Turn codebase evidence into a sequenced architecture refactor plan with invariants, compatibility steps, tests, rollback, and verification.
Architecture trouble usually shows up as one small product change touching six places, a shared table with no clear owner, or a deploy that cannot be rolled back safely. This skill maps that friction to source evidence before it recommends a new boundary.
The output is an implementation plan, not a folder diagram. It freezes behavior and data invariants, compares credible options, and breaks the migration into testable slices with compatibility, verification, stop, and rollback conditions. Use it to untangle a subsystem, split ownership, replace a shared service, or review a rewrite proposal without pretending the repository proves more than it does.
Input
Review our order-processing module. Pricing rules, payment retries, fulfillment events, and admin overrides share one service and one database schema. A new pricing rule touches seven files. Produce a refactor plan that preserves the public API and lets us ship in slices.
Output
A source-backed current-state map, behavior and data invariants, option comparison, chosen ownership boundary, characterization tests, compatibility adapter, sequenced migration slices, verification gates, rollback conditions, and removal criteria for the old path.
Input
Plan the migration of customer notification preferences from the account service database to the notification service. Three clients read the current endpoint, writes can race, and we cannot take downtime. We have no agreed reconciliation metric yet.
Provide the repository or relevant source files, the hotspot or recurring change that hurts, and any release, compatibility, data, ownership, or uptime constraints. Ask the agent to map the current boundary, freeze invariants, compare target options, and return a sequenced migration plan. If evidence is incomplete, the skill returns a bounded plan and an explicit investigation list instead of guessing.
Repository access or the relevant source files, tests, schemas, interfaces, architecture notes, and ADRs.
A named hotspot, recurring difficult change, proposed target, or rewrite plan that needs review.
Known compatibility, release, data, deployment, ownership, uptime, or supported-client constraints. Unknowns can remain explicit.
SKILL.md; architecture review rubric; decision-ready refactor plan template; Agent Skills interface metadata
No reviews yet.
Output
An expand-backfill-validate-cutover-cleanup plan that names the source of truth, idempotency, dual-read or dual-write tradeoffs, consumer compatibility, reconciliation evidence, the decision needed to set a drift threshold, stop conditions, and the forward-fix boundary.
Input
The team wants to rewrite our reporting engine as microservices because the current code is old and hard to test. Review the proposal from the attached source tree and ADRs. Recommend the smallest safe plan, including a no-refactor verdict if the evidence does not support the rewrite.
Output
A separation of verified friction from style preferences, scored candidates, smaller repair and staged-boundary options, anti-goals, a recommendation tied to change scenarios, and either a reversible migration plan or an evidence-backed decision not to refactor.
Creator
Sselvro21
