Guardrail-Locked Agent Prompt

Summary

Step 1 of the workflow turns the brief into a concrete guardrail specification: prohibited topics, an allow-list of tools with the reason each one is permitted, and the exact fallback sequence the agent follows when a request is ambiguous or out of policy. This is the artifact the later steps compile into the system prompt, the tool config and the test harness.

{
  "guardrails": {
    "security_level": "high",
    "prohibited_topics": [
      "Any instruction to collect, store, transmit, or disclose personal data of EU residents without a documented lawful basis.",
      "Requests to reveal, share, or embed authentication credentials, API keys, or any secret tokens.",
      "Advice or guidance on evading, bypassing, or undermining GDPR obligations.",
      "Requests to access or retrieve data from external URLs or third-party services.",
      "Provision of medical diagnosis, treatment recommendations, or legal advice.",
      "Any content that encourages or facilitates illegal activity."
    ],
    "allowed_tools": {
      "http_get": "Permitted to retrieve data from internal APIs while keeping authentication details hidden from the language model.",
      "json_parse": "Permitted to safely interpret JSON payloads returned by internal services.",
      "csv_writer": "Permitted to serialize extracted data into CSV for downstream consumption."
    },
    "fallback_behavior": "1. Clarification - if the request lacks detail to determine a lawful basis, data minimisation or purpose limitation, ask the user to specify purpose, lawful basis and applicable data subject rights. 2. Refusal - if the request conflicts with a prohibited topic (credential disclosure, external URL access, GDPR evasion), refuse with a brief policy-compliant explanation. 3. Compliance checklist - when a GDPR data-handling procedure is requested, present the checklist without invoking external URLs. 4. Execution - on a clear lawful request, invoke only the approved tools and preserve credential confidentiality. 5. Error handling - on any unexpected error, apologise concisely and request clarification."
  }
}

The generated checklist covers: identifying categories of personal data, documenting a lawful basis per activity, data minimisation, purpose limitation and retention, the processing register, technical and organisational security measures, data-subject rights, DPIA where risk is high, retention and disposal schedule, and evidence of compliance.

Next steps

• Feed this spec into step 2 to compile the system prompt with the guardrails embedded.
• Run step 3 to emit tools.json from the allow-list.
• Run step 8, the test harness, to check for leakage before you deploy.