# MCP and evidence path

Use this reference when deciding how to ground `azure-rbac-review` guidance.

## Evidence order

1. Microsoft Learn documentation through the user's configured documentation MCP for documented Azure behavior.
2. Sampled read-only Azure evidence when the user has configured it and current-state confirmation is necessary.
3. Sanitized user-provided evidence when no read-only evidence path is available.
4. Clearly labeled inference when evidence is incomplete.

## Boundaries

- Documentation evidence does not prove the user's tenant, subscription, RBAC, quotas, deployed resources, private connectivity, incident state, or production readiness.
- Sampled read-only evidence proves only the sampled configured environment and time window.
- User-provided evidence can be incomplete or stale; preserve uncertainty.
- Never ask for credentials, tokens, secrets, tenant IDs, subscription IDs, resource IDs, customer data, private keys, or raw incident payloads.

## Required phrasing

Use generic phrasing such as "Microsoft Learn documentation through the user's configured documentation MCP". Do not expose internal tool names, profile names, environment names, or local identifiers in committed docs.

# Official sources

Use this reference when grounding current Azure behavior for `azure-rbac-review`.

## Microsoft Learn sources

- https://learn.microsoft.com/azure/role-based-access-control/overview
- https://learn.microsoft.com/azure/role-based-access-control/best-practices
- https://learn.microsoft.com/azure/role-based-access-control/scope-overview
- https://learn.microsoft.com/azure/role-based-access-control/built-in-roles
- https://learn.microsoft.com/azure/role-based-access-control/custom-roles
- https://learn.microsoft.com/azure/role-based-access-control/conditions-overview
- https://learn.microsoft.com/entra/id-governance/privileged-identity-management/pim-configure

## Current documentation refresh (2026-06-04)

- Microsoft Learn documentation through the user's configured documentation MCP is the primary source for documented Azure behavior.
- Documentation evidence is not live customer-state evidence. It does not prove the user's tenant, subscription, RBAC, quotas, deployed resources, incident posture, private connectivity, automation state, or production readiness.
- Use sampled read-only Azure evidence only when the user has configured it and the task requires current-state confirmation. Label it as sampled evidence, not broad proof.

## Grounding rule

Docs explain service behavior. Current-state claims require sampled read-only evidence or sanitized user-provided evidence. If current state was not queried or shown, say so.

# Azure RBAC Review Operations

> Version note: Azure service behavior and tooling change over time. Verify exact command syntax, permissions, and feature availability against Microsoft Learn documentation through the user's configured documentation MCP before production use. Do not paste secrets or sensitive identifiers into commands, files, or chat.

Use this reference for current, source-grounded service behavior and the hard review gates that the lean `SKILL.md` intentionally does not carry.

## What people get wrong

- Giving Owner because the exact job-function role was not checked.
- Assigning broad roles at subscription or management-group scope for convenience.
- Creating wildcard custom roles that silently inherit future permissions.
- Assigning roles directly to users instead of groups or managed identities.
- Treating RBAC as static and ignoring PIM, reviews, and privileged assignment governance.

## Officially grounded service shape

- Microsoft Learn evidence says Azure RBAC defines who can access Azure resources, what actions they can perform, and where they can perform them.
- Best practices emphasize least privilege, narrow scope, limiting subscription owners, limiting privileged administrator roles, PIM for time-bound access, assigning roles to groups, using unique role IDs in automation, and avoiding wildcard custom-role permissions.
- Privileged administrator roles require special scrutiny; where role assignment delegation is needed, conditions can constrain what assignees may grant.
- Built-in job-function roles should be preferred before custom roles, and custom roles should specify explicit Actions and DataActions.

Documentation evidence proves documented Azure service behavior. It does not prove the user's tenant, subscription, RBAC, quotas, deployed resources, incident state, or production readiness.

## Non-negotiable design rules

- Start from task-required actions and scope, not from a desired role name.
- Prefer built-in job-function roles and narrow scopes before custom roles or privileged administrator roles.
- Use groups or managed identities instead of direct user assignments where operationally possible.
- Use PIM or time-bound access for elevated privileges.
- Reject wildcard custom roles unless there is a documented, exceptional justification and compensating review.

## Minimal safe implementation flow

- Scope principal, resource, required actions, data-plane needs, assignment duration, and approval path.
- Inventory existing assignments and inherited assignments at management group, subscription, resource group, and resource scopes.
- Compare required actions to built-in roles, then custom role only if necessary.
- Assess privileged role, PIM, condition, group assignment, and review requirements.
- Return least-privilege recommendation, risks, rollback/removal plan, and verification query targets.

