Invented Quote Removal Prompt for Meeting Summaries

This prompt is designed for a single, high-stakes job: detecting and removing fabricated direct or indirect quotes from AI-generated meeting summaries before those summaries reach end users, databases, or downstream systems. The ideal user is a product team building meeting intelligence, collaboration, or conversation analytics software where transcripts are the source of truth. The prompt expects a complete meeting transcript and a model-generated summary as inputs. It is not designed for general factuality checks, RAG citation audits, or hallucination detection in free-form prose without a transcript anchor. Use it when your product must guarantee that every quoted statement attributed to a participant can be traced back to the transcript, and when presenting invented speech as real carries reputational, legal, or compliance risk.

The prompt works by instructing the model to perform a structured comparison: extract every quoted or indirectly attributed statement from the summary, search for that statement in the transcript, and classify each as grounded, fabricated, or paraphrased. A paraphrased quote is one where the meaning is preserved but the exact wording differs—this is legitimate and should be retained. A fabricated quote is one where neither the wording nor the semantic content appears in the transcript. The prompt requires the model to output a clean version of the summary with fabricated quotes removed and a separate audit log documenting every removal decision. This separation of cleaned output and audit trail is critical for downstream logging, quality monitoring, and human review workflows.

Do not use this prompt when the transcript is unavailable, incomplete, or poorly formatted. The prompt's accuracy depends entirely on the quality of the transcript input. If speaker diarization is missing or unreliable, the model may struggle to attribute quotes correctly. Do not use this prompt as a substitute for a full factuality pipeline—it only addresses quote fabrication, not invented statistics, hallucinated action items, or spurious entity references. For those concerns, pair this prompt with sibling playbooks like Factual Claim Extraction and Hallucination Flagging or Entity Hallucination Detection for Summaries. In regulated or high-risk domains such as legal depositions, clinical handoffs, or compliance hearings, always route the audit log to a human reviewer before the cleaned summary is published or stored. The prompt is a filter, not a guarantee—treat it as one layer in a defense-in-depth strategy against hallucinated content.

This prompt template is designed to be integrated into a post-generation validation step for meeting intelligence applications. It takes a meeting transcript and a generated summary as inputs and instructs the model to identify any direct or indirect quotes in the summary that cannot be verified against the transcript. The output is a cleaned summary with fabricated quotes removed and a structured audit log of changes.

textCopy
You are an output validator for a meeting intelligence system. Your task is to compare a generated meeting summary against the original transcript and remove any fabricated quotes.

[INPUT]

Transcript:
[TRANSCRIPT]

Generated Summary:
[SUMMARY]

[CONSTRAINTS]

1. Identify every direct quote (text inside quotation marks) and indirect quote (attributed speech like "Alice said that...") in the summary.
2. For each quote, search the transcript for an exact or near-exact match. A near-exact match preserves the core meaning and most keywords, allowing for minor grammatical adjustments.
3. If a quote has no match in the transcript, flag it as FABRICATED.
4. Remove all fabricated quotes from the summary. Replace them with a neutral, non-quote paraphrase of the surrounding context if necessary to maintain grammatical flow. If removal leaves a sentence broken, rewrite the sentence to be grammatically complete.
5. Do not remove or alter any content that is a legitimate paraphrase of the transcript's discussion, even if it is attributed to a speaker. Only remove unverifiable quoted speech.
6. Preserve all action items, decisions, and non-quote factual content.

[OUTPUT_SCHEMA]

Return a valid JSON object with the following structure:
{
  "cleaned_summary": "The full summary text with all fabricated quotes removed.",
  "audit_log": [
    {
      "original_quote": "The fabricated quote found in the summary.",
      "speaker_attribution": "The person the quote was attributed to, or null.",
      "reason": "FABRICATED",
      "action_taken": "REMOVED_AND_REPLACED"
    }
  ],
  "fabricated_quote_count": 0,
  "total_original_quote_count": 0
}

[RISK_LEVEL]

HIGH. Fabricated quotes in meeting summaries can create false records of what individuals said, leading to misattribution, compliance issues, and trust erosion. Always log all removals for auditability. If the summary contains a high volume of fabricated quotes (more than 20% of total quotes), flag the entire summary for human review instead of returning a cleaned version.

To adapt this template, replace the [TRANSCRIPT] and [SUMMARY] placeholders with your actual data at runtime. The [OUTPUT_SCHEMA] section defines a strict JSON contract that your application can parse programmatically. The [RISK_LEVEL] section includes a critical gating rule: if the model detects systemic fabrication, it should escalate for human review rather than silently patching a deeply flawed output. This prevents a false sense of security when the underlying generation process is broken.

Integrating this prompt into a meeting intelligence application requires a pre-processing step that extracts the transcript and the summary into the [TRANSCRIPT] and [SUMMARY] placeholders. The transcript should be speaker-attributed and timestamped if possible, as the model uses speaker identity and temporal context to verify quote authenticity. The summary must be the raw model output before any post-processing or formatting has been applied. If your pipeline generates summaries incrementally or in chunks, assemble the full summary before invoking this prompt to avoid partial-quote false positives.

The implementation harness should wrap the prompt call in a validation layer that checks the output schema before accepting the result. Expect a JSON response with a flagged_quotes array, each containing quote_text, line_number, fabrication_confidence, and reason. Validate that fabrication_confidence is a float between 0.0 and 1.0, that line_number maps to an actual line in the input summary, and that quote_text is a non-empty string. If validation fails, retry once with the same prompt and a stronger [CONSTRAINTS] block that emphasizes strict JSON output. After two failures, escalate the summary to a human review queue rather than silently passing potentially fabricated quotes to end users.

For high-stakes meeting domains—legal proceedings, board meetings, earnings calls, or clinical handoffs—add a mandatory human approval step for any quote flagged with fabrication_confidence above 0.7. Log every removal decision with the original summary text, the flagged quote, the confidence score, and the reviewer's action (confirmed removal, reinstated, or edited). This audit trail is essential for compliance and for debugging model behavior over time. Model choice matters here: use a model with strong instruction-following and JSON output reliability, such as GPT-4o or Claude 3.5 Sonnet, and set temperature=0 to minimize variance in fabrication confidence scoring. Do not use this prompt as the sole defense if fabricated quotes carry legal or regulatory risk—always pair it with human review for high-severity outputs.