Generate a phased migration execution plan for Azure OpenAI model transitions. Covers assessment, testing, pilot, gradual rollout, validation, and cleanup phases with checklists, gate criteria, risk registers, and rollback strategies. USE FOR: migration plan, project plan, rollout plan, migration timeline, migration checklist, phased migration, canary deployment, rollback strategy, migration risk, cutover plan, migration steps, migration template. DO NOT USE FOR: choosing a target model (use aoai-model-lifecycle), code-level API changes (use aoai-model-migration), evaluation setup (use aoai-migration-evaluation).
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Generate structured, phased migration execution plans for Azure OpenAI model transitions. Use this skill when the user already knows the source and target models and needs the operational rollout plan: scope, owners, phases, gates, risks, rollback criteria, and cleanup activities.
When to Use
Creating a project plan for an Azure OpenAI model migration
Building a rollout checklist for production model cutover
Planning a phased canary or blue-green migration
Translating retirement deadlines into an execution timeline
Preparing stakeholder-ready migration plans with go/no-go gates
Defining rollback criteria before production rollout
Coordinating app, platform, operations, and business owners during migration
Decision Logic
Choose the migration plan shape based on time remaining before the source model retirement date.
Time until retirement
Planning mode
How to adapt the plan
Less than 30 days
Accelerated
Compress phases, run assessment and testing in parallel where possible, shorten approval windows, and skip the pilot only if risk is understood and rollback is proven.
30-90 days
Standard
Use the full six phases with abbreviated timelines, one pilot window, and shorter document review cycles.
More than 90 days
Full
Use the full six phases with comfortable buffers, broader stakeholder review, multiple evaluation rounds, and time for prompt tuning or code remediation.
📝 Note: If the source model has already retired or auto-upgrade is imminent, switch immediately to an accelerated plan and treat rollback readiness as mandatory.
What to Produce
When using this skill, generate a migration plan document with:
A clear source model, target model, deadline, and owner
A six-phase plan with tasks, dates, owners, dependencies, and gate criteria
A risk register with mitigations and contingency actions
A cost impact table comparing current and projected run rates
A rollback strategy with explicit triggers and verification steps
A final decision summary with next milestone and blockers
Goal: Validate feasibility, confirm capacity, identify all impacted deployments, and estimate cost impact.
Checklist
Confirm the retirement date and replacement path
Inventory all Azure OpenAI deployments using the source model
Identify every application, environment, workflow, and region using the deployment
Confirm quota, capacity, and regional availability for the candidate model
Identify required code, prompt, configuration, or infrastructure changes
Estimate expected cost impact and budget implications
Identify owners, approvers, support contacts, and change windows
Decide whether the migration can use pilot plus progressive rollout or needs an accelerated path
Outputs
Deployment inventory
Scope and dependency list
Initial cost impact estimate
Draft timeline and owner matrix
Gate — Go/No-Go
Proceed only when feasibility is confirmed, candidate capacity is available, and stakeholders agree on the migration path.
Phase 2: Testing
Goal: Prove that the target model meets quality, compatibility, and performance expectations in non-production.
Checklist
Create a test deployment for the candidate model
Run the evaluation suite against representative scenarios
Compare quality metrics for source versus target
Test prompt compatibility and required prompt updates
Validate parameter compatibility and API behavior
Run load and latency tests under expected traffic patterns
Confirm monitoring, alerting, and dashboards are ready for pilot
Record known issues, accepted deviations, and follow-up actions
Outputs
Test report and comparison summary
Prompt or configuration adjustments
Performance and reliability results
Acceptance criteria baseline for pilot
Gate — Quality and Performance
Proceed only when quality, reliability, and latency meet acceptance criteria for the workload.
💡 Tip: Use the sibling aoai-migration-evaluation skill to define the evaluation approach, datasets, and quality thresholds used in this phase.
Phase 3: Pilot
Goal: Validate the candidate model on a small slice of real production traffic before full rollout.
Checklist
Route 5-10% of production traffic to the candidate deployment
Monitor error rates, latency, safety signals, and user feedback
Keep the source deployment fully available for immediate rollback
Run the pilot for a minimum of 7 days unless an accelerated plan is required
Review support tickets, product feedback, and business KPIs
Compare pilot behavior with baseline production metrics
Hold a pilot review meeting with go/no-go stakeholders
Outputs
Pilot scorecard
Production-observed issue log
Updated risk register
Recommendation to expand, hold, or roll back
Gate — Pilot Stability
Proceed only when pilot metrics are stable, within agreed tolerance, and no unresolved high-severity issues remain.
⚠️ Important: In an accelerated migration, skipping the pilot is a risk decision, not a default. Only skip it if rollback is proven, monitoring is strong, and the retirement deadline leaves no safe alternative.
Phase 4: Migration
Goal: Gradually shift all traffic to the target deployment while preserving rollback capability.
Checklist
Shift traffic to 25% and monitor for 24-48 hours
Shift traffic to 50% and monitor for 24-48 hours
Shift traffic to 75% and monitor for 24-48 hours
Shift traffic to 100% and continue heightened monitoring
Keep the source deployment active until validation is complete
Reconfirm incident response contacts at each increment
Pause progression immediately if rollback criteria are met
Outputs
Rollout checkpoint log
Traffic split record
Incident and mitigation log
Final cutover timestamp
Gate — Controlled Cutover
Proceed to the next traffic increment only when the current increment is stable and within tolerance.
Phase 5: Validation
Goal: Confirm the migration is complete, stable, and accepted by technical and business stakeholders.
