Shield pack: 27 compliance specialists and 11 workflows
Shield routes compliance questions to 27 framework specialists and 11 GDPR and AI Act workflows, and adds assurance cases, an AI register and review rules. Not legal advice.
Shield turns BMAD+ into a governance, risk and compliance (GRC) assistant. One orchestrator reads your question, spots the framework or frameworks it concerns, and routes it to a specialist agent: 27 of them, each dedicated to one regulation or standard, in six categories. Eleven further workflows cover the recurring GDPR and EU AI Act documents (impact assessments, breach response, privacy notices, cookie policies, AI risk classification). On top of that, Shield ships three evidence tools backed by real CLI commands: an assurance case, an AI processing register and compliance review rules.
Warning· Not legal advice
Every Shield report carries an AI-generated marking and a not-legal-advice notice. The output may be incomplete or out of date and does not replace a lawyer, a data protection officer or a certified auditor. A qualified person must review it before any compliance decision, filing or attestation.
You must show that a product meets a regulation or a standard (GDPR, SOC 2, ISO 27001, PCI DSS, HIPAA, NIS2, DORA, EU AI Act and others) and want a structured gap analysis rather than a blank page.
You need a standard GDPR or AI Act document: DPIA, breach notification, legitimate-interest test, privacy notice, cookie policy, AI risk class.
You want audit evidence that is checked by commands, not asserted: an assurance case tied to a commit.
You want a code review that says which compliance controls a change touches.
Shield is a packaged pack: everything sits in .agents/skills/pack-shield/. That folder holds the pack skill SKILL.md, the orchestrator shield-orchestrator.md, categories/ (the 27 specialists and the 11 workflows), references/ (79 regulatory reference files), shared/ (cross-framework mapper, gap-analysis and audit-report templates, assurance-case and AI-register procedures with their templates) and review-rules/ (the six compliance review rules and their index). bmad-plus doctor checks that SKILL.md and review-rules/index.yaml are present.
ISO/IEC 27001, NIST CSF 2.0, NIST SP 800-53 Rev. 5, CIS Controls v8, NIS2, Australian ISM
Industry Compliance
6
SOC 2, PCI DSS v4.0.1, HIPAA, SWIFT CSP, DORA, FedRAMP
Defense & Export
4
CMMC 2.0, ITAR, EAR, TSA Security Directives
AI Governance
3
EU AI Act 2024/1689, ISO/IEC 42001, NIST AI RMF 1.0
Accessibility & ESG
3
WCAG 2.2, Section 508, CSRD
The pack has 27 specialist agents, one per framework. The registry’s list of compliance tags names all 27, and the pack list of the installer reads “GRC compliance (27 frameworks)”.
A claim counts as supported only if the check behind it ran at the commit concerned, recorded in a hash-chained ledger. Missing, failed or stale evidence leaves the claim unsupported.
AI processing register
bmad-plus ai-register init, check
Compares _bmad/ai-processing-register.yaml with the AI integrations found in the project and warns about unregistered ones. Warnings never fail the command.
Compliance review rules
bmad-plus review scope, review rules <path>
Six path-scoped checklists tagged with controls such as ISO27001:A.8.5, SOC2:CC6.1 or GDPR:Art.32, so a review lists the controls a change touches.
The six review rules are shield-access-control, shield-personal-data, shield-cryptography, shield-logging, shield-ai-integrations and shield-change-and-supply-chain. They add to the built-in review rules and never replace them; a project can replace or disable them in _bmad/review-rules.yaml.
Core is added automatically. Shield needs only Node.js 20 for the installer; the evidence commands run from the same CLI.
In your project terminal
npx bmad-plus install --packs shield
02
Ask the orchestrator
Name the framework if you know it, or describe your situation; Shield asks two or three clarifying questions (jurisdiction, industry, data types) when the framework is ambiguous.
In your project terminal
Shield, audit my app for GDPR compliance
03
Check the install
The diagnostic confirms that the pack skill and the review rules index are present.
The shortest route to a single-framework code review. Shield recognises GDPR from the wording and hands over to the GDPR specialist. A general request for a “GDPR audit” or a “compliance assessment” goes to the privacy-advisor workflow (an assessment of the whole privacy programme) instead: to review code, ask about named files.
In your AI assistant
Shield, our mobile backend stores user profiles and push tokens. Review the code in services/api/ and the table definitions in db/migrations/ against the GDPR: list the personal data you find, the lawful basis you can infer for each, and the gaps.
What you should get The request goes to gdpr-agent, workflow 1 (code and system audit). The report follows the audit format: personal data identified, lawful-basis assessment, findings and a summary. It carries the AI-generated marking and the not-legal-advice notice.
When a request names several frameworks, Shield runs each specialist and merges the results.
In your AI assistant
Shield, gap analysis NIS2 vs DORA for our payment platform, which serves banks in three EU countries. We have a security incident process but no register of ICT providers. Show which obligations overlap, such as incident reporting and third-party risk, so we build one process and satisfy both, and give a prioritised remediation roadmap.
What you should get Both nis2-agent and dora-agent run, then Shield uses the cross-framework mapper (NIS2 ↔ DORA is one of the pairings it lists) to build the overlapping-control mapping and produces one consolidated report in the audit-report format: common controls (implement once, satisfy many), framework-specific gaps and a priority roadmap.
Use it ahead of an audit window, once you know which trust criteria are in scope.
