Spurious Claim Removal Prompt for Executive Summaries

This prompt is designed for post-generation validation in executive briefing and decision-support systems. It takes an AI-generated executive summary and the source documents used to produce it, then identifies and strips every factual assertion that cannot be traced back to the input material. The output is a clean summary containing only grounded claims, plus an audit trail of what was removed and why. Use this when your product surfaces executive summaries to decision-makers who act on the information, and when a single fabricated statistic or unsupported projection can damage trust or cause a bad decision.

This is a repair prompt, not a generation prompt. It assumes you already have a draft summary and the original source documents. It does not rewrite for style or fluency; it only removes ungrounded content. The ideal user is a product engineer or AI architect building a system where a draft summary is generated first (by another model or pipeline), and this prompt acts as a safety net before the summary reaches a human decision-maker. Required inputs are the draft summary text and the complete set of source documents that were used to produce it. Without both, the prompt cannot perform reliable grounding checks.

Do not use this prompt when you need to generate a summary from scratch, when you need to improve writing quality or narrative flow, or when the source documents are unavailable or incomplete. It is also not a substitute for human review in high-stakes domains—the audit trail it produces should inform a human reviewer's decision, not replace it. For regulated industries, always pair this prompt's output with a human-in-the-loop approval step before the cleaned summary is acted upon.

The prompt below is designed to be pasted into your system prompt or user message. It takes a draft executive summary and a set of source documents, then returns a cleaned summary with an audit trail of every removed claim. Replace the square-bracket placeholders with your actual inputs before running. The prompt enforces strict traceability: any assertion that cannot be mapped to a source passage is flagged and stripped, not softened or reworded.

textCopy
You are an executive briefing editor. Your task is to remove every unsupported assertion from the draft executive summary below.

## INPUTS
[DRAFT_SUMMARY]

[SOURCE_DOCUMENTS]

## CONSTRAINTS
- Preserve only claims that can be directly traced to at least one passage in [SOURCE_DOCUMENTS].
- Do not reword, soften, or hedge unsupported claims. Remove them entirely.
- Do not add new facts, interpretations, or connective tissue.
- Maintain the original summary structure where possible, leaving placeholders for removed content.
- If a claim is partially supported, keep only the supported portion and discard the rest.

## OUTPUT_SCHEMA
Return a JSON object with the following fields:
{
  "cleaned_summary": "string (the executive summary with all unsupported claims removed)",
  "removals": [
    {
      "removed_text": "string (the exact text that was removed)",
      "reason": "unsupported_claim | partial_match_trimmed | contradicted_by_source",
      "source_evidence": "string or null (the closest matching passage if partial_match_trimmed, otherwise null)"
    }
  ],
  "retained_claim_count": "integer",
  "removed_claim_count": "integer"
}

## EXAMPLES
[EXAMPLES]

## RISK_LEVEL
[HIGH] — Executive summaries drive decisions. Fabricated claims can cause financial, legal, or reputational harm. Every removal must be auditable.

To adapt this prompt, replace [DRAFT_SUMMARY] with the full text of the summary under review and [SOURCE_DOCUMENTS] with the complete set of source material, clearly delimited by document title or ID. The [EXAMPLES] placeholder should be populated with 2–3 few-shot examples showing a draft summary, the source passages that support or fail to support claims, and the resulting cleaned output. For high-stakes domains such as legal, financial, or clinical briefings, route every removal record through a human review queue before the cleaned summary is delivered to decision-makers. Pair this prompt with a validation step that checks the removed_claim_count against a threshold; if removals exceed a configured limit, escalate the entire summary for manual review rather than delivering a heavily redacted output.

This prompt is designed to sit in a post-generation validation pipeline, not as a standalone chat interaction. After your primary model generates an executive summary from source documents, route the summary and the original source documents through this prompt before the summary reaches any downstream system, database, or user-facing surface. The prompt expects two structured inputs: the generated summary text and the full source document text. If your system synthesizes across multiple documents, concatenate them with clear document boundary markers (e.g., [DOCUMENT: filename] ... [/DOCUMENT]) so the prompt can trace claims to specific sources.

Input validation should happen before the prompt is called. Verify that both the summary and source documents are non-empty strings. If the source documents exceed the model's context window, chunk them and run the prompt once per chunk, then deduplicate the removal lists. Output validation must parse the JSON response and confirm it contains the required fields: cleaned_summary, removed_claims (an array of objects with claim, reason, and severity), and audit_summary. If the model returns malformed JSON, use a retry loop with the error message injected into the next attempt. After three failed retries, escalate to a human review queue rather than silently passing the uncleaned summary downstream.

Model choice matters here. Use a model with strong instruction-following and JSON output capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent). Avoid smaller or older models that may struggle with the dual task of identifying unsupported claims while preserving the summary's narrative flow. Logging and audit trails are critical: store the original summary, the cleaned summary, the full removal list, and the model version for every invocation. This creates an evidence trail for compliance reviews and lets you measure hallucination rates over time. Human review should be gated on severity: if any removed claim has a severity of high or critical, route the output to a human reviewer before publication. For lower-severity removals, you can auto-approve but still log for periodic quality sampling.

Integration pattern: wrap this prompt in a function that accepts (summary: str, sources: str) -> CleanedSummaryResult. Call it after your primary summarization step and before any caching, distribution, or user notification. If your system uses streaming, buffer the full summary before running this check—partial summaries will produce unreliable results. Cost optimization: cache the source documents if they're reused across multiple summaries, and consider running this prompt only on summaries flagged by a cheaper, faster pre-filter (e.g., a regex-based entity mismatch check or a lightweight confidence classifier). What to avoid: do not use this prompt as your primary summarization step. It is a repair and validation tool, not a generation tool. Running it on every output without caching or pre-filtering will double your inference costs. Also, never silently discard the removal list—the audit trail is the primary value for trust and compliance.