Unanswered Question Re-Surfacing Prompt

This prompt is for chat assistant builders who need to recover dropped user questions without disrupting the current conversation flow. The core job-to-be-done is: given the current conversation state and a list of previously unanswered questions, generate a natural re-surfacing message that reintroduces the question at an appropriate moment. The ideal user is an engineering lead or developer integrating this into a production chat system where dropped questions are a known failure mode—common in support agents, task-oriented copilots, and any multi-turn assistant where users ask multiple questions across turns.

Use this prompt when you have a reliable upstream detection system that identifies unanswered questions with turn references and severity indicators. The prompt requires [CURRENT_CONVERSATION], [UNANSWERED_QUESTIONS] (a structured list with question text, turn reference, and age), and [CONVERSATION_PHASE] to determine timing appropriateness. Do not use this prompt when the current turn is mid-clarification, when the user is actively correcting a prior assistant error, or when the unanswered question is older than the session's relevance window. The prompt is designed for single-question re-surfacing per invocation; batching multiple unanswered questions requires a separate prioritization step before calling this prompt.

Before wiring this into production, implement a timing gate that checks whether the current conversation phase is appropriate for interruption. The prompt includes a [TIMING_CHECK] constraint that should be evaluated programmatically before invocation—do not rely on the model alone to decide when to re-surface. Pair this prompt with the Pending Question Detection Prompt to build the unanswered question list, and use the Unanswered Question Severity Scoring Prompt to prioritize which question to re-surface first. For high-stakes domains like healthcare or legal support, require human review of the re-surfacing message before it reaches the user, and log every re-surfacing event with the question reference, timing decision, and user response for auditability.

The prompt template below is designed to be copied directly into your prompt management system or codebase. It accepts the current conversation state and a structured list of previously unanswered questions, then produces a single re-surfacing message that reintroduces the dropped question in a way that feels helpful rather than disruptive. The template uses square-bracket placeholders for all dynamic inputs—replace these with your application's actual data before sending the request to the model.

textCopy
You are an assistant that helps users without dropping their questions. Your task is to reintroduce a previously unanswered question into the current conversation in a natural, non-disruptive way.

## CURRENT CONVERSATION STATE
[CONVERSATION_HISTORY]

## UNANSWERED QUESTIONS
Below is a list of questions the user asked earlier that have not yet been answered. Each entry includes the original question, the turn when it was asked, and a severity score indicating how important it is to address.

[UNANSWERED_QUESTIONS_JSON]

## CONSTRAINTS
- Do NOT reintroduce a question if the user's current turn is actively addressing a high-urgency issue or if the user is mid-task on something that would be disrupted by an interruption.
- If the current conversation topic naturally connects to an unanswered question, weave the re-surfacing into the flow. Otherwise, wait for a natural pause or transition point.
- If multiple questions are pending, surface the highest-severity question first. Do not dump a list of forgotten questions all at once.
- If a question has become irrelevant due to later conversation turns, do not re-surface it. Instead, note that it has been implicitly resolved.
- Acknowledge the delay briefly without over-apologizing. One short acknowledgment is sufficient.
- Maintain the same tone and helpfulness level as the rest of the conversation.

## OUTPUT FORMAT
Return a JSON object with the following fields:
- `should_surface` (boolean): Whether now is an appropriate time to re-surface any unanswered question.
- `question_to_surface` (string or null): The ID of the question being surfaced, from the input list.
- `surface_message` (string or null): The exact message to send to the user. Write this as a natural conversational turn.
- `implicitly_resolved` (array of strings): List of question IDs that appear to have been resolved by later conversation turns.
- `reasoning` (string): Brief explanation of the decision, for logging and debugging.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [CONVERSATION_HISTORY] with the recent turns formatted as a clear dialogue transcript. [UNANSWERED_QUESTIONS_JSON] should be a JSON array of objects, each containing at minimum an id, question_text, turn_number, and severity_score. The [FEW_SHOT_EXAMPLES] placeholder should be populated with 2–3 examples showing both appropriate surfacing moments and appropriate deferral decisions. Set [RISK_LEVEL] to low, medium, or high to adjust the model's caution—high-risk domains like healthcare or finance should bias toward surfacing rather than silently dropping questions. Before deploying, validate that the output JSON parses correctly and that surface_message is never injected when should_surface is false.

Integrating the re-surfacing prompt into an application requires a stateful middleware layer that tracks unanswered questions across turns. The application must maintain a structured list of pending questions—each with a unique ID, the original turn reference, a severity score, and a staleness timestamp—outside the model's context window. Before each assistant response generation, the harness checks this pending queue. If items exist and a set of gating conditions are met (e.g., the current turn is not itself a clarification, the user hasn't explicitly asked to pause, and a minimum number of turns have passed since the question was dropped), the harness injects the re-surfacing prompt with the current conversation context and the oldest or highest-severity pending question. The model's output is then parsed to extract the re-surfacing message, which is prepended to the main assistant response.

Validation is critical because a poorly timed or tonally off re-surfacing message erodes user trust faster than a dropped question. The harness should validate the model's output against a strict schema: the response must contain a re_surface_message string and a should_surface boolean. If should_surface is false, the harness respects the model's decision and leaves the question in the pending queue for a future turn. If true, the harness logs the re-surfacing attempt, marks the question as surfaced with a timestamp, and increments a retry counter. If a question has been surfaced more than a configured maximum (e.g., 2 attempts) without user engagement, the harness should escalate it to a human review queue or append it to a session closeout checklist instead of retrying. Implement a circuit breaker: if the user explicitly dismisses the re-surfaced question or changes the topic sharply, immediately suppress further re-surfacing for that session to prevent harassment.

Model choice matters here. This prompt requires strong instruction-following and nuanced tone control. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro are all suitable. Avoid smaller or older models that may conflate re-surfacing with answering the question inline or produce overly apologetic language. Set temperature low (0.0–0.2) to keep the should_surface decision stable. Log every re-surfacing decision—including the pending question ID, the model's boolean output, the generated message, and the user's subsequent response—to a structured observability pipeline. This data is essential for tuning the gating conditions, evaluating the prompt's effectiveness, and detecting regressions when the prompt or model version changes. Never surface a question that contains PII or sensitive content without redaction; apply a PII scrub to the pending question text before it enters the prompt template.