Gradual Topic Drift Detection Prompt

This prompt is designed for AI engineers and product teams managing long-running, multi-turn assistant sessions where the user's subject changes incrementally over 10, 20, or more turns. Unlike a sudden topic shift, which is easy to catch with a simple classifier, gradual drift silently pollutes the context window with irrelevant history, degrading response quality, increasing latency, and wasting token budget. The core job-to-be-done is automated context hygiene: detecting that cumulative semantic change has crossed a meaningful threshold so your application can trigger a context pruning, summarization, or system prompt switch before the assistant starts blending unrelated topics. The ideal user is a production engineer responsible for session state management in a multi-domain assistant, support bot, or copilot that handles conversations spanning multiple subtopics.

You should use this prompt when you have a session history of at least 10 turns and need a structured signal—a drift score, a turn index where the threshold was crossed, and a recommended action—that your application code can act on. The prompt expects a full conversation transcript as input, along with a configurable drift threshold and a sensitivity parameter that controls how much incremental change is tolerated before a reset is triggered. It is not a single-turn classifier; it requires the full turn history to compute cumulative semantic distance. Wire this into your session management middleware so that after every N turns (or at a configurable interval), the prompt evaluates the conversation and returns a decision. Log the drift score and threshold crossing events for observability, and use the output to drive downstream actions like context window truncation or summarization.

Do not use this prompt for detecting abrupt, single-message topic jumps—a simpler shift detection prompt with lower latency and token cost is a better fit for that case. Do not use it for real-time, per-turn classification in latency-sensitive chat streams where you cannot afford the context overhead of processing the full history. This prompt is also not a replacement for a conversation phase classifier if your workflow requires mapping turns to a predefined phase taxonomy; it measures cumulative drift, not discrete phase membership. Finally, avoid using this prompt in sessions with fewer than 8-10 turns, as the drift signal will be noisy and unreliable without sufficient history to establish a baseline. For high-stakes applications where a false positive could drop critical context, always pair the prompt's output with a lightweight confirmation check before executing a destructive context reset.

This prompt template is designed to be pasted directly into your application's model call. It instructs the model to analyze a multi-turn conversation transcript and detect when the topic has drifted significantly from the original subject, even if the change was gradual. The output is a structured JSON object containing a drift score, the critical turn index, and a rationale, which your application can use to trigger context resets, summarization, or user confirmations.

textCopy
You are a conversation analyst. Your task is to examine the provided multi-turn conversation transcript and detect gradual topic drift. Gradual drift occurs when the subject changes incrementally over several turns, rather than in a single abrupt shift.

Analyze the conversation from the first turn to the last. Identify the primary initial topic. For each subsequent turn, assess its semantic distance from the initial topic. Determine the turn index where the cumulative shift first crosses a significant threshold, indicating a new, distinct topic has become dominant.

[CONVERSATION_TRANSCRIPT]

Output a single JSON object with the following schema:
{
  "initial_topic": "string (concise label for the starting subject)",
  "drift_detected": boolean,
  "drift_score": "float between 0.0 and 1.0, where 1.0 represents a complete topic change",
  "critical_turn_index": "integer or null (the first turn index where the cumulative drift crossed the threshold)",
  "new_dominant_topic": "string or null (label for the new topic if drift is detected)",
  "rationale": "string (brief explanation of the evidence for the drift score and critical turn)"
}

[CONSTRAINTS]
- A turn index is the 0-based position in the transcript.
- A "significant threshold" is a cumulative semantic shift of 0.7 or higher on a conceptual 0.0 to 1.0 scale.
- Do not confuse a brief digression that returns to the main topic with a genuine drift.
- If no drift is detected, the critical_turn_index and new_dominant_topic should be null.

To adapt this template, replace the [CONVERSATION_TRANSCRIPT] placeholder with your formatted dialogue history. You can adjust the drift threshold in the [CONSTRAINTS] section to make the detection more or less sensitive to change. For high-stakes applications, such as legal or healthcare consultations, always route the output for human review before automatically resetting context, and log the rationale field for auditability. Before deploying, test this prompt against a golden dataset of conversations with known drift points to calibrate the threshold and minimize both false positives on deep-dive clarifications and false negatives on subtle subject changes.

The Gradual Topic Drift Detection prompt is designed to run as a background analysis step within a long-running conversation loop, not as a direct user-facing response. It should be invoked periodically—for example, every N turns or when a turn-level coherence score drops below a threshold—to assess whether the cumulative semantic distance from a reference anchor point has crossed a defined drift boundary. The prompt expects a structured conversation transcript, a reference turn index marking the last confirmed topic, and a configurable drift threshold. The output is a drift score, the turn index where the threshold was first crossed, and a recommended context action (e.g., reset, compress, or retain). This output feeds into your context window manager, not directly to the user.

To integrate this into an application, implement a stateless function that assembles the prompt with the required inputs and calls your model API. Validate the output against a strict schema before acting on it: the drift score must be a float between 0.0 and 1.0, the crossing turn index must be an integer within the transcript range or null if no crossing occurred, and the recommended action must be one of a predefined enum (reset, compress, retain). If validation fails, retry once with the validation error appended to the prompt as a correction hint. If the retry also fails, default to a safe fallback—typically retain—and log the failure for offline review. For high-stakes applications where incorrect context resets could lose critical user state, route reset recommendations with confidence below 0.85 to a human approval queue before execution.

Model choice matters here. This prompt requires strong semantic comparison capabilities over long contexts, making it a good fit for models with large context windows and strong instruction-following on structured output tasks (e.g., Claude 3.5 Sonnet, GPT-4o). Avoid running this on small or fast models where drift scoring accuracy degrades, as false positives on drift detection will cause unnecessary context resets that frustrate users. Log every invocation with the drift score, crossing index, recommended action, and the final action taken by your system. This audit trail is essential for tuning the drift threshold and diagnosing cases where the system either over-resets on natural conversation evolution or under-resets on genuine topic changes. Next, pair this prompt with a Context Carry-Forward vs Reset Decision prompt to handle the actual context surgery once drift is confirmed.