plan-ai-model-rollout
Installation
SKILL.md
Plan AI Model Rollout
Contract
An AI change is a versioned product release even when the interface does not change. Offline evaluation is an entry gate, not production proof. Keep model, prompt, retrieval, tools, policy, and data versions independently identifiable.
Workflow
- Define the exact change, owner, affected jobs and populations, current baseline, intended improvement, known regressions, dependencies, and rollback target.
- Record immutable versions for model or provider, prompts, retrieval indexes, tools, policies, parameters, safety controls, and evaluation set.
- Confirm offline quality, safety, latency, cost, capacity, privacy, security, and compatibility gates by relevant slice.
- Choose staged exposure: internal, shadow, replay, staff, opt-in, canary, percentage, segment, geography, or broad release. Prevent contamination where comparison matters.
- Define eligibility, routing, sticky assignment, fallbacks, timeout, partial failure, provider outage, quota, and data-residency behavior.
- Set online success, quality, safety, latency, cost, override, complaint, and incident measures with owners and observation windows.
- Predefine pause, rollback, containment, and escalation thresholds. Test rollback and state compatibility before customer exposure.
- Record approvals, release evidence, exact state, monitoring, customer communication needs, and the decision required at each expansion gate.