## High-risk assumptions to kill

- Owner or Contributor is not a default troubleshooting role; job-function built-in roles and narrower scopes must be checked first.
- A role assignment at management-group or subscription scope multiplies blast radius even if the principal only needs one resource.
- Direct user assignments and standing privileged access are operational debt unless there is a documented exception.
- Custom roles with wildcards can silently gain future permissions and are not least privilege.
- RBAC evidence is incomplete unless inherited assignments, data-plane permissions, PIM eligibility, conditions, and review cadence are considered.

## Safe command/code verification targets

- Inventory direct and inherited assignments for the principal at management group, subscription, resource group, and resource scopes.
- Compare required management-plane and data-plane actions against built-in job-function roles before custom role design.
- Check privileged administrator roles, assignment conditions, PIM/time-bound controls, group-based assignment, and access review evidence.
- Verify custom role definitions use explicit Actions/DataActions and stable role IDs for automation.
- Provide removal/expiry verification so access can be cleanly revoked after the task.

## Safe verification targets

- Chosen role grants only required management-plane and data-plane actions.
- Assignment scope is the narrowest workable scope.
- Privileged administrator roles are absent or explicitly justified with PIM/time-bound controls.
- Custom roles avoid wildcards and use explicit Actions/DataActions.
- Role assignment can be removed or expires cleanly after the task.

## When to push back

- The user asks for Owner without proving why narrower roles fail.
- The request grants permissions to an individual user for a standing operational pattern.
- The custom role uses wildcard permissions.
- The target scope is broader than the resource set named in the task.

# Safety checklist

Use before recommending production Azure changes, access grants, network connectivity changes, deployment automation, resilience claims, or incident conclusions for `azure-rbac-review`.

## Non-negotiables

- Do not ask for or print credentials, client secrets, certificates, private keys, access tokens, tenant IDs, subscription IDs, resource IDs, customer data, raw incident payloads, or environment-specific identifiers.
- Prefer Microsoft Learn documentation through the user's configured documentation MCP for documented Azure behavior.
- Use sampled read-only Azure evidence only for current-state claims and label it as sampled evidence.
- Require explicit approval before recommending live mutation, broad access, destructive remediation, production deployment, DNS changes, failover, failback, or alert suppression.
- Keep recommendations least-privilege, reversible where possible, and scoped to the named resource or workload.
- Separate documentation-based claims, sampled evidence, user-provided evidence, and inference.

## Component risks

- **Identity and RBAC:** broad privileged roles, direct user grants, wildcard custom roles, missing PIM/time-bound controls, inherited scope surprises.
- **Automation and IaC:** missing preview, unreviewed delete/modify changes, overbroad deployment identities, unsafe secret handling, no rollback path.
- **Networking and Private Link:** DNS misconfiguration, duplicate private DNS zones, missing VNet links, resolver/forwarder gaps, route surprises, broken application connectivity.
- **Resilience and BCDR:** fantasy RTO/RPO, untested restore, undocumented failback, inaccessible DR assets, hidden single-region dependencies.
- **Health triage:** false provider attribution, unsupported resource health, ignored activity-log changes, sensitive incident payload exposure, broad remediation before blast-radius evidence.

## Evidence labels

Use `documentation-based`, `sampled read-only evidence`, `repo evidence`, `user-provided evidence`, or `inference`. Documentation alone never proves the user's live Azure environment.

# Workflow and output contract

Use this reference for full execution of `azure-rbac-review`.

## Workflow

1. **Classify the request**
   - Identify service/domain, resource scope, environment, production impact, and whether mutation is requested.
   - Identify whether the task needs documentation-only guidance, sampled read-only current-state evidence, or sanitized user evidence.

2. **Ground in current sources**
   - Prefer Microsoft Learn documentation through the user's configured documentation MCP.
   - Read the component operations guide before issuing design, safety, or readiness conclusions.
   - Treat current-state claims as unproven unless supported by sampled read-only evidence or sanitized user-provided evidence.

3. **Stress-test the plan**
   - Kill broad permissions, vague ownership, missing rollback, missing validation, and unsupported production-readiness claims.
   - Separate facts from inference.
   - State blockers before recommendations.

4. **Recommend minimal safe action**
   - Prefer read-only inspection, preview, what-if, dry run, diagnostic query, or staged rollout before mutation.
   - Require explicit approval for live or destructive actions.
   - Keep the recommendation scoped and reversible where possible.

5. **Validate and hand off**
   - Name verification targets and evidence gaps.
   - Provide safe next actions and escalation criteria.
   - Do not claim tenant, subscription, resource, quota, or incident state that was not observed.

## Output contract

Return:

1. Scope and target
2. Evidence level: documentation-based, sampled read-only evidence, user-provided evidence, repo evidence, or inference
3. Key findings and risks
4. Blockers or missing evidence
5. Minimal safe next actions
6. Verification targets
7. Rollback, cleanup, or reversal path where applicable