auditor

Master prompts that turn any AI coding agent into a swarm of specialist auditors — for security, engineering, frontend, API, performance, data, infrastructure, AI/LLM, compliance, accessibility, documentation, content, and lean.

Website: auditor.rapold.io · Repository: rapold-io/mra-auditor · Design: product signet from Atlas, the Rapold design system.

[!NOTE] Management summary. auditor is a library of 13 reusable "master prompt" audit templates plus the standards that govern their output. You point one at a repository, app, website, API, datastore, or infrastructure; it runs a parallel swarm of specialist agents, adversarially verifies every finding, scores the result, and files GitHub issues (German or English) — led by a single priority-sorted tracking issue. With a readiness target set, the same run also yields a control-by-control gap matrix for SOC 2, ISO 27001, ISO 42001 / EU AI Act, or NIS2 / CRA, so the backlog doubles as certification evidence. It exists because most "audit my code" prompts return vague, unverified opinions; these enforce evidence, self-challenge, and an actionable 30/60/90 roadmap.

[!NOTE] Kurzfassung (DE). Wiederverwendbare Master-Prompts für die tiefe, parallelisierte Durchleuchtung von Repos, Apps, APIs, Daten und Infrastruktur. Viele Agenten arbeiten gleichzeitig, challengen ihre Ergebnisse adversariell, decken blinde Flecken auf und öffnen am Ende GitHub-Issues auf Deutsch — zuerst ein nach Priorität sortiertes Tracking-Issue, dann pro Befund ein Ticket mit eigener Management-Summary.

Figure 1. How auditor works: a standard plus a template guide an AI agent swarm over a target, producing a scored report and GitHub issues (German or English).

flowchart LR
    Standards["Standards<br/>(documentation,<br/>issue-output)"] --> Prompt
    Target["Target repo / app / API"] --> Agent
    Prompt["Audit template<br/>(1 of 13)"] --> Agent["AI agent swarm<br/>(parallel + adversarial)"]
    Agent --> Report["Report +<br/>scorecard"]
    Agent --> Issues["GitHub issues<br/>(tracker + findings, DE/EN)"]

---

Contents

• Overview
• The audit library
• Run it from your AI agent
• Certification readiness
• Outputs
• Continuous compliance
• Worked example: auditing this repo
• Quickstart
• How it works
• Issue output
• Standards
• Read-only and authorization
• Repository layout
• Contributing
• Changelog and license

---

Overview

Every prompt in this repo enforces the same engineering discipline a top-tier review board would demand:

• Evidence or it didn't happen. Every finding cites a concrete artifact — file:line, a query plan, a request/response, a config value, a measured metric. No evidence means it is discarded.
• Adversarial self-challenge. No finding survives until independent skeptic agents have tried to refute it. This is the strongest defense against the false positives that make most AI audits useless.
• Blind-spot hunting. A completeness critic asks each round which surface, use-case, or assumption was not examined. Unchecked areas are declared, never hidden.
• Maximum parallelism. Work fans out across many specialist agents at once (Claude Code ultracode / Workflow mode), then pipelines per-finding verification.
• A standardized procedure. Every audit runs the same phases and maps findings to recognized standards (OWASP, CWE, MITRE ATT&CK, WCAG, CIS, DORA, RFCs, GDPR).
• Actionable output. A severity-scored register, a 30/60/90 roadmap, and GitHub issues (German or English) led by a priority-sorted tracker.

The bar is "Google-grade": precise, reproducible, no panic, no hand-waving, every recommendation immediately actionable.

The audit library

All templates are stack-, framework-, and product-agnostic and share one methodology, so findings compose cleanly when you run several against the same target.

[!NOTE] Terminology. Each audit ships as a master prompt — a self-contained template you paste into your AI agent. This README uses "audit", "template", and "master prompt" for the same artifact; "audit" names what it does, "template/master prompt" names the file you copy.

