MCPGateway: Skill for Codex

MCAF: ML/AI Delivery

Trigger On

• the repo contains model training, inference, experimentation, or data-science workflow
• ML work needs explicit process, testing, or responsible-AI guidance
• delivery discussion is mixing product, data, and model concerns

Value

• produce a concrete project delta: code, docs, config, tests, CI, or review artifact
• reduce ambiguity through explicit planning, verification, and final validation skills
• leave reusable project context so future tasks are faster and safer

Do Not Use For

• generic software delivery with no ML or data-science component
• loading all ML references when only one stage is active

Inputs

• the current ML stage: framing, data exploration, experimentation, training, inference, or operations
• product assumptions, data assumptions, and model assumptions
• current verification and responsible-AI expectations

Quick Start

1. Read the nearest AGENTS.md and confirm scope and constraints.
2. Run this skill's Workflow through the Ralph Loop until outcomes are acceptable.
3. Return the Required Result Format with concrete artifacts and verification evidence.

Workflow

1. Separate product assumptions, data assumptions, and model assumptions.
2. Keep experimentation traceable and testable.
3. Treat responsible AI, data quality, and ML-specific verification as first-class requirements.
4. Load only the references that match the current ML stage.

Deliver

• clearer ML/AI delivery guidance
• better links between data, experimentation, verification, and responsible AI
• docs that match how the ML system is built and validated

Validate

• the active ML stage is explicit
• experimentation and evaluation are traceable
• responsible-AI and data-quality requirements are not bolted on at the end

Ralph Loop

Use the Ralph Loop for every task, including docs, architecture, testing, and tooling work.

1. Plan first (mandatory):
   • analyze current state
   • define target outcome, constraints, and risks
   • write a detailed execution plan
   • list final validation skills to run at the end, with order and reason
2. Execute one planned step and produce a concrete delta.
3. Review the result and capture findings with actionable next fixes.
4. Apply fixes in small batches and rerun the relevant checks or review steps.
5. Update the plan after each iteration.
6. Repeat until outcomes are acceptable or only explicit exceptions remain.
7. If a dependency is missing, bootstrap it or return status: not_applicable with explicit reason and fallback path.