Data pipelines for ML are not ETL pipelines. They are product systems with customers, a spec, a quality bar, and a lifecycle — and they are increasingly the part of the AI stack that determines whether a model works in production or only on benchmarks. This kit is a structured reference for product owners managing that system: a domain primer, thirteen capability cards drawn from SAFe 6 PO practice, a self-scored maturity rubric, and a downloadable agent pack you can deploy in Claude or ChatGPT to get hands-on help with your specific pipeline.
The most consequential mistake a product owner can make in an ML data pipeline is treating it like a software delivery project — scope it, build it, ship it, close it. A labeling pipeline or a data curation system does not get to be done, because the model it feeds does not get to be done. The world keeps generating scenarios the training data has never seen.
The spec is the product. A labeling specification that defines what counts as a correctly annotated item is the closest equivalent to a software architecture document. Without a versioned, enforced spec, quality is unmeasurable and every review cycle becomes a renegotiation. The first question to ask of any pipeline is not "how much data do we have" but "how precisely have we defined what correct looks like."
If your quality measurement system is the model's downstream performance, you have measured the output and missed the input. By the time the model fails in production, the labeling decision that caused it was made months earlier.
The customers are ML engineers. Their failure modes — workarounds in training scripts, inconsistent model behaviour, evaluation gaps — are the most direct signal that the pipeline is not serving its purpose. Treating ML engineers as consumers of the pipeline rather than as co-designers is one of the most common structural errors in data platform product ownership.
Each capability maps a PO responsibility area to its common failure mode and concrete outputs. Categories reflect SAFe 6 PO practice areas and ML data pipeline specifics. Click a card for the full failure mode description, what good looks like, and what to produce.
Eight dimensions, each worth 0–2 points. 0 = not at all, 1 = partially or informally, 2 = systematically and reproducibly. Tap a cell to score yourself. Total out of 16. Honest scoring tends to matter more than a high number — the point is finding the structural gaps you did not know you had.
The agent pack is a single Markdown file containing the domain primer, all thirteen capability cards in structured form, the maturity rubric, and a set of guided prompts that walk a model through producing a prioritised pipeline health summary tailored to your specific context.
Paste it into a Claude project, a ChatGPT custom GPT, or any LLM that accepts long-context system prompts. The agent will ask about your domain, your internal customers, your pipeline's current state, and your most pressing problem — then produce a structured assessment grounded in this kit. A JSON version is also available for any integration that prefers machine-readable input.