Template	Audits	Maps to	Output
security	Security across 14 domains: injection, authN/authZ, secrets, supply chain, IaC, CI/CD, API, business logic, privacy, LLM	OWASP Top 10 (2025) / API / ASVS / LLM, CWE-25, MITRE ATT&CK, CIS, GDPR; controls: ISO 27001, SOC 2, NIS2, CRA	Issues + scorecard (DE/EN)
repo	Whole-repo engineering: architecture, stack consistency, docs, tests, deps, CI/CD, observability, git hygiene	Google Eng Practices, SRE, SLSA, OWASP	Board-ready report
frontend	16-agent frontend sweep: usability, psychology, visual design, a11y, performance, SEO, copy, CRO, IA, forms, trust	Nielsen, WCAG 2.2, Core Web Vitals	Prioritized backlog + scorecard
api	API design: resource modeling, HTTP semantics, error model, versioning, idempotency, rate limits, schema, DX	RFC 9110/9457, OpenAPI 3.1, GraphQL, Google AIP	Contract-drift matrix + roadmap
performance	Performance & scalability: hotspots, N+1, caching, concurrency, leaks, load behavior, resilience, FinOps	SRE practice, DORA, load/latency SLOs	Bottleneck map + per-fix metric gains
data	Data & database: modeling, types, constraints, migration safety, transactions, integrity, backup/DR	Normalization, ACID/CAP, RLS, expand/contract

[!NOTE] Acronyms. DX = developer experience · CRO = conversion-rate optimization · IA = information architecture · RLS = row-level security · FinOps = cloud-cost engineering · RoPA = Records of Processing Activities · DAST = dynamic application security testing.

Run it from your AI agent

Point any capable agent at the live entry point — no install:

Audit github.com/your/repo using auditor.rapold.io

The agent fetches the orchestrator (full-audit-master-prompt.md, surfaced at auditor.rapold.io/llms.txt), asks your output language (Deutsch / English) and which audits to run (or the full repo), then files a consolidated, priority-sorted issue backlog.

Certification readiness

Most organisations do not run an audit for its own sake — they run it because a customer asked for a SOC 2 report, a tender demands ISO 27001, procurement wants an AI Act answer, or a regulatory deadline (NIS2, CRA) is approaching. The orchestrator therefore offers a readiness target; set one and the same run produces, on top of the backlog, what an external auditor or a compliance tool needs:

READINESS_TARGET	Assesses against	Extra deliverables
soc2	SOC 2 Trust Services Criteria	Control gap matrix (CC1–CC9 + in-scope categories), evidence register, readiness score
iso27001	ISO/IEC 27001:2022 Annex A	Gap matrix, Statement of Applicability draft, risk-register seed
iso42001-ai-act	ISO/IEC 42001:2023 + EU AI Act	AI-system inventory with risk tier, gap matrix, impact-assessment skeleton
nis2-cra	NIS2 Art. 21/23 + CRA Annex I / Art. 14	Gap matrix, SBOM and disclosure readiness, reporting-capability check

How it works: every finding of every audit carries three business fields — controls (IDs from CONTROL-CROSSWALK.md, e.g. ISO27001:A.8.3, SOC2:CC6.1), deal_blocker (would this fail an audit or block enterprise procurement?), and fine_exposure (GDPR, revDSG, NIS2, CRA, AI Act tiers). The orchestrator inverts the mapping into one status per control (implemented · partial · missing · not-assessable · n/a), classifies nonconformities the way certification auditors do (Major / Minor / OFI), and computes a technical control readiness score with a time-to-audit-ready estimate. Swiss revDSG and GDPR are always mapped by the compliance-privacy audit. Regulated sectors add an overlay via SECTOR: finance (PCI DSS 4.0.1, EU-DORA, FINMA-RS 2023/1), health (HIPAA Security Rule), swiss (ISG reporting duty, BWL ICT minimum standard, eCH-0059).

[!WARNING] Readiness is not certification. ISO certificates come from accredited bodies, SOC 2 reports from CPA firms. Roughly half of every framework is organisational (policies, HR, physical security) and is reported as not assessable from code, never as missing. What you get is the gap assessment and the evidence pack that make the external audit short.

Framework editions are tracked in FRAMEWORK-VERSIONS.md and reviewed quarterly.

