Create ZenML Skills repository for AI coding agents

[strickvl]

Summary

Inspired by HuggingFace Skills and their blog post on teaching Claude to fine-tune LLMs, we propose creating a zenml-skills repository that packages ZenML domain expertise for AI coding agents.

What Are Skills?

Skills are Agent Context Protocol (ACP) definitions that teach coding agents (Claude Code, OpenAI Codex, Google Gemini CLI) how to perform specialized tasks. Each skill bundles:

• Instructions (SKILL.md) - When and how to use the skill
• Scripts - Executable CLI tools
• Templates - Common configurations and patterns
• References - Troubleshooting and best practices

When a user installs a skill, their AI coding agent gains domain knowledge about ZenML workflows, enabling conversations like:

"Set up a ZenML stack for my team on AWS with Kubernetes orchestration"

"Build an LLM agent that uses RAG and deploy it with ZenML"

"Create a churn prediction pipeline with proper evaluation and model registry"

Proposed Skills

1. zenml-stack-configurator

Target users: Platform engineers setting up ZenML infrastructure

Helps configure stacks (local, cloud, hybrid), connect orchestrators, artifact stores, container registries, secrets backends, etc. Encodes best practices for production stack setup.

2. zenml-org-setup

Target users: Platform engineers managing ZenML for teams

Handles organization permissions, RBAC setup, team onboarding, service accounts, and multi-tenant configurations. Useful for enterprise deployments.

3. zenml-agent-builder

Target users: ML/AI engineers building LLM applications

Covers the full agent lifecycle: creating agent pipelines, tool integration, deployment to endpoints, and monitoring. Teaches agents how to build agents with ZenML.

4. zenml-llm-finetuning

Target users: ML engineers fine-tuning LLMs

Extends/builds on top of the HuggingFace hf-llm-trainer skill but wraps training in ZenML pipelines. Adds experiment tracking, artifact versioning, model registry integration, and reproducibility.

5. zenml-mlops-templates (or split into sub-skills)

Target users: Data scientists and ML engineers

Traditional MLOps use cases with proper ZenML patterns:

• Churn prediction
• Recommendation systems
• Time-series forecasting
• Classification/regression pipelines

Each could be a targeted sub-skill for specific use cases rather than one generic "pipeline builder" (specificity = better agent guidance).

Technical Implementation

Following the HuggingFace pattern:

zenml-skills/
├── README.md
├── AGENTS.md                    # Auto-generated for Codex
├── gemini-extension.json        # Gemini CLI support
├── scripts/
│   ├── generate_agents.py       # Parses SKILL.md → AGENTS.md
│   └── AGENTS_TEMPLATE.md
├── .github/workflows/
│   └── generate-agents.yml      # CI validation
└── [skill-folders]/
    ├── zenml-stack-configurator/
    │   ├── SKILL.md
    │   ├── scripts/
    │   └── templates/
    ├── zenml-org-setup/
    ├── zenml-agent-builder/
    ├── zenml-llm-finetuning/
    └── zenml-mlops-templates/

Multi-Agent Support

Agent	Integration
Claude Code	/plugin install skill@zenml-skills
OpenAI Codex	Reads AGENTS.md automatically
Gemini CLI	Uses gemini-extension.json
Cursor, Windsurf	Planned by HF, we'd get for free

Key Patterns to Adopt

1. YAML frontmatter in SKILL.md for metadata parsing
2. Auto-generated AGENTS.md from single source of truth
3. CI validation to keep agent index in sync
4. PEP 723 inline dependencies for self-contained scripts
5. Validation-first workflows (check before expensive operations)

Benefits

1. Lower barrier to entry - Users can describe what they want in plain English
2. Encoded best practices - Skills teach agents the "right way" to do things
3. Cross-agent compatibility - Works with Claude Code, Codex, Gemini CLI
4. Community contribution - Users can submit their own skill patterns
5. Marketing/awareness - Positions ZenML as AI-agent-native

Open Questions

• Should this be a separate zenml-io/zenml-skills repo or live within the main repo?
• Which skill should we build first as a proof-of-concept?
• Should we coordinate with HuggingFace on the zenml-llm-finetuning skill since it extends their work?

References

• HuggingFace Skills repo
• Blog: We Got Claude to Fine-Tune an Open Source LLM
• Simon Willison on Claude Skills