Data Retention Enforcement Prompt for Session Cleanup

This prompt is for platform operators and SREs who need to enforce data retention policies at the moment an agent session terminates. The job-to-be-done is producing a complete, auditable cleanup action plan that identifies every artifact—tool-cached data, temporary files, context artifacts, and derived datasets—that must be deleted based on a specific retention policy. The ideal user is an AI infrastructure engineer integrating this prompt into a session lifecycle manager or a compliance engineer building automated evidence collection for data deletion workflows. You need a defined data retention policy document, a session manifest listing all tools invoked and their output destinations, and the agent's execution trace to use this prompt effectively.

Do not use this prompt for real-time data minimization during active sessions—that belongs to the Data Minimization Prompt for Search Tool Queries. Do not use it for PII redaction before tool dispatch, which is covered by the PII Redaction Prompt. This prompt is specifically for the terminal cleanup phase, where the session is ending and you need a defensible record that all data subject to retention limits has been identified and scheduled for deletion. It assumes the session has already been logged and that tool outputs are traceable to storage locations. If your system lacks session-level artifact tracking, you will need to implement that instrumentation first before this prompt can produce reliable results.

The prompt requires a structured retention policy as input—not vague principles. You must supply concrete rules like 'delete all temporary files after 30 days' or 'remove cached API responses immediately upon session close.' The output is a machine-readable cleanup plan, not a conversational summary. You should wire the output directly into a deletion orchestrator that executes the plan and logs confirmation. Human review is required when the policy matching confidence is below your threshold or when the agent identifies artifacts in unregistered storage locations. Start by ensuring your session manifest includes tool name, output path, data classification tag, and creation timestamp for every artifact before invoking this prompt.

This prompt template is designed to be copied directly into your agent's system instructions or a dedicated cleanup step at session termination. It forces the model to act as an auditor, systematically identifying all artifacts created during a session and mapping each one to a specific rule in your data retention policy. The output is a structured action plan, not just a summary, making it suitable for automated execution or human review before deletion.

textCopy
You are a data retention enforcement agent. Your task is to produce a session cleanup action plan at termination.

## INPUTS
- Session Artifact Inventory: [SESSION_ARTIFACT_INVENTORY]
- Data Retention Policy: [DATA_RETENTION_POLICY]
- Session Metadata: [SESSION_METADATA]

## OUTPUT_SCHEMA
Return a JSON object with the following structure:
{
  "session_id": "string",
  "cleanup_plan": [
    {
      "artifact_id": "string",
      "artifact_type": "string",
      "storage_location": "string",
      "contains_pii": boolean,
      "applicable_policy_rule": "string",
      "required_action": "DELETE | ARCHIVE | ANONYMIZE | RETAIN",
      "retention_deadline_utc": "ISO8601 timestamp or null",
      "justification": "string"
    }
  ],
  "completeness_check": {
    "total_artifacts_in_inventory": number,
    "artifacts_in_plan": number,
    "unmatched_artifacts": ["string"]
  }
}

## CONSTRAINTS
1. Every artifact in the Session Artifact Inventory MUST appear in the cleanup plan. If no policy rule matches, set required_action to "RETAIN" and note the gap in justification.
2. For artifacts containing PII, the justification must cite the specific regulatory basis (e.g., GDPR Art. 17, CCPA 1798.105).
3. If the Data Retention Policy specifies a duration, calculate the retention_deadline_utc from the session end time in Session Metadata.
4. Do not invent artifacts. Only process what is in the inventory.
5. If the inventory is empty, return a plan with an empty cleanup_plan array and a note in completeness_check.

## RISK_LEVEL
HIGH — Incorrect deletion can cause data loss; incorrect retention can cause compliance violations. Output must be reviewed by a human before execution.

To adapt this template, replace the placeholders with data from your application layer. [SESSION_ARTIFACT_INVENTORY] should be a structured list generated by your agent harness—do not rely on the model to recall what happened. [DATA_RETENTION_POLICY] is your organization's policy document, injected as text or a structured schema. [SESSION_METADATA] must include at minimum the session_end_time_utc and tenant_id. The output schema is designed to be validated programmatically: check that total_artifacts_in_inventory equals artifacts_in_plan and that every artifact_id from the input appears in the output. Any mismatch is a critical failure that should block automated deletion and escalate for human review.

This prompt is designed to sit at the end of an agent session lifecycle, triggered by a session termination event in your platform. The harness should invoke the prompt after the agent has completed its work but before the session context is destroyed. Pass the full session transcript, tool call history, and any file or cache manifests as [SESSION_CONTEXT]. The [RETENTION_POLICY] input should be a structured policy document—ideally a JSON or YAML object—that maps data categories to retention durations, deletion methods, and any legal-hold overrides. Do not rely on the model to recall policy details from training; always supply the policy inline to prevent drift when policies change.

The output of this prompt is a structured cleanup action plan. Your harness must validate this plan before executing any destructive actions. Implement a schema validator that checks each action item for required fields: artifact_id, artifact_type, storage_location, policy_rule_applied, action (e.g., DELETE, ARCHIVE, HOLD), and verification_step. If the plan references artifacts not present in the session manifest, flag them for human review. For high-risk environments, insert a mandatory human approval gate before executing any DELETE action. Log the full prompt input, output, and validation result to an immutable audit store for compliance evidence. Use a model with strong structured output support—GPT-4o or Claude 3.5 Sonnet with JSON mode enabled—and set temperature=0 to maximize consistency across runs.

Retry logic should be conservative. If validation fails due to missing fields or malformed JSON, retry once with the validation errors appended to the prompt as [PREVIOUS_ERRORS]. If the retry also fails, escalate to a human operator and do not proceed with automated cleanup. For artifact discovery completeness, run a post-execution verification step that queries your storage systems directly to confirm that all artifacts identified in the plan were actually removed or archived. Any mismatch between the plan and observed state should generate an alert and a follow-up remediation ticket. Avoid wiring this prompt into a fully autonomous deletion loop—always keep a human-in-the-loop for the final execution stage, especially when the session involves multi-tenant data stores or regulated data categories.