Outputs

Every run ends in one canonical file, auditor-out/audit-run.json (schemas/audit-run.schema.json); everything a board, an auditor, a compliance tool or an engineering tracker needs is derived from it by a dependency-free script — never re-typed by the agent:

node scripts/export-findings.mjs auditor-out/audit-run.json --out auditor-out --repo . [--docx]

File	For	Format
EXECUTIVE-REPORT.md → .docx / .pdf	management, board, external auditor	fixed board structure; pandoc or a document skill converts it
findings.sarif	engineering, GitHub Code Scanning, IDEs	SARIF 2.1.0 (P0/P1 error, P2 warning, P3 note)
assessment-results.oscal.json	GRC platforms	OSCAL 1.1.2 assessment results, one finding per finding × control
gap-matrix.csv, findings.csv	Vanta / Drata / Secureframe, Jira import, spreadsheets	CSV
evidence-manifest.json	external auditor	sha256 + size per cited artifact, no content copied
caiq-answers.csv	security questionnaires (CAIQ v4, SIG)	one pre-filled answer per CSA CCM domain; "Not assessed" where the run is silent
acr-wcag22.csv	procurement, accessibility statements	Accessibility Conformance Report, one row per WCAG 2.2 A/AA criterion (VPAT 2.5 vocabulary)
trust-center.html	prospects, security reviews	aggregate-only readiness page, no finding detail
checks.json, audit-diff.json	CI	executable re-audit criteria and the run-to-run diff — see Continuous compliance
Tracker issues	engineering	GitHub, Jira, Linear or ServiceNow via ISSUE_TARGET

The contract is REPORT-OUTPUT-STANDARD.md. Output language is chosen per run — English and German are equally first-class; German follows the de-CH rules of the issue standard.

Continuous compliance

An audit that runs once decays. Three pieces keep it alive between audits:

• Executable re-audit checks. A finding whose fix is machine-verifiable carries a check (a shell command that passes once the finding is fixed). node scripts/verify-checks.mjs auditor-out/audit-run.json --repo . runs them all and exits non-zero while a finding is open.
• Run-to-run diff. node scripts/diff-runs.mjs baseline.json current.json --fail-on new-p0p1,regression answers "better or worse since the last audit?" — added, fixed, worsened, score and control deltas — and gates CI on it.
• GitHub Action. rapold-io/mra-auditor/.github/actions/[email protected] validates the run, exports everything, writes the executive report into the job summary, diffs against a baseline, runs the checks and uploads SARIF to Code Scanning — no third-party action involved.

- uses: rapold-io/mra-auditor/.github/actions/[email protected]
  with:
    run-file: auditor-out/audit-run.json
    baseline: auditor-out/baseline/audit-run.json
    fail-on: new-p0p1,regression

The MCP server exposes the same crosswalk to agents: get_readiness_checklist(target) returns every technically assessable control of a target with the audits to run, and list_control_themes(audit?) is the lookup for a finding's controls field.

Worked example: auditing this repo

The library is dogfooded on itself. Running the orchestrator end-to-end against auditor produced a consolidated, cross-audit backlog — a concrete trace of the input-to-issues loop:

• Input. Point the agent at the repo. Phase 0 recon selected the seven applicable audits and declared four not applicable, with reasons (no runtime API, no database, no in-repo LLM runtime, no personal data) rather than skipping them silently.
• Scorecard. documentation A (94), accessibility A (93), performance A (92), security A (91), repo / frontend / infrastructure A- (90) — zero P0, one P1.
• Cross-audit dedupe. "CHECKSUMS.txt is not verified in CI" was found independently by the repo, infrastructure, and security lenses and merged into a single backlog item with all three citations — the payoff of running the audits together.
• Output. One master tracking issue plus one issue per finding, each with a management summary, evidence (file:line), a before/after fix, an effort estimate, and a re-audit criterion.

See the run on GitHub: master tracker #97 (scorecard, consolidated backlog, and a 30/60/90 roadmap) and the per-finding issues it links.

Quickstart

[!TIP] Claude Code with ultracode or Workflow mode unlocks the real multi-agent parallelism. Any capable agent also works single-threaded — it runs the same phases sequentially.

Prerequisites: a capable AI coding agent; the GitHub CLI (gh) installed and authenticated if you target a GitHub URL or want the audit to file issues; and write access to the target repository for the issue phase.

