlops-problem-framing/         # Problem framing deep-dive
    ├── mlops-architecture/            # Architecture design deep-dive
    ├── mlops-data-and-features/       # Data & features deep-dive
    ├── mlops-training-eval/           # Training & evaluation deep-dive
    └── mlops-deploy-monitor/          # Deployment & monitoring deep-dive

---

Adding New Skills

This repo is designed for multiple skills. To add a new one:

1. Create skills/<skill-name>/
2. Add SKILL.md with frontmatter (name, description, allowed-tools)
3. Add references/ for knowledge base
4. Symlink to ~/.claude/skills/<skill-name>
5. Update this README

---

Contributing

Contributions welcome — especially:

• New orchestration framework support (Prefect, Airflow, Kubeflow, etc.) in references/tooling/
• New skill types (time series, NLP, computer vision pipelines)
• Bug fixes and capability reference improvements

---

Author

Ayush Singh — MLOps Engineer

Built with the conviction that AI should handle the engineering while humans bring the judgment. Every skill in this stack is designed to produce real, deployable, production-grade ML systems — not AI slop.

---

License

MIT