Compliance Guard

Purpose

Use this skill to validate artifacts (specs, API contracts, data models, architectures, ADRs, workflows, generated code) against a defined set of organizational or architectural rules. The skill produces structured compliance verdicts that distinguish blocking violations from warnings, support waiver workflows, and provide rule-level traceability.

This skill is domain-generic. It must work for any software system, API design, data model, or architecture without embedding project-specific assumptions.

When to Use

Use this skill when the user asks to:

• Validate an artifact against a rule catalog or compliance checklist.
• Check if a spec, API, or architecture violates organizational constraints.
• Enforce naming conventions, structural patterns, or architectural principles.
• Run a compliance gate before approving an artifact for implementation.
• Audit generated code or synthetic artifacts for rule adherence.
• Identify which rules pass, which warn, and which fail.
• Recommend waivers or exceptions when a rule cannot be met.

Do not use this skill to invent governance rules, define organizational policy from scratch, or perform general quality review (use reviewer for that). This skill evaluates against provided or inferred rules — it does not define them unilaterally.

Core Operating Rules

1. Never invent governance. If the user does not provide rules, ask for them before proceeding. Only infer universally common standards when the user explicitly requests inference.
2. Classify every rule correctly. Distinguish constraint (hard requirement), convention (agreed pattern), and policy (organizational rule with possible exceptions). Misclassification leads to wrong verdicts.
3. Three-level verdict only. Use pass, pass with warnings, or fail for each rule. Add not verifiable when evidence is insufficient to evaluate.
4. Distinguish blocking from warnable. Constraint violations are blocking; convention violations may be warnings; policy violations depend on waiver status.
5. Recommend waivers when relevant. If a rule cannot be met, suggest a waiver with rationale and traceability reference rather than failing silently.
6. Be evidence-based. Reference the specific artifact section, line, or pattern that triggered each finding.
7. Use unique rule IDs. Every rule has an ID (e.g., API-001, ARCH-002, DATA-003) for traceability.
8. Preserve existing compliance state. Acknowledge what passes correctly — a compliance review is not only about failures.

Rule Classification

Verdict Definitions

Verdict Aggregation

Rule Catalog Format

A rule catalog is a YAML list of rules. Each rule has:

- id: "API-001"           # Unique ID (prefix categorizes: API, ARCH, DATA, OPS, NOMEN, etc.)
  class: constraint        # constraint | convention | policy
  description: "RFC 7807 problem responses"
  check: |
    Every error response MUST use RFC 7807 application/problem+json format.
    Must include "type", "title", "status", "detail", and may include "instance".
  severity: blocking       # blocking | warning — only for convention/policy
  example_good: |
    {"type": "https://example.com/probs/tenant-not-found", "title": "Tenant not found", "status": 404, "detail": "Tenant 'acme' does not exist"}
  example_bad: |
    {"error": "Tenant not found"}
  reference: "RFC 7807"   # Optional: external standard, ADR, or policy doc

When the user provides a catalog, apply it as-is. Do not modify rule definitions.

Inference Rules (Only When Explicitly Requested)

If the user says "infer common standards" or "apply default rules", apply the following universally common rules only:

Never infer constraint rules beyond the list above. Never infer policy rules.

Evaluation Procedure

Phase 1: Intake

1. Identify the artifact type (spec, API contract, data model, architecture, ADR, workflow, generated code).
2. Confirm whether the user provided a rule catalog.

• If yes: load and apply the catalog.
• If no: ask for the rule catalog. Do not proceed without rules unless the user explicitly requests inference.

1. Note which rules are verifiable with the provided evidence.

Phase 2: Rule-by-Rule Evaluation

For each rule in the catalog:

1. Locate relevant content in the artifact.
2. Evaluate against the rule's check condition.
3. Assign verdict: pass, pass with warnings, fail, or not verifiable.
4. Record evidence (quote the specific section or pattern that triggered the finding).
5. If fail and a waiver is plausible, note the waiver recommendation.

Phase 3: Synthesis

1. Aggregate per-rule verdicts to overall verdict.
2. List blocking failures first.
3. List warnings with context.
4. List not verifiable rules with explanation.
5. Produce the structured compliance report.

