Compliance Gate Review Prompt for Tool Actions

This prompt is designed for compliance engineers and platform operators who need to insert a pre-execution regulatory review gate into an agent's tool-use workflow. The primary job-to-be-done is surfacing the regulatory context of a proposed action—such as a database mutation, a cross-border data transfer, or a financial transaction—so that a human reviewer or an automated policy engine can make a consistent, auditable decision. The ideal user is someone who understands the control obligations under frameworks like SOX, GDPR, HIPAA, or PCI-DSS but needs a structured, repeatable way to enforce them before an agent acts. The required context includes the proposed tool name, its full input arguments, the data classifications involved, and the applicable jurisdictional or regulatory framework.

Use this prompt when an agent proposes an action that touches regulated data, crosses a jurisdictional boundary, or triggers a control obligation that cannot be fully automated. Concrete examples include: a support agent preparing to export user data in response to a DSAR, an infrastructure agent proposing a firewall rule change that affects a PCI-DSS in-scope network segment, or a database agent about to execute a DROP TABLE on a schema containing PHI. In each case, the prompt forces the model to produce a structured compliance assessment with applicable controls, required evidence, and a clear pass/fail determination. The output is not a replacement for a human compliance officer; it is a pre-execution filter that reduces the cognitive load on reviewers and prevents actions from slipping through without a regulatory check.

Do not use this prompt for low-risk, non-regulated actions where the overhead of a structured review gate would slow down legitimate automation. If an agent is updating a non-sensitive configuration flag in a development environment or fetching read-only telemetry data with no PII, a full compliance assessment adds friction without benefit. Similarly, do not rely on this prompt as the sole control for actions that require real-time human judgment in crisis scenarios—such as incident response break-glass procedures—where a structured gate could introduce unacceptable latency. In those cases, pair the prompt with a parallel escalation path that logs the action for post-hoc review. The next step after understanding when to apply this prompt is to wire it into your agent's tool-calling middleware so that governed actions are intercepted, assessed, and routed to the appropriate review queue before execution proceeds.

This prompt template is designed to be pasted directly into your agent harness. It instructs the model to act as a compliance review gate, analyzing a proposed tool action against a provided control framework before execution. The model will produce a structured assessment, including a pass/fail determination, applicable controls, and required evidence, rather than simply approving or denying the request.

textCopy
You are a Compliance Gate Review Agent. Your sole function is to evaluate a proposed tool action against a predefined set of regulatory controls and internal policies. You must produce a structured, auditable assessment. Do not execute the tool. Do not offer opinions outside the provided framework.

## INPUT
- **Proposed Action:** [ACTION_DESCRIPTION]
- **Actor:** [ACTOR_IDENTITY]
- **Target Resource:** [TARGET_RESOURCE]
- **Context:** [CONTEXT]

## CONTROL FRAMEWORK
[CONTROL_FRAMEWORK]

## CONSTRAINTS
- You must map the proposed action to specific controls in the framework.
- For each applicable control, state the evidence required for compliance.
- If any required evidence is missing, the determination must be "FAIL - INSUFFICIENT EVIDENCE."
- If the action violates a control, the determination must be "FAIL - CONTROL VIOLATION."
- If all controls are satisfied with provided evidence, the determination must be "PASS."
- Do not infer or assume evidence that is not explicitly provided in the [CONTEXT].

## OUTPUT_SCHEMA
{
  "assessment_id": "string",
  "timestamp": "string (ISO 8601)",
  "determination": "PASS | FAIL - INSUFFICIENT EVIDENCE | FAIL - CONTROL VIOLATION",
  "mapped_controls": [
    {
      "control_id": "string",
      "control_description": "string",
      "applicability_reasoning": "string",
      "required_evidence": ["string"],
      "evidence_provided": ["string"] | null,
      "compliance_status": "COMPLIANT | NON_COMPLIANT | EVIDENCE_MISSING"
    }
  ],
  "summary": "A concise summary of the assessment for an auditor.",
  "recommended_action": "PROCEED | ESCALATE_FOR_REVIEW | BLOCK"
}

To adapt this template, replace the square-bracket placeholders with real data. The [CONTROL_FRAMEWORK] should be a structured list of rules, such as 'SOX Section 404: All financial data modifications require a documented and pre-approved change ticket.' The [CONTEXT] should contain the evidence gathered so far, like the change ticket ID or the data classification. The prompt's strict constraints and output schema are designed to minimize false positives by forcing the model to link a control to missing evidence, rather than issuing a vague denial. Always log the raw input and the structured JSON output for your audit trail.

The Compliance Gate Review prompt is not a standalone chat interaction; it is a synchronous control point inserted into an agent's tool-calling loop. When the agent proposes a regulated action—such as a financial transaction, a data export, or a configuration change—the application layer intercepts the tool call, assembles the prompt with the proposed action details, and sends it to a fast, instruction-following model (e.g., GPT-4o, Claude 3.5 Sonnet) for a structured compliance assessment. The model returns a JSON object containing a decision field (PASS, FAIL, or REVIEW), a list of applicable controls, and a risk summary. The harness reads this output and either allows the tool call to proceed, blocks it, or routes it to a human review queue.

To implement this reliably, wrap the prompt call in a lightweight gate function that enforces a strict output contract. Use a JSON schema validator (such as jsonschema in Python or zod in TypeScript) to confirm the model's response contains the required fields before acting on the decision. If validation fails, retry once with a repair prompt that includes the validation error. Log every assessment—including the proposed action, the model's decision, the controls cited, and the raw response—to an append-only audit store. For high-risk domains, require a second reviewer model or a human-in-the-loop step when the REVIEW decision is returned or when the risk score exceeds a configurable threshold. The harness should also enforce a hard timeout (e.g., 5 seconds) to prevent the gate from becoming a bottleneck in latency-sensitive workflows.

When integrating this into an agent framework like LangChain, CrewAI, or a custom loop, treat the compliance gate as a tool-wrapper or middleware rather than a separate agent. The gate should be stateless and idempotent: the same proposed action should produce the same assessment. Avoid placing the compliance prompt inside the agent's main system message, as that increases the risk of prompt injection and makes the gate harder to test in isolation. Instead, keep it as a separate, versioned prompt resource that can be updated, A/B tested, and rolled back independently of the agent's core behavior. Start with a narrow set of regulated action types, monitor the false-positive rate (actions incorrectly blocked), and tune the prompt's examples and risk thresholds before expanding to broader tool coverage.