Session Summary Injection Prompt on Resume

This prompt is designed for AI engineers and platform teams building persistent chat products where conversations span multiple sessions. The primary job-to-be-done is to reconstruct a minimal but sufficient context block from a stored session summary when a user returns. You need this when your application serializes conversation state on session end and must rehydrate it on resume. The ideal user is a developer integrating this into a state management layer—someone who already has a session store, a summary generation pipeline, and a context assembly function.

Use this prompt when you have a structured session summary containing prior turns, pending actions, unresolved questions, and user preferences. The prompt decides what to inject into the current context window and what to discard. It is not a replacement for full conversation history when the context budget allows it. Do not use this prompt for real-time turn-by-turn state tracking—that belongs in dialogue state management. Do not use it for long-term memory across weeks or months—that requires a separate personalization memory system. Do not use it when the session summary itself is stale or corrupted; run a session continuity validation check first.

The prompt requires a pre-generated session summary as input, typically produced by a session summarization pipeline. It expects the summary to include structured fields: a dialogue synopsis, a list of pending actions with status, unresolved questions, active constraints or preferences, and a timestamp. The output is a context-injection block ready to be prepended to the system prompt or inserted as a system message. Before deploying, validate that the injected context does not reintroduce stale facts, duplicate completed actions, or contradict the current user message. For high-stakes domains like healthcare or finance, require human review of the injected summary before the assistant responds. If the summary is missing critical fields, fall back to asking the user what they remember rather than injecting incomplete state.

This prompt template is designed to be called by your application logic immediately after a session resume event. Its job is to take a structured session summary—produced by your serialization layer—and convert it into a compact, context-efficient block that can be prepended to the model's system or developer message. The template uses square-bracket placeholders for all dynamic inputs, making it safe to store in a prompt library and instantiate at runtime. The output is a single text block, not a JSON object, because it is intended for direct injection into the next model request.

textCopy
You are resuming a conversation with a user after a pause. Below is a summary of the prior session. Use this context to maintain continuity, but do not repeat information the user already knows unless it is directly relevant to the current turn.

[SESSION_SUMMARY]

Active constraints from the prior session:
[ACTIVE_CONSTRAINTS]

Pending actions that were not yet completed:
[PENDING_ACTIONS]

Unresolved questions that still need answers:
[UNRESOLVED_QUESTIONS]

User preferences observed in the prior session:
[USER_PREFERENCES]

Instructions:
- Treat the summary as ground truth for what happened previously.
- If the user contradicts the summary, trust the user and update your understanding.
- Do not re-execute completed actions unless the user explicitly asks.
- Surface pending actions naturally when the conversation context makes them relevant.
- If an unresolved question becomes relevant, ask it once and only once.
- If [ACTIVE_CONSTRAINTS] is empty, ignore this section.
- If [PENDING_ACTIONS] is empty, do not invent follow-ups.
- If [UNRESOLVED_QUESTIONS] is empty, do not fabricate questions.
- If [USER_PREFERENCES] is empty, do not assume preferences.

To adapt this template, replace each bracketed placeholder with the corresponding field from your session state object. The [SESSION_SUMMARY] field should contain a concise narrative of the prior conversation, ideally generated by a dedicated summarization prompt. The [ACTIVE_CONSTRAINTS] field should list any time-sensitive rules, compliance boundaries, or user-imposed limits that remain in effect. [PENDING_ACTIONS] should enumerate tasks the assistant committed to but did not complete, while [UNRESOLVED_QUESTIONS] captures explicit questions the assistant asked that went unanswered. [USER_PREFERENCES] should include only preferences that were stated or demonstrated in the prior session, not inferred from user profiles. If any field has no content, insert the literal string "None" or an empty bullet list to prevent the model from hallucinating missing information. Before injecting this block into production, validate that the combined token count of all populated fields does not exceed your context budget allocation for session history—typically 20-30% of the total context window for most chat applications.

The session summary injection prompt is not a standalone component—it is a critical link in the conversation lifecycle chain. In production, this prompt fires when a user reconnects after a session boundary (timeout, explicit pause, device switch, or agent handoff). The application must first retrieve the serialized session state from your persistence layer, validate its integrity, and then pass the relevant fields into the prompt's placeholders. Do not inject raw, unvalidated state directly into the context window. A corrupted or stale summary will cause the assistant to confidently continue from a broken premise, which is worse than starting fresh.

Wire this prompt into your resume pipeline with a pre-injection validation gate. Before calling the model, verify that the session summary is non-empty, that the last_active_turn timestamp is within your acceptable staleness window, and that the pending_actions array does not contain malformed or duplicate entries. If validation fails, fall back to a cold-start resume that acknowledges the gap without injecting suspect state. On the model response side, implement a post-generation continuity check: compare the assistant's first response against the injected summary to confirm that critical pending actions or unresolved questions were acknowledged. A simple eval prompt can score this on a pass/fail basis—if the assistant ignores a high-priority pending action, flag the turn for human review or trigger a corrective re-prompt.

For model selection, prefer models with strong instruction-following and long-context reliability (e.g., Claude 3.5 Sonnet or GPT-4o) because the prompt requires precise adherence to the injection contract and the ability to distinguish summary context from live user input. Log every injection event with the session ID, summary version, validation result, and continuity check outcome. This audit trail is essential for debugging 'resume amnesia' bugs where the assistant appears to forget prior context. Avoid injecting summaries that exceed 30-40% of your available context budget—if the summary is too large, compress it further before injection rather than starving the model of room for the current turn and any tool outputs.