Context-Stripping Flagging Prompt

This prompt is designed for verification analysts, editorial QA teams, and AI pipeline builders who need to detect when a quoted passage has been stripped of its surrounding context in a way that materially changes its meaning. The job-to-be-done is not simple quote matching; it is a semantic integrity check. The prompt compares a presented quote against its full source document to identify missing qualifying statements, contradictory adjacent sentences, or omitted caveats that would lead a reasonable reader to a different conclusion. The ideal user is someone integrating this into a content verification workflow—either as a manual review aid or as an automated step in a CMS or fact-checking pipeline—where the cost of a missed context-stripping incident is high, such as in news publishing, legal document review, or executive communications.

Use this prompt when you have both the extracted quote and the complete, unabridged source text available. It is most effective on non-fiction, argumentative, or declarative source material where authorial intent is carried across multiple sentences. The prompt requires few-shot examples to calibrate the distinction between benign truncation (e.g., removing a redundant introductory clause) and malicious context removal (e.g., omitting a subsequent 'however' clause that reverses the claim). Do not use this prompt for verifying quotes against audio transcripts where disfluencies are the primary concern; that is a separate quote-to-transcript alignment task. Do not use it when the source text is unavailable or when you are only checking if a quote is verbatim—this prompt's value is in meaning distortion detection, not string matching.

Before wiring this into an automated pipeline, you must define a distortion severity threshold for routing. Low-severity findings (e.g., minor omissions that do not change the core claim) can be logged for audit; high-severity findings (e.g., removal of a direct negation) should block publication and route to a human reviewer. The prompt's output includes a structured severity score, the stripped context, and the missing material, which serves as the evidence packet for that review. If you are operating in a regulated domain such as legal or financial communications, always require human sign-off on any context-stripping flag before taking action, and retain the prompt's output as part of the audit trail.

This prompt template is designed for verification analysts and AI pipeline builders who need to flag context-stripping in attributed quotes. It compares a quoted passage against its original source material, identifies what surrounding context was removed, and assesses whether the omission distorts the intended meaning. The output includes the stripped context, the missing material, and a distortion severity score. Use this template as the core instruction block in a verification workflow, adapting the placeholders to your specific source documents, quote formats, and output requirements.

textCopy
You are a quote-verification analyst specializing in context-stripping detection. Your task is to compare a quoted passage against its original source material and determine whether surrounding context has been removed in a way that changes the quoted statement's meaning.

## INPUT

**Quoted Passage:**
[QUOTED_TEXT]

**Original Source Excerpt:**
[SOURCE_EXCERPT]

**Attribution Context (optional):**
[ATTRIBUTION_CONTEXT]

## INSTRUCTIONS

1. Locate the quoted passage within the original source excerpt. If the quote cannot be found, report a provenance failure.
2. Identify what text immediately precedes and follows the quoted passage in the original source.
3. Determine whether the omitted surrounding context materially changes the meaning, tone, or implications of the quoted passage.
4. Classify the omission using one of these categories:
   - **Benign Truncation:** Context removed for brevity without meaning distortion.
   - **Qualification Removal:** Hedging, conditions, or limitations removed.
   - **Contradiction Omission:** Directly contradictory statements omitted.
   - **Tone Shift:** Emotional or rhetorical framing altered by omission.
   - **Speaker Position Reversal:** The speaker's actual stance is opposite to what the isolated quote suggests.
5. Assign a distortion severity score from 1 to 5, where:
   - 1 = No meaningful distortion (routine excerpting)
   - 2 = Minor nuance loss, overall meaning preserved
   - 3 = Moderate distortion, some implications changed
   - 4 = Significant distortion, meaning substantially altered
   - 5 = Severe distortion, quote now conveys the opposite of the original intent

## OUTPUT SCHEMA

Return a JSON object with this exact structure:
{
  "quote_found": boolean,
  "quote_location": {
    "start_index": number or null,
    "end_index": number or null
  },
  "omitted_preceding_context": string or null,
  "omitted_following_context": string or null,
  "omission_category": "Benign Truncation" | "Qualification Removal" | "Contradiction Omission" | "Tone Shift" | "Speaker Position Reversal" | "None",
  "distortion_severity_score": number (1-5),
  "distortion_explanation": string,
  "original_meaning_summary": string,
  "isolated_quote_impression": string,
  "recommendation": "Publish as-is" | "Restore context" | "Remove quote" | "Escalate for review"
}

## CONSTRAINTS

- Base all analysis strictly on the provided source excerpt. Do not infer missing context.
- If the quote is a paraphrase rather than a direct quotation, note this in the distortion explanation.
- If multiple instances of the quote appear in the source, analyze the most relevant match and flag ambiguity.
- Do not evaluate whether the quote itself is factually accurate. Focus only on context-stripping distortion.
- If the source excerpt is insufficient to make a determination, set distortion_severity_score to null and explain why.

## EXAMPLES

**Example 1: Benign Truncation**
Quoted: "The product is revolutionary."
Source: "While early results are promising, the product is revolutionary. We expect strong adoption."
Output: omission_category: "Benign Truncation", distortion_severity_score: 1

**Example 2: Qualification Removal**
Quoted