ai-tool-compliance（AI 工具合规性）

ai-tool-compliance - Internal AI Tool Compliance Automation

When to use this skill

• Starting a new AI project: When scaffolding the compliance foundation (RBAC, Gateway, logs, cost tracking) from scratch
• Pre-deploy P0 full verification: When automatically evaluating all 13 P0 mandatory requirements as pass/fail and computing a compliance score
• RBAC design and permission matrix generation: When defining the 5 roles (Super Admin/Admin/Manager/Viewer/Guest) + granular access control per game/menu/feature unit
• Auditing existing code for compliance: When inspecting an existing codebase against the guide and identifying violations
• Implementing cost transparency: When building a tracking system for model/token/BQ scan volume/cost per action
• Designing a behavior log schema: When designing a comprehensive behavior log recording system (Firestore/BigQuery)
• Role-based verification workflow: When configuring the release approval process based on Section 14 (ServiceStability/Engineer/PM/CEO)
• Building a criteria verification system: When setting up the Rule Registry + Evidence Collector + Verifier Engine + Risk Scorer + Gatekeeper architecture

Installation

npx skills add https://github.com/supercent-io/skills-template --skill ai-tool-compliance

Quick Reference

Slash Mode Router

Mode slash commands are mapped as follows.

• /quick, /compliance-quick -> Quick Scan (/compliance-scan)
• /full, /compliance-full -> Full Verify (/compliance-verify)
• /improve -> Improve (/compliance-improve)

3 Execution Modes

1. Quick Scan (quick-scan)

Statically analyzes the codebase to quickly identify potential P0 violations.

How to run: /compliance-scan, /compliance-quick, /quick or trigger keywords compliance scan, quick scan

What it does:

• Grep/Glob-based code pattern search
• Detect direct external API calls (whether Gateway is bypassed)
• Detect direct Firestore client access
• Detect hardcoded sensitive data
• Check for missing Guest role

Output: List of suspected violations (file path + line number + rule ID)

Duration: 1–3 minutes

2. Full Verify (full-verify)

Fully verifies all 11 P0 rules and computes a quantitative compliance score.

How to run: /compliance-verify, /compliance-full, /full or trigger keywords P0 verification, full verify, deploy verification

What it does:

• Collect Evidence and evaluate pass/fail for each of the 11 P0 rules
• Compute scores per 4 domains (Security 40pts / Auth 25pts / Cost 20pts / Logging 15pts)
• Calculate total compliance score (out of 100)
• Determine deploy gate grade (Green/Yellow/Red)
• Generate role-based approval checklist

Output: Compliance report (compliance-report.md)

## Compliance Report
- Date: 2026-03-03
- Project: my-ai-tool
- Score: 92/100 (Green)

### Rule Results
| Rule ID | Rule Name | Result | Evidence |
|---------|-----------|--------|----------|
| AUTH-P0-001 | Force Guest for New Signups | PASS | signup.ts:45 role='guest' |
| AUTH-P0-002 | Block Guest Menu/API Access | PASS | middleware.ts:12 guestBlock |
| ... | ... | ... | ... |

### Score Breakdown
- Security: 33/40
- Auth: 25/25
- Cost: 17/20
- Logging: 12/15
- Total: 92/100

### Gate Decision: GREEN - Deploy Approved

Duration: 5–15 minutes (varies by project size)

3. Improve (improve)

Provides specific fix guides for violations and runs a re-verification loop.

How to run: /compliance-improve, /improve or trigger keywords compliance improvement, fix violations

What it does:

• Code-level fix suggestions for each FAIL item (file path + before/after code)
• Re-verify the rule after applying the fix
• Track score changes (Before -> After)
• Guide for gradually introducing P1 recommended requirements after passing P0

Output: Fix proposal + re-verification results

Improve Mode Auto-Fix Logic

/compliance-improve runs
       |
  1. Load latest verification-run.json
       |
  2. Extract FAIL items (rule_id + evidence)
       |
  3. For each FAIL:
       |
     a. Read violation code from evidence file:line
     b. Derive fix direction from rule.remediation + rule.check_pattern.must_contain
     c. Generate before/after code diff
     d. Apply fix via Write (after user confirmation)
     e. Re-verify only that rule (re-run Grep pattern)
     f. Confirm transition to PASS
       |
  4. Full re-verification (/compliance-verify)
       |
  5. Output Before/After score comparison
       |
  6. If no remaining FAILs → present guide for introducing P1 recommended requirements

Fix application priority:

1. must_not_contain violations (requires immediate removal) → delete the code or replace with server API call
2. must_contain unmet (pattern needs to be added) → insert code per the remediation guide
3. Warning (partially met) → apply supplement only to unmet files

P0 Rule Catalog

11 P0 rules based on the internal AI tool mandatory implementation guide v1.1:

Scoring System

Per-rule Automatic Verification Logic

Verification for each rule is performed based on the check_pattern defined in rules/p0-catalog.yaml. The core mechanism is Grep/Glob static analysis.