Required Output Structure

# Compliance Report: <Artifact Name>

## 1. Verdict

**Overall:** <pass | pass with warnings | fail>

<One-paragraph summary: how many rules passed, failed, warned, not verifiable.>

## 2. Compliance Summary

| Rule ID | Class | Rule Description | Verdict | Evidence |
| --- | --- | --- | --- | --- |
| API-001 | constraint | RFC 7807 problem responses | pass | §4.2 uses correct format |
| ARCH-001 | constraint | No circular dependencies | fail | Module A → Module B → Module C → Module A |
| OPS-001 | convention | Healthcheck endpoint | pass | GET /health returns 200 |
| DATA-001 | constraint | Multi-tenant tenant_id | not verifiable | No entity definitions present |

3. Blocking Violations

## 4. Warnings and Recommendations

| ID | Rule | Concern | Suggested Action |
| --- | --- | --- | --- |
| OPS-003 | Retry backoff | Exponential backoff not defined | Add jittered exponential backoff with circuit breaker |
| NOMEN-002 | Field naming | Mixed snake_case and camelCase | Standardize to snake_case |

5. Not Verifiable Rules

## 6. Waiver Recommendations

| ID | Rule | Justification | Traceability Reference |
| --- | --- | --- | --- |
| API-010 | Sync-only API | Vendor requires synchronous response; async not supported by protocol | ADR-042, Waiver-2024-017 |

7. Rule Catalog Applied

8. Quality Checklist

## Rule ID Prefixes

Use these prefixes for consistent categorization:

| Prefix | Category |
| --- | --- |
| `API-` | API contracts, HTTP semantics, error formats, versioning |
| `ARCH-` | Architecture, module dependencies, system design |
| `DATA-` | Data models, entity schemas, tenant isolation, persistence |
| `OPS-` | Observability, healthchecks, retries, timeouts, circuit breakers |
| `NOMEN-` | Naming conventions, terminology, URL structure, field formats |
| `SEC-` | Security, auth, authorization, data protection |
| `WORKFLOW-` | Process, state machines, orchestration, workflow definitions |
| `ADR-` | Architecture Decision Records (checks ADR conventions) |

## Example Rule Catalog (YAML)

rules:

• id: "API-001"

class: constraint description: "RFC 7807 problem+json for all error responses" check: "Every error response MUST use RFC 7807 application/problem+json format with type, title, status, detail." severity: blocking reference: "RFC 7807"

• id: "API-002"

class: constraint description: "Idempotency keys on mutating operations" check: "All POST, PUT, PATCH requests MUST accept an Idempotency-Key header and be safely retryable." severity: blocking

• id: "API-003"

class: constraint description: "Timeout on external calls" check: "Every call to an external service or database MUST specify a timeout." severity: blocking

• id: "ARCH-001"

class: constraint description: "No circular module dependencies" check: "Module dependency graph must be acyclic. No module may transitively import itself." severity: blocking

• id: "ARCH-002"

class: constraint description: "No synchronous blocking calls in async contexts" check: "In async service contexts, external calls must be non-blocking. Sync-over-async patterns are forbidden." severity: blocking reference: "Ops policy OP-03"

• id: "DATA-001"

class: constraint description: "Multi-tenant entities include tenant_id" check: "Every entity accessible across tenants MUST have a tenant_id field as the first attribute." severity: blocking

• id: "OPS-001"

class: convention description: "Healthcheck endpoint" check: "Service MUST expose a GET /health endpoint returning 200 when healthy." severity: warning

• id: "OPS-002"

class: convention description: "Retry with exponential backoff" check: "Transient failures MUST be retried with exponential backoff and jitter." severity: warning

• id: "OPS-003"

class: convention description: "Observability: structured logging" check: "All service logs MUST be structured (JSON) with trace_id, span_id, tenant_id where applicable." severity: warning

• id: "NOMEN-001"

class: convention description: "URL paths use kebab-case" check: "URL path segments MUST use kebab-case (e.g., /user-profiles, not /userProfiles or /user_profiles)." severity: warning

• id: "NOMEN-002"

class: convention description: "JSON fields use snake_case" check: "JSON request/response fields MUST use snake_case (e.g., created_at, not createdAt)." severity: warning

## Present Results to User

Lead with the overall verdict and the count of blocking failures. Present blocking violations first with clear evidence and required fixes. Then present warnings with suggested actions. If waivers are recommended, explain why and provide the traceability reference format the user should complete.

If any rules are `not verifiable`, explain what evidence is needed to close the compliance gate and ask whether to proceed with what can be verified or wait for complete evidence.

## Troubleshooting

- **No rules provided:** Ask the user for a rule catalog before proceeding. Do not invent governance.
- **Artifact is too large:** Ask which sections or rule categories to prioritize. Apply rules in order of severity (constraint > policy > convention).
- **Rule conflicts with evidence:** Re-examine the artifact section. If the rule genuinely cannot be evaluated, mark `not verifiable` with the specific reason.
- **Waiver candidate identified:** Flag it in the report with recommendation. Do not assume the waiver is granted.
- **User asks to "check everything":** Ask for a rule catalog or confirm which common rule sets to apply. Do not evaluate against rules the user has not seen or approved.
- **Rule inference requested:** Apply only the inferred common standards table above. Do not add constraint-level rules beyond that list. State clearly which rules were inferred.

(End of file — total 471 lines)