1. Open your AI coding agent in the target's working directory, or give it a GitHub URL.

2. Paste the chosen master prompt in full (open the file on GitHub and use "Copy raw file"), then fill in the config block at its top — every prompt starts with one:

   TARGET:       <local repo path or GitHub URL>
   SCOPE:        <whole repo | specific flows/paths>
   STACK:        <languages, frameworks, cloud — or let Phase 0 infer>
   AUDIENCE:     <who the target serves, if known>
   DATA_ACCESS:  <can run / active-test? or read-only>
   OUTPUT_LANG:  <English | Deutsch>  →  chosen per run; ask if unset, never assume
   ISSUE_TARGET: <github:owner/repo | jira:KEY | linear:TEAM | servicenow:<url> — preview-first, on approval>
   

3. Let it run. It builds a fact sheet, fans out specialist agents, cross-pollinates and dedupes, adversarially verifies every serious finding, benchmarks against best-in-class, then synthesizes the report and roadmap.

4. Review the issues. Each audit previews a priority-sorted tracking issue plus one issue per finding (in your chosen language), and creates them only on your explicit approval.

To get the full picture of a production system, run repo + security + performance + data + infrastructure (and frontend / api / ai-llm / compliance-privacy / accessibility / documentation where applicable). Each produces an independent report and issue set.

How it works

Every template runs the same phases:

Phase 0  Reconnaissance      factual inventory + attack/surface map (no opinions yet)
Phase 1  Specialist swarm    many domain experts run IN PARALLEL, each evidence-bound
Phase 2  Cross-pollination   merge + dedupe + surface compound findings
Phase 3  Adversarial verify  independent skeptics try to REFUTE each P0/P1; ≥2/3 to survive
Phase 4  Benchmark           compare against named best-in-class references & standards
Phase 5  Synthesis           report, scorecard, issues, 30/60/90 roadmap

Figure 2. The six-phase method: parallel specialists, then adversarial verification before anything reaches the report.

flowchart LR
    P0["Phase 0<br/>Recon &<br/>surface map"] --> P1

    subgraph P1["Phase 1 — parallel specialist swarm"]
        direction TB
        A1["Specialist A"]
        A2["Specialist B"]
        A3["Specialist C"]
        A4["Specialist …N"]
    end

    P1 --> P2["Phase 2<br/>merge ·<br/>dedupe ·<br/>compound"]
    P2 --> P3{"Phase 3<br/>adversarial<br/>verify · ≥2/3?"}
    P3 -- "refuted" --> KILL["Killed<br/>(appendix +<br/>refutation)"]
    P3 -- "survives" --> P4["Phase 4<br/>benchmark vs<br/>best-in-class"]
    P4 --> P5["Phase 5<br/>report · scorecard ·<br/>issues · roadmap"]
    P5 -. "blind-spot critic:<br/>what did we miss?" .-> P1

Severity is standardized (P0 critical to P3 polish), each finding carries an effort estimate and an ICE/priority score, and the roadmap is always Quick Wins → 30 → 60 → 90 days, dependency-aware, with finding IDs traced through. Every report ends with an honest "Coverage and limitations" section; an unaudited area reported as "fine" is treated as an audit failure.

Issue output

After verification, every audit produces GitHub issues per ISSUE-OUTPUT-STANDARD.md:

1. A tracking issue first — the index of all sub-tasks, sorted by priority (P0→P3), with a management summary, a scorecard, and the 30/60/90 roadmap. Each checklist line links its child issue.
2. One issue per confirmed finding — German, top-notch documented, each opening with its own management summary, then severity, evidence, a concrete before/after fix, effort, and a re-audit criterion.

Issues are written in the language chosen per run (English or German), can target GitHub, Jira, Linear or ServiceNow, are preview-first, and are created only on explicit authorization.

Standards

Four normative standards govern the audits and are reusable on their own:

• DOCUMENTATION-STANDARD.md (German) and DOCUMENTATION-STANDARD.en.md (English) — a Google-grade documentation standard with five repo profiles and a 0–100 scoring rubric. The documentation audit uses it as its yardstick.
• ISSUE-OUTPUT-STANDARD.md — the mandatory tracking-issue + per-finding issue contract that every audit follows.
• CONTROL-CROSSWALK.md — the control mapping (SOC 2, ISO 27001, ISO 42001 + AI Act, NIS2, CRA, revDSG/GDPR) behind the controls field of every finding and the orchestrator's readiness mode; editions in FRAMEWORK-VERSIONS.md.
• REPORT-OUTPUT-STANDARD.md — the business-output contract: the canonical audit-run.json, the executive report, SARIF / OSCAL / CSV exports, the evidence manifest, issue targets (GitHub, Jira, Linear, ServiceNow) and the per-run language rule.