Verdict Algorithm (per rule):

1. Glob(check_targets) → collect target files
2. grep_patterns matching → identify files using that feature
   - 0 matches → N/A (feature not used, no penalty)
3. must_not_contain check (excluding exclude_paths)
   - Match found → immediate FAIL + record evidence
4. must_contain check
   - All satisfied → PASS
   - Partially satisfied → WARNING
   - Not satisfied → FAIL

Key Grep Patterns per Rule:

Detailed schema: see rules/p0-catalog.yaml and the "Judgment Algorithm" section in REFERENCE.md

Verification Scenarios (QA)

5 key verification scenarios run in Full Verify mode (/compliance-verify). Each scenario groups related P0 rules for end-to-end verification.

Verification Flow by Scenario

SC-001: grep signup/register -> assert role='guest' -> grep guestBlock/guestDeny -> assert exists
SC-002: grep fetch(https://) in client -> assert 0 hits -> grep gateway log -> assert cost fields -> assert cache check
SC-003: grep firebase/firestore in client/ -> assert 0 hits -> grep authMiddleware in functions/ -> assert exists
SC-004: grep catch blocks -> assert logAction in each -> grep roleChange -> assert auditLog
SC-005: grep res.json for raw text -> assert 0 hits -> grep guest response -> assert no model/cost info

Role-based Go/No-Go Checkpoints

After the deploy gate verdict, the role's Go/No-Go checkpoints must be cleared based on the grade. 4 roles × 5 items = 20 checkpoints total.

Service Stability (5 items)

Engineer (5 items)

PM (5 items)

CEO (5 items)

Report Format

compliance-report.md, generated when /compliance-verify runs, consists of 6 sections.

Report Section Structure (6 sections)

# Compliance Report

## 1. Summary
- Project name, verification date/time, verification mode (quick-scan / full-verify)
- Total compliance score / 100
- Deploy gate grade (Green / Yellow / Red)
- P0 FAIL count
- Verification duration

## 2. Rule Results
| Rule ID | Category | Rule Name | Result | Score | Evidence |
|---------|----------|-----------|--------|-------|----------|
| AUTH-P0-001 | Auth | Force Guest for New Signups | PASS | 10/10 | signup.ts:45 |
| SEC-P0-005 | Security | Enforce External API Gateway | FAIL | 0/15 | client/api.ts:23 direct fetch |
| ...

## 3. Score Breakdown
| Domain | Score | Max | % |
|--------|-------|-----|---|
| Security | 20 | 40 | 50% |
| Auth | 25 | 25 | 100% |
| Cost | 17 | 20 | 85% |
| Logging | 12 | 15 | 80% |
| **Total** | **79** | **100** | **79%** |

## 4. Failures Detail
For each FAIL item:
- Violation code location (file:line)
- Description of the violation
- Recommended fix (remediation)
- Related verification scenario ID (SC-001–SC-005)

## 5. Gate Decision
- Verdict grade + basis for verdict
- List of required approval roles
- Role-based Go/No-Go checkpoint status (unmet items out of 20 shown)

## 6. Recommendations
- Immediate action: Fix P0 FAILs (file path + fix guide)
- Short-term improvement: Path from Yellow to Green
- Mid-term adoption: Order for introducing P1 recommended requirements

Report Generation Rules

1. Summary is always first: Decision-makers must be able to immediately see the score and grade
2. Evidence required: Attach code evidence (file:line) to all PASS/FAIL items in Rule Results
3. Failures Detail contains only FAILs: PASS items appear only in the Rule Results table
4. Role mapping in Gate Decision: Auto-display required approval roles based on grade
5. Recommendations priority: Sorted as Immediate > Short-term > Mid-term

Deploy Gate Policy

Grade Verdict Criteria

Core Rules

1. P0 Absolute Rule: If even one P0 FAILs, the verdict is Red regardless of total score. Deploy automatically blocked
2. Yellow conditional: Total score passes but not perfect. PM reviews the risk and decides approve/reject
3. Green auto-approve: All P0 passed + score ≥ 90 allows deploy without additional approval

Gate Execution Flow

/compliance-verify runs
       |
  Full verification of 11 P0 rules
       |
  Score computed (Security+Auth+Cost+Logging)
       |
  +----+----+----+
  |         |         |
Green     Yellow    Red
  |         |         |
Auto-Approve  PM-Approve  Deploy-Block
  |       Pending   |
  v         |      After Fix
Deploy       v      Re-verify
        PM Review     |
        |    |      v
      Approve  Reject  /compliance-improve
        |    |
        v    v
      Deploy  After Fix
            Re-verify

Role-based Approval Process

Based on Section 14 of the internal AI tool mandatory implementation guide. The required approval roles vary by deploy grade.

