HewlettPackard

Behavioral rules that make AI coding agents more reliable — verification, self-correction, and scope discipline.

Debugging guidance and structured report template for thorough bug investigation, evidence gathering, and root cause documentation.

Cold-audit a completed delivery — verify the spec was actually delivered, challenge performative claims, and produce a what-landed report.

Guidelines for generating self-contained HTML mockups that look professional and are easy to iterate on.

Shared implementation guidance for making minimal, correct, maintainable code changes that fit existing systems.

The doctrine for `/roadmap-review` — interactive triage of roadmap-shape drift, walked one item at a time, with paste-ready resolution phrasing for the active sizing lens.

When work would produce 2+ related specs, prefer initiative-first composition over `/design` × N siblings. Carries the canonical "multiple related specs" phrasing that both interactive triage (`/roadmap-review`) and design-time routing (`/design`) quote.

Defines the spec document structure and conventions for feature designs and bug diagnoses in the hero workflow.

Use when setting up CMF in a new or existing project. Covers creating a Python environment (conda, uv, or venv), installing cmflib (requires Python 3.9–3.11, recommended 3.10), running `cmf init` to configure a storage backend (local, S3, MinIO, SSH, or OSDF), and verifying the setup. Linux only — cmflib does not support macOS.

Use when adding CMF metadata tracking to existing Python ML pipeline code, or writing new CMF-instrumented pipeline stages from scratch. Covers the Cmf() API: create_context, create_execution, log_dataset, log_model, log_metric, commit_metrics, log_execution_metrics, and finalize(). Includes copy-paste templates for common ML pipeline stages.

Use when setting up the CMF MCP server so an AI coding assistant can query CMF metadata using natural language, or when the CMF MCP server is already running and the user wants to use it from their coding agent. Covers Docker deployment, per-agent configuration (Claude Code, GitHub Copilot, Cursor, Codex), and example prompts.

Use when querying CMF metadata programmatically — listing pipelines, stages, executions, artifacts; tracing lineage; comparing hyperparameters across runs; or exporting metadata for analysis. Uses the CmfQuery Python API or CLI listing commands.

Use when deploying the CMF Server locally with Docker Compose, or connecting to an existing shared CMF Server. Covers creating the .env file, starting the container stack, verifying the web UI, and pointing cmflib clients at the server. If a server is already running and credentials are available, skip to the connection section.

Use when working with the Common Metadata Framework (CMF) — initializing CMF in a project, instrumenting ML pipeline code, syncing metadata and artifacts with a CMF Server, querying lineage and artifacts, or setting up the CMF MCP server for AI assistant integration. Routes to the appropriate specialized skill based on the task.

Use when pushing or pulling CMF metadata and artifacts between a local environment and a CMF Server or central artifact repository (S3, MinIO, SSH, OSDF, local). Covers `cmf metadata push/pull`, `cmf artifact push/pull`, `cmf repo push/pull`, and collaborative team workflows. Assumes `cmf init` has been run.