Streaming Interruption Continuation Prompt Template

This prompt is for real-time application developers and streaming infrastructure engineers who need to recover from a mid-stream disconnection or timeout without losing the model's generated context. The core job-to-be-done is to request a clean continuation from the exact truncation point, preserving the original task's tone, reasoning chain, and output format. The ideal user is someone building a production system—such as a chat interface, a coding copilot, or a real-time data extraction pipeline—where a dropped connection should not force the end-user to restart the entire interaction from scratch.

Use this prompt when you have a partial, well-formed output that was interrupted by a network error, a server-side timeout, or a client-side cancellation, and you need to stitch the remaining content seamlessly. The prompt is designed to work with the raw truncated text and the original system instructions. It is not suitable for repairing malformed JSON, fixing token-limit cutoffs where the model stopped naturally, or reassembling multiple independently streamed chunks. For those cases, use the sibling playbooks for 'Truncated JSON Completion Repair' or 'Streaming Chunk Reassembly.' Do not use this prompt if the interruption point is ambiguous, if the partial output is corrupted, or if the original task context is no longer available—the model will hallucinate a continuation rather than recover the intended output.

Before implementing this prompt in your harness, ensure you have captured the exact last 50–100 characters of the streamed output to use as the [TRUNCATION_POINT] and that you can reconstruct the full [ORIGINAL_SYSTEM_PROMPT] and [ORIGINAL_USER_MESSAGE]. The primary failure mode is duplication: the model may repeat the last few tokens already sent to the user. Your application layer must implement a deduplication check by comparing the overlap between the end of the partial output and the beginning of the continuation. If the workflow involves regulated data, customer-facing financial advice, or clinical documentation, a human must review the stitched output before it is delivered. Proceed to the 'Prompt Template' section for the copy-ready instruction set.

This template is designed for real-time application developers who need to recover from a streaming interruption without losing context, tone, or output format. It instructs the model to continue generating from the exact point of truncation, using the provided partial output and original task context. The prompt is structured to prevent duplication, maintain schema adherence, and preserve the original reasoning chain.

textCopy
SYSTEM:
You are a continuation assistant. Your task is to complete a response that was interrupted mid-generation. You will receive the original task, the partial output that was generated before the interruption, and the exact last token where generation stopped. You must continue from that exact point without repeating any content, preserving the original tone, format, and structure. Do not add introductory phrases, summaries, or commentary about the continuation. Output only the completion text that should be appended directly to the partial output.

USER:
Original Task: [ORIGINAL_TASK]
Partial Output (truncated): [PARTIAL_OUTPUT]
Last Token Before Interruption: [LAST_TOKEN]
Output Format Required: [OUTPUT_FORMAT]
Context Preservation Notes: [CONTEXT_NOTES]

Continue from the exact truncation point. Output only the completion.

To adapt this template, replace the square-bracket placeholders with your actual values. [ORIGINAL_TASK] should contain the full prompt or instruction that initiated the original generation. [PARTIAL_OUTPUT] is the complete text received before the stream disconnected. [LAST_TOKEN] should be the final few characters or tokens of the partial output to help the model align its continuation precisely. [OUTPUT_FORMAT] specifies the expected structure (e.g., JSON, markdown, code block). [CONTEXT_NOTES] can include any additional state such as conversation history, tool call results, or reasoning steps that must be preserved. Before deploying, validate that the continuation stitches cleanly without duplication by checking the boundary between the partial output and the new completion. For high-stakes applications, implement a post-generation check that compares the last 50 characters of the partial output against the first 50 characters of the completion to detect overlap.

Integrating the streaming interruption continuation prompt into a real-time application requires treating the prompt as one component inside a larger recovery harness. The application must detect the interruption, capture the exact truncation point and surrounding context, invoke the continuation prompt, and then stitch the result onto the original partial output without duplication or context loss. This is not a fire-and-forget operation—the harness must validate the boundary between the original fragment and the continuation, handle cases where the model restates content, and fall back gracefully when continuation fails.

The harness should maintain a structured record of the original request parameters (model, temperature, system prompt, messages, tools) alongside the partial output. When an interruption is detected—via a streaming error, timeout, or token-limit signal—capture the last complete sentence or structural boundary as the anchor point. Pass the original system prompt, conversation history, and the partial output up to the anchor into the continuation prompt's [PARTIAL_OUTPUT] placeholder. Set [OUTPUT_FORMAT] to match the original expected schema. After receiving the continuation, the harness must strip any repeated content by comparing the last N tokens of the original fragment with the first N tokens of the continuation and removing overlaps. Validate the stitched output against the original schema or format contract before returning it downstream.

For high-reliability systems, implement a retry policy with exponential backoff when the continuation prompt fails or produces invalid output. Log every recovery attempt with the truncation position, anchor text, continuation response, and stitch result for debugging. If the continuation introduces hallucinated content beyond the original task scope, the harness should flag the output for human review rather than silently accepting it. Avoid using continuation prompts for security-sensitive or compliance-critical content without a human-in-the-loop review step, as the model may infer incorrect context from the partial output. The harness should also enforce a maximum continuation length to prevent runaway generation, and if the stitched output still fails validation after two retries, escalate to a fallback workflow rather than looping indefinitely.