Service Stability

Responsibility: Verify incident impact, performance degradation, and rollback feasibility

Checklist:

• Does the new deployment not impact existing service SLAs?
• Is the rollback procedure documented?
• Has performance testing (load/stress) been completed?
• Are incident alert channels configured?

Approval trigger: Required for Yellow/Red grade

Engineer

Responsibility: Root cause analysis of failed rules + code-level fix + re-verification

Checklist:

• Has the cause of all FAIL rules been identified?
• Does the fixed code accurately reflect the intent of the rule?
• Has the rule transitioned to PASS in re-verification?
• Does the fix have no negative impact on other rules?

Approval trigger: Required for Red grade (responsible for re-verification after fix)

PM (Product Manager)

Responsibility: User impact, schedule risk, scope change approval

Checklist:

• Is the impact of non-compliant items on user experience acceptable?
• Is the schedule impact of fixes acceptable within the overall release timeline?
• If scope reduction/deferral is needed, has stakeholder agreement been reached?

Approval trigger: Required for Yellow grade

CEO

Responsibility: Cost cap, business risk, final approval

Checklist:

• Is AI usage cost within the pre-approved budget?
• Is the security risk at an acceptable level for the business?
• Are legal/regulatory risks identified and managed?

Approval trigger: When cost cap is exceeded or when approving a security P0 exception

Project Initialization (/compliance-init)

Generated File Structure

project/
├── compliance/
│   ├── rbac-matrix.yaml          # 5-role × game/menu/feature permission matrix
│   ├── rules/
│   │   └── p0-rules.yaml         # 11 P0 rule definitions
│   ├── log-schema.yaml           # Behavior log schema (Firestore/BigQuery)
│   └── cost-tracking.yaml        # Cost tracking field definitions
├── compliance-config.yaml        # Project metadata + verification settings
└── compliance-report.md          # Verification result report (generated on verify run)

Each YAML File Schema

compliance-config.yaml (project root):

project:
  name: "my-ai-tool"
  type: "web-app"           # web-app | api | mobile-app | library
  tech_stack: ["typescript", "firebase", "next.js"]

verification:
  catalog_path: "compliance/rules/p0-rules.yaml"   # default
  exclude_paths:                                     # paths to exclude from verification
    - "node_modules/**"
    - "dist/**"
    - "**/*.test.ts"
    - "**/*.spec.ts"

scoring:
  domain_weights:            # total = 100
    security: 40
    auth: 25
    cost: 20
    logging: 15

gate:
  green_threshold: 90        # >= 90 = auto approve
  yellow_threshold: 75       # 75-89 = PM review required
  p0_fail_override: true     # Red verdict on P0 FAIL regardless of score

compliance/log-schema.yaml (behavior log schema):

log_schema:
  version: "1.0.0"
  storage:
    primary: "firestore"           # for real-time access
    archive: "bigquery"            # for analytics/audit
    retention:
      hot: 90                      # days (Firestore)
      cold: 365                    # days (BigQuery)

  fields:
    - name: userId
      type: string
      required: true
    - name: action
      type: string
      required: true
      description: "action performed (ai_call, role_change, login, etc.)"
    - name: timestamp
      type: timestamp
      required: true
    - name: model
      type: string
      required: false
      description: "AI model name (gemini-1.5-flash, etc.)"
    - name: inputTokens
      type: number
      required: false
    - name: outputTokens
      type: number
      required: false
    - name: estimatedCost
      type: number
      required: false
      description: "estimated cost in USD"
    - name: status
      type: string
      required: true
      enum: [success, fail, timeout, error]
    - name: errorMessage
      type: string
      required: false
    - name: metadata
      type: map
      required: false
      description: "additional context (bytesProcessed, cacheHit, etc.)"