In your AI assistant
Shield, generate a SOC 2 evidence checklist for Security and Availability. We run on AWS with GitHub Actions. For each criterion, name the artifact to collect, who owns it and how often it must be refreshed.
What you should get Routed to soc2-agent, whose audit-evidence section organises evidence by criterion. Expect a checklist you can turn into tickets. It does not certify anything: only an auditor can issue a SOC 2 report.
Use it when a processing is likely to carry a high risk for people, for example the systematic monitoring of employees or profiling.
In your AI assistant
Shield, run a DPIA for the driver-fatigue camera we are adding to our delivery vans. It films the driver’s face continuously, computes a fatigue score on the device and uploads only the score and a timestamp to the fleet dashboard. Footage stays on the device for 24 hours. Ask me for anything you need to describe the processing.
What you should get The dpia-sentinel workflow walks the steps of Article 35 (threshold, description, necessity and proportionality, risks, mitigation, DPO consultation, prior consultation) and produces a report in nine sections: project information, threshold assessment, processing description, necessity and proportionality, risks to data subjects, mitigation measures, DPO opinion, conclusion and sign-off. Sections 7 (DPO opinion) and 9 (sign-off) are left for the people concerned to fill in.
React to a personal data breach (breach-sentinel)#
Use it as soon as you suspect a breach: the 72-hour clock of Article 33 starts when you become aware of it.
In your AI assistant
Shield, we had a data breach. An exported customer CSV (names, emails, order history, about 4,200 people in the EU) was sent to the wrong external address this morning at 09:40. Walk me through the first hours and draft the authority notification.
What you should get The breach-sentinel workflow follows a timeline: detection and initial assessment (first hours), severity classification, authority notification by the 72-hour mark with the four Article 33(3) elements, data-subject notice when the severity requires it, then documentation and lessons learned. The drafts are for your DPO or lawyer to review before anything is sent.
Use it first, before you decide which obligations apply. Follow with the roles and FRIA workflows when the result is high risk.
In your AI assistant
Shield, classify my AI system under the AI Act. It ranks job applicants for a staffing agency from CVs and a short questionnaire. We build it ourselves and sell it to agencies in France and Germany.
What you should get The ai-act-classifier workflow identifies the system, checks the prohibited practices of Article 5, then high-risk status through Article 6 and Annex III, then limited and minimal risk, and returns a classification report with the reasoning, the applicable obligations and recommended next steps. Whether a given case is high risk is a legal conclusion: have it confirmed.
Use it for a release you must defend to an auditor. Run the commands from a clean working tree at the commit concerned.
In your AI assistant
Shield, build a security assurance case for release 2.3 of our mobile API. Claims: the container image scan reports no critical vulnerability, the API refuses TLS versions below 1.2, and last night’s backup restores into an empty database. For each claim, pick a command that exits non-zero when the claim is false, then run and verify the case and tell me which claims are supported.
What you should get The case file _bmad/assurance/<id>.yaml is started with bmad-plus assurance init, bmad-plus assurance run executes each declared command without a shell and appends the outcome to a hash-chained ledger, and bmad-plus assurance verify reports each claim as supported or unsupported, judging the check behind it as passed, failed, stale or missing. Set BMAD_PLUS_ASSURANCE_KEY (32 characters or more) and keep the ledger head from run to make deliberate edits detectable. The verdict says what the evidence is; it is not a certification.
Use it when you must answer which AI tools see your data and on what legal basis.
In your AI assistant
Shield, this quarter we added a code-assistant plugin, a support-chatbot MCP server and an OCR API. Create the AI processing register, enter these three and ask me for each one’s data, legal basis, retention and any transfer outside the EU. Then check the register and tell me which integrations the scan found that I did not list.
What you should getbmad-plus ai-register init creates _bmad/ai-processing-register.yaml from the Shield template; bmad-plus ai-register check then prints the review date and how many AI integrations are registered, with a warning for each unregistered one (BMAD+ adapters, tool folders, MCP servers). The check is a soft gate: warnings do not make it fail, and bmad-plus doctor reports the same.
No Python, no API key and no extra tool: Shield is instructions plus CLI commands. Your AI tool supplies the model.
One off-machine flow concerns the evidence commands, and it belongs to the bmad-plus CLI rather than to the pack: assurance-checks. The assurance command exists without Shield (the pack ships the procedure and the templates). bmad-plus assurance run executes the commands written in the case, with the same trust as the project’s own scripts, so a check may reach the network if its command does. Read a case written by someone else before running it. The key BMAD_PLUS_ASSURANCE_KEY stays in the environment and is never given to a check.
What you paste into a Shield conversation (code, data flows, incident details) goes to your AI tool’s model like any other prompt. Do not paste live personal data into an audit request.
Specialists are expert prompts built on reference files. Their regulatory content can be out of date, and a supported claim in an assurance case means only that a named check passed at a commit. Treat every output as a draft for a qualified person.
Tip· Installing the pack
--packs takes a comma-separated list of pack ids: npx bmad-plus install --packs core,shield. Core is always included, and all installs every pack.
Note· Ledger limits
Without BMAD_PLUS_ASSURANCE_KEY and a kept --ledger-head, the hash chain shows accidental or careless edits only: whoever can write the ledger can append a record that verifies. Say which protection applied when you present the result.
Note· Provenance
The specialist agents are adapted from an MIT-licensed open-source collection of GRC skills, and the GDPR and AI Act workflows were enriched with ideas from a professional skills catalogue. The notices are in THIRD-PARTY-LICENSES.md of the package.