Session-Long Persona Consistency Prompt

This prompt is for AI product teams and engineering leads who ship conversational assistants with a defined persona—a specific tone, role, expertise boundary, and interaction style. The job-to-be-done is automated quality assurance: you need to detect when a long-running session causes the assistant's persona to drift, erode, or contradict itself. The ideal user is a developer or QA engineer integrating this check into a regression test suite, a CI/CD pipeline for prompt changes, or a periodic production sampling job. You must provide a sample of assistant turns from a single session, the original persona specification or system prompt, and the evaluation criteria you care about most (e.g., tone, expertise claims, refusal style).

Do not use this prompt for real-time enforcement during a live user session—it is an offline evaluation tool, not a guardrail. It is not a replacement for system prompt engineering or instruction hierarchy design. If your assistant has no defined persona beyond generic helpfulness, this prompt will produce low-signal results. The prompt works best when the persona is explicitly documented with concrete traits (e.g., 'speaks like a senior staff engineer, never apologizes for correct answers, defers to human judgment on architecture decisions'). Vague personas produce vague drift reports. For high-risk domains like healthcare or legal, always pair the output with human review before acting on drift findings.

After running this prompt, you should expect a structured consistency report with specific drift examples and re-anchoring suggestions. Use the report to decide whether to adjust your system prompt, add re-injection triggers, or update your few-shot examples. If the report shows no drift but users are complaining about inconsistency, your persona specification may be too loosely defined—tighten it before re-running the evaluation. Avoid treating this as a one-time check; run it against production sessions weekly and after every system prompt change.

This template provides the core evaluation prompt for detecting persona drift across a multi-turn conversational session. It is designed to be inserted into an evaluation harness that samples turns from a production conversation, compares them against the assistant's defined persona contract, and produces a structured consistency report. The template uses square-bracket placeholders that you must replace with your specific persona definition, session transcript, and output requirements before use.

codeCopy
You are an AI persona consistency auditor. Your task is to evaluate whether an assistant maintained its defined persona across a multi-turn conversation session.

## PERSONA CONTRACT
[PERSONA_DEFINITION]

## SESSION TRANSCRIPT
[SAMPLED_TURNS]

## EVALUATION INSTRUCTIONS
Analyze the assistant's responses across all sampled turns and evaluate consistency against the persona contract. For each persona dimension defined in the contract, determine whether the assistant's behavior remained stable, drifted, or violated the contract.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "session_id": "[SESSION_ID]",
  "turns_evaluated": [TURN_COUNT],
  "overall_consistency_score": [0.0-1.0],
  "dimension_results": [
    {
      "dimension": "[DIMENSION_NAME]",
      "score": [0.0-1.0],
      "status": "stable|minor_drift|significant_drift|violation",
      "drift_examples": [
        {
          "turn_number": [INTEGER],
          "assistant_text": "[EXCERPT]",
          "deviation_description": "[EXPLANATION]",
          "severity": "low|medium|high|critical"
        }
      ],
      "re_anchoring_suggestion": "[CORRECTIVE_GUIDANCE]"
    }
  ],
  "global_drift_patterns": ["[PATTERN_DESCRIPTION]"],
  "re_anchoring_priority": ["[DIMENSION_NAME]"],
  "requires_immediate_correction": [BOOLEAN]
}

## CONSTRAINTS
[EVALUATION_CONSTRAINTS]

## EXAMPLES
[FEW_SHOT_EXAMPLES]

Adaptation guidance: Replace [PERSONA_DEFINITION] with your assistant's complete persona contract, including tone, expertise boundaries, interaction style, and behavioral policies. [SAMPLED_TURNS] should contain a representative sample of turns from the session, not necessarily every turn—sample at regular intervals and at known drift points. [EVALUATION_CONSTRAINTS] should specify any domain-specific rules, such as tolerance thresholds for minor tone variation or strictness levels for regulated industries. [FEW_SHOT_EXAMPLES] should include at least two annotated examples showing both stable and drifted behavior to calibrate the evaluator. For high-risk deployments, add a human review step before accepting automated consistency scores below 0.8. Do not use this prompt as a real-time guardrail; it is designed for offline audit and monitoring, not for inline intervention during active conversations.

The Session-Long Persona Consistency Prompt is designed as an offline evaluation tool, not a real-time interceptor. It should be wired into a post-session or periodic sampling pipeline that pulls completed conversation transcripts and runs them through the prompt to produce a consistency report. The prompt expects a full session transcript and the original persona definition as inputs. It returns a structured report with drift examples, severity scores, and re-anchoring suggestions. This output is consumed by product and QA teams, not by the assistant itself during a live session.

For implementation, build a scheduled job or event-driven trigger that fires when a session reaches a configurable turn threshold (e.g., 20 turns) or a session-close event. The job should fetch the session transcript from your conversation store, retrieve the active persona definition from your prompt configuration system, and assemble the prompt with both inputs. Use a model with strong instruction-following and structured output capabilities—GPT-4o or Claude 3.5 Sonnet are appropriate. Set temperature to 0 or very low (0.0–0.1) to maximize consistency in the evaluation itself. Parse the output against a defined JSON schema that includes persona_traits, drift_instances (with turn numbers, trait violated, severity, and evidence), overall_stability_score, and re_anchoring_suggestions. If the output fails schema validation, retry once with a repair prompt that includes the validation error. Log all evaluation runs, including the raw prompt, model response, parse success/failure, and final report, to your observability platform for audit and trend analysis.

Do not run this prompt on every turn in production—it is too expensive and adds latency that degrades the user experience. Instead, use it as a sampling and alerting mechanism. Set thresholds on overall_stability_score to trigger alerts when persona drift exceeds acceptable limits. Feed drift instances back into your prompt engineering workflow to identify weak instructions that erode over long sessions. The re-anchoring suggestions are diagnostic artifacts for prompt authors, not instructions to inject mid-session. If you need real-time drift prevention, pair this evaluation prompt with a separate Instruction Re-Injection Trigger Prompt that runs inline and decides when to re-insert persona instructions into the active context window.