A ready-to-use README skeleton lives at templates/README.template.md.

Read-only and authorization

These audits are non-destructive and read-only by default:

• Static analysis of code, config, schema, and IaC needs no special permission.
• Active or dynamic testing (DAST, load tests, injection/jailbreak probes, live cloud or DB access) requires explicit, documented authorization from the owner. Without it, the prompts fall back to static reasoning and flag the limitation.
• No destructive techniques, no DoS, no data exfiltration. Real secrets and PII are never copied into reports — location is cited and the value redacted.
• Creating issues is preview-first and gated on explicit approval.

Use these only on systems you own or are authorized to assess.

Repository layout

auditor/
├── README.md                       you are here
├── ARCHITECTURE.md                 monorepo structure + design decisions
├── DOCUMENTATION-STANDARD.md       the doc standard (German, canonical)
├── DOCUMENTATION-STANDARD.en.md    the doc standard (English mirror)
├── ISSUE-OUTPUT-STANDARD.md        mandatory issue output for all audits
├── CONTROL-CROSSWALK.md            findings → SOC 2 / ISO 27001 / ISO 42001 / NIS2 / CRA / revDSG controls
├── FRAMEWORK-VERSIONS.md           dated register of every framework edition in use
├── REPORT-OUTPUT-STANDARD.md       business outputs: audit-run.json, executive report, SARIF/OSCAL/CSV, issue targets
├── schemas/                        JSON Schemas for the canonical run file and a finding
├── scripts/                        CI gates, checksums, version pins, export-findings / diff-runs / verify-checks (+ tests)
├── .github/actions/verify/         the reusable GitHub Action (validate, export, diff, checks, SARIF upload)
├── CONTRIBUTING.md                 how to contribute templates and docs
├── CODE_OF_CONDUCT.md              Contributor Covenant
├── SECURITY.md                     how to report a vulnerability
├── CHANGELOG.md                    Keep a Changelog format
├── templates/
│   └── README.template.md          canonical README skeleton
├── .github/
│   ├── workflows/docs.yml          docs CI (markdown lint, link-check, emoji guard)
│   ├── PULL_REQUEST_TEMPLATE.md
│   └── ISSUE_TEMPLATE/             findings + new-template + chooser config
├── web/                            landing page (Next.js 16) → auditor.rapold.io
├── mcp/                            stdio MCP server (exposes the prompts as agent tools)
└── audit-prompts/
    ├── full-audit-master-prompt.md                (orchestrator)
    ├── security-audit-master-prompt.md
    ├── repo-audit-master-prompt.md
    ├── frontend-audit-master-prompt.md
    ├── api-audit-master-prompt.md
    ├── performance-audit-master-prompt.md
    ├── data-audit-master-prompt.md
    ├── infrastructure-audit-master-prompt.md
    ├── ai-llm-audit-master-prompt.md
    ├── compliance-privacy-audit-master-prompt.md
    ├── accessibility-audit-master-prompt.md
    ├── documentation-audit-master-prompt.md
    ├── content-audit-master-prompt.md
    └── lean-audit-master-prompt.md

The landing page lives in web/ — its stack, local development, and deployment are documented in web/README.md.

Contributing

New templates follow the house structure so they compose with the rest: Mission + Universality, operating principles, a P0–P3 severity scale, the Phase 0–5 pipeline, a shared finding schema, a mandatory issue-output block, and a Definition of Done. Keep them provider-agnostic, map findings to recognized standards, and make every recommendation immediately actionable. See CONTRIBUTING.md, and follow DOCUMENTATION-STANDARD.md for all Markdown — including no emojis in headings.

Changelog and license

Changes are recorded in CHANGELOG.md (Keep a Changelog; SemVer). Licensed under MIT — see LICENSE.