compliance/cost-tracking.yaml (cost tracking fields):

cost_tracking:
  version: "1.0.0"

  ai_models:
    required_fields:
      - model              # model identifier
      - inputTokens        # input token count
      - outputTokens       # output token count
      - estimatedCost      # estimated cost in USD
    optional_fields:
      - cacheHit           # whether cache was hit
      - latencyMs          # response latency (ms)

  bigquery:
    required_fields:
      - queryId            # query identifier
      - bytesProcessed     # bytes scanned
      - estimatedCost      # estimated cost in USD
    optional_fields:
      - slotMs             # slot usage time
      - cacheHit           # BQ cache hit indicator

  cost_formula:
    gemini_flash: "$0.075 / 1M input tokens, $0.30 / 1M output tokens"
    gemini_pro: "$1.25 / 1M input tokens, $5.00 / 1M output tokens"
    bigquery: "$5.00 / TB scanned"

RBAC Matrix Base Structure

# compliance/rbac-matrix.yaml
roles:
  - id: super_admin
    name: Super Admin
    description: Full system administration + role assignment rights
  - id: admin
    name: Admin
    description: Service configuration + user management
  - id: manager
    name: Manager
    description: Team/game-level management
  - id: viewer
    name: Viewer
    description: Read-only access
  - id: guest
    name: Guest
    description: Minimum access (tool name/model name/cost/structure not exposed)

permissions:
  - resource: "dashboard"
    actions:
      super_admin: [read, write, delete, admin]
      admin: [read, write, delete]
      manager: [read, write]
      viewer: [read]
      guest: []  # no access
  # ... expand per game/menu/feature

Gateway Pattern Example

// functions/src/gateway/ai-gateway.ts
// Direct external API calls forbidden - must route through this Gateway

import { onCall, HttpsError } from "firebase-functions/v2/https";
import { verifyRole } from "../auth/rbac";
import { logAction } from "../logging/audit";
import { checkCache } from "../cache/cost-cache";

export const callAIModel = onCall(async (request) => {
  // 1. Server-side auth verification (AUTH-P0-003)
  const user = await verifyRole(request.auth, ["admin", "manager"]);

  // 2. Block Guest access (AUTH-P0-002)
  if (user.role === "guest") {
    throw new HttpsError("permission-denied", "Access denied");
  }

  // 3. Cache-first lookup (COST-P0-011)
  const cached = await checkCache(request.data.prompt);
  if (cached) {
    await logAction({
      userId: user.uid,
      action: "ai_call",
      source: "cache",
      cost: 0,
    });
    return { result: cached, fromCache: true };
  }

  // 4. AI call via Gateway (SEC-P0-005)
  const result = await callGeminiViaGateway(request.data.prompt);

  // 5. Record cost log (COST-P0-006)
  await logAction({
    userId: user.uid,
    action: "ai_call",
    model: result.model,
    inputTokens: result.usage.inputTokens,
    outputTokens: result.usage.outputTokens,
    estimatedCost: result.usage.estimatedCost,
  });

  // 6. Sensitive raw content processed server-side only; return reference ID to client (SEC-P0-009)
  const responseRef = await storeResponse(result.text);
  return { responseId: responseRef.id, summary: result.summary };
});

Relationship with Other Skills

Integration with bmad-orchestrator

Where bmad-orchestrator manages the full project development phases (Analysis -> Planning -> Solutioning -> Implementation), ai-tool-compliance acts as a companion tool that verifies whether each phase's outputs meet compliance requirements.

Recommended usage order:

1. /workflow-init (bmad) -- establish project structure
2. /compliance-init (compliance) -- generate compliance foundation
3. After Phase 3 Architecture completes, /compliance-scan -- architecture-level compliance check
4. After Phase 4 Implementation completes, /compliance-verify -- full verification + deploy gate verdict

Relationship with security-best-practices

security-best-practices provides general web security patterns (OWASP, HTTPS, XSS, CSRF). ai-tool-compliance assumes these as prerequisites and focuses on organization-specific requirements (5-role RBAC, Gateway enforcement, cost transparency, comprehensive behavior logging).

Relationship with code-review

Where code-review subjectively reviews code quality/readability/security, ai-tool-compliance provides a quantitative verdict (pass/fail, out of 100) on "Does it pass the internal AI tool guide?" The /compliance-scan result can be used as reference material during code review.