Checklist
Verify all production traffic is on the target deployment
Run the full evaluation suite after cutover
Confirm error rate, latency, and safety metrics are within expected range
Verify projected versus actual cost impact is acceptable
Confirm no hidden dependencies still reference the source deployment
Collect stakeholder sign-off from engineering, operations, and business owners
Outputs
Post-cutover validation report
Final cost review
Stakeholder sign-off record
Decision to decommission source deployment
Gate — Sign-Off
Proceed only when all traffic is stable on the target deployment, post-cutover checks pass, and stakeholders approve closure.
Phase 6: Cleanup
Goal: Remove obsolete assets and capture lessons learned.
Checklist
Decommission the source deployment after the rollback window closes
Archive migration evidence, reports, and approvals
Update runbooks, dashboards, and operational documentation
Update infrastructure-as-code, environment variables, and inventory records
Close the migration ticket or change record
Document lessons learned for the next model transition
Outputs
Updated runbooks and inventory
Archived migration package
Closed change record
Retrospective notes
Gate — Closure
Close the migration only after source assets are retired, documentation is updated, and ownership is clear for steady-state operations.
Timeline Guidance
Use the six phases in all modes, but compress them differently.
Phase
Accelerated (<30 days)
Standard (30-90 days)
Full (>90 days)
Assessment
1-3 days
3-7 days
1-2 weeks
Testing
3-7 days
1-2 weeks
2-4 weeks
Pilot
0-3 days or skipped if necessary
1 week
1-2 weeks
Migration
2-5 days
1-2 weeks
2-3 weeks
Validation
1-3 days
3-5 days
1 week
Cleanup
1-2 days
2-5 days
1 week
📝 Note: These are planning ranges, not guarantees. Adapt them to business criticality, approval cycles, and regional deployment complexity.
Risk Register Template
Populate and maintain this table throughout the migration.
Risk
Likelihood
Impact
Mitigation
Quality regression in production responses
Medium
High
Run representative evaluations early, pilot on real traffic, and define rollback triggers before cutover.
Capacity shortage or quota unavailability
Medium
High
Validate quota and regional availability during assessment; request increases before testing completes.
Prompt incompatibility with the target model
Medium
Medium
Test prompts and parameters in non-production; isolate and version prompt changes.
Latency increase during peak traffic
Medium
High
Run load tests, monitor pilot latency, and pause rollout at the first unstable checkpoint.
Cost overrun after migration
Medium
Medium
Estimate projected usage up front, compare pilot cost signals, and verify actual run rate after cutover.
Cost Impact Template
Do not hard-code pricing in the plan. Use placeholders and attach the latest pricing or FinOps estimate separately.
Cost Category
Current State
Projected State
Variance
Notes
Model deployment / token usage
TBD
TBD
TBD
Compare current source-model spend versus expected target-model spend.
Supporting infrastructure
TBD
TBD
TBD
Include any changes to caching, gateways, monitoring, or storage.
Evaluation and test environments
TBD
TBD
TBD
Include temporary test or pilot environments if relevant.
Operational support
TBD
TBD
TBD
Include added on-call coverage or change-management effort if tracked.
Total
TBD
TBD
TBD
Validate after pilot and after full rollout.
Rollback Strategy
Define rollback before any production traffic moves.
Rollback Triggers
Roll back immediately if any of the following occur:
Error rate exceeds the agreed tolerance for the workload
Latency materially exceeds the approved range during rollout or pilot
Quality regression is observed in user-facing outputs or evaluation results
Safety, compliance, or policy violations increase after cutover
A Sev 1 or Sev 2 production incident is attributed to the migration
Business stakeholders reject pilot or rollout outcomes
Rollback Procedure
Stop any further traffic progression.
Route traffic back to the source deployment.
Verify the source deployment is healthy and serving expected volume.
Freeze candidate rollout changes until root cause is understood.
Capture logs, metrics, prompts, and traffic split state for analysis.
Open or update the incident/change record.
Decide whether to retry, remediate, or select a different target model.
Rollback Verification Checklist
Traffic has returned to the source deployment
Error rate and latency have returned to baseline
Customer-facing impact has stabilized
Incident owner and remediation plan are assigned
Next decision point is scheduled
Cross-References
Use sibling skills together rather than overloading this one.
Use aoai-model-lifecycle for retirement dates, lifecycle status, and target model selection
Use aoai-model-migration for code-level API changes, SDK updates, and parameter adaptation
Use aoai-migration-evaluation for evaluation methodology, golden datasets, A/B comparisons, and acceptance criteria
Plan Generation Guidance
When drafting the final plan for a user:
Prefer a phased checklist over narrative paragraphs
Include named owners for each phase when the user provides teams or roles
Convert vague requests into concrete milestones, gates, and rollback triggers
Call out assumptions explicitly if the user did not provide traffic, region, or deadline details
Keep the source deployment active until validation is complete
Use fictional examples such as Contoso if an example organization is needed
Avoid specific benchmark numbers or pricing claims unless the user supplies them
Must Not
Choose a target model without validating retirement status and replacement options
Treat testing as optional before production rollout
Decommission the source deployment before rollback readiness is confirmed
Collapse assessment, testing, and rollout into one undocumented change window unless an emergency retirement forces it
Publish pricing or benchmark claims without fresh authoritative sources
Use real customer names or identifiable production details in the plan
Repository Resources
💡 Tip: This skill provides rollout planning guidance. For the latest model dates, migration paths, and evaluation details, always check the repo documentation — it is updated more frequently than this skill.
