ML AI Project Planning

Plan ML and AI work from the decision, data, risk, and operating model before choosing a model. Use this skill to decide whether ML is appropriate, define success, and create an actionable project card.

Process

1. Start from the business decision, user, prediction unit, prediction moment, available data, target, latency, cost, audit, privacy, and owner.
2. Check whether deterministic rules, search, a parser, a heuristic, or the current process can solve the problem with lower risk than ML.
3. Quantify false positive, false negative, abstain/no-decision, latency, compute, operational, and review costs.
4. Define how uncertainty is represented and when the system falls back, asks for human review, or refuses to decide.
5. Establish a baseline before proposing complex models. Do not approve large training work without a stable evaluator and baseline.
6. Choose business metrics, model metrics, segment reporting, guardrails, minimum useful improvement, and release acceptance criteria together.
7. Produce a project card or planning report with scope, assumptions, data needs, evaluation plan, phases, risks, and explicit non-goals.

Coordination

Use with ml-data-evaluation when datasets, splits, leakage, metrics, eval datasets, or error analysis are in scope. Use ml-serving-mlops-review when rollout, monitoring, retraining, rollback, model cards, or production review are in scope. For Rust/Rig/Candle implementation details, route to the existing rust-ml-llm-* skills first.

Reference Routing