Relationship with workflow-automation

Where workflow-automation provides general CI/CD patterns (npm scripts, Makefile, GitHub Actions), ai-tool-compliance provides the domain-specific verification step inserted into that pipeline.

scripts/ Directory — Detailed Implementation

install.sh -- Skill installation and initialization

bash scripts/install.sh [options]
  --dry-run        preview without making changes
  --skip-checks    skip dependency checks

What it does:

1. Dependency check (yq, jq -- optional; falls back to default parsing if absent)
2. Apply chmod +x to all scripts/*.sh
3. Validate YAML syntax of rules/p0-catalog.yaml
4. Print installation summary

verify.sh -- Full P0 rule verification

bash scripts/verify.sh [--rule RULE_ID] [--output JSON_PATH]

What it does:

1. Parse rules/p0-catalog.yaml (yq or grep-based)
2. For each rule:

- Collect target files via check_targets Glob - Detect feature usage via grep_patterns (N/A if not used) - Detect must_not_contain violations (excluding exclude_paths) - Verify must_contain compliance - Determine Pass/Fail/Warning/N/A + collect evidence

1. Output results in templates/verification-run.json format
2. Print summary table to console

score.sh -- Compliance score computation

bash scripts/score.sh [--input VERIFY_JSON] [--verbose]

What it does:

1. Load verify.sh result JSON (or run verify.sh directly)
2. Calculate scores by domain:

- Pass = 100% of score, Warning = 50%, Fail = 0%, N/A = excluded from denominator - Security: sum of SEC rule scores / Security max (35) - Auth: sum of AUTH rule scores / Auth max (30) - Cost: sum of COST rule scores / Cost max (20) - Logging: sum of LOG rule scores / Logging max (15)

1. Compute total score (out of 100)
2. Render templates/risk-score-report.md
3. Generate templates/remediation-task.json (per FAIL item)

gate.sh -- Deploy gate check

bash scripts/gate.sh
# exit 0 = Green (deploy approved)
# exit 1 = Red (deploy blocked)
# exit 2 = Yellow (conditional -- PM review required)

What it does:

1. Run verify.sh + score.sh sequentially
2. Check for P0 FAIL existence

- If any exist → Red (exit 1)

1. Determine grade based on total score

- 90+ → Green (exit 0) - 75–89 → Yellow (exit 2) - ≤ 74 → Red (exit 1)

1. Print grade + score + list of items requiring fixes to console

CI/CD integration example (GitHub Actions):

# .github/workflows/compliance-gate.yml
name: Compliance Gate
on: [pull_request]
jobs:
  compliance:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Run compliance gate
        run: bash .agent-skills/ai-tool-compliance/scripts/gate.sh

P1 Recommended Requirements (Gradual Adoption)

Items recommended for gradual adoption after passing all P0 rules:

P1 v1.1 Rule Catalog (Extended)

Standard Columns for Notion Table Sorting (Additional)

Criteria Verification System Design (Additional)

Operating Modes (Additional, preserving existing flow)

Constraints

Mandatory Rules (MUST)

1. P0 Absolute Principle: All 11 P0 rules are verified without exception. Partial verification is not allowed
2. Server final verification: All auth decisions are made server-side. Client-side checks alone cannot result in PASS
3. Gateway enforcement: Any direct external AI API call discovered results in unconditional FAIL. No bypass allowed
4. Guest default: If a role other than Guest is assigned on signup, FAIL
5. Evidence-based verdict: All pass/fail verdicts must include code evidence (file path + line number)

Prohibited Actions (MUST NOT)

1. P0 exception abuse: P0 exceptions are never allowed without CEO approval
2. Score manipulation: PASS verdict without Evidence is forbidden
3. Gateway bypass: Direct external API calls are not allowed for any reason including "testing purposes"
4. Selective logging: Logging only successful requests while omitting failed requests is not allowed

Best practices

1. Shift Left: At project start, generate the foundation with /compliance-init before implementing business logic
2. Gradual adoption: Adopt in order: pass all P0 first -> then P1. Do not start with P1
3. Re-verification loop: After fixing violations, always re-run /compliance-verify to confirm score changes
4. BMAD integration: Adopt running compliance-verify after bmad-orchestrator Phase 4 as the standard workflow
5. CI/CD integration: Automate by adding a compliance-gate step to GitHub Actions

References

• Internal AI Tool Mandatory Implementation Guide v1.1 (Notion)
• OWASP Top 10
• Firebase Security Rules
• Cloud Functions for Firebase

Metadata

Version

• Current version: 1.0.0
• Last updated: 2026-03-03
• Compatible platforms: Claude, Gemini, Codex, OpenCode

Related Skills

• bmad-orchestrator: Development phase orchestration
• security-best-practices: General web security patterns
• code-review: Code quality/security review
• workflow-automation: CI/CD automation patterns
• authentication-setup: Authentication/authorization system setup
• firebase-ai-logic: Firebase AI integration

Tags

#compliance #RBAC #security #cost-tracking #audit-log #gateway #firestore #deploy-gate #P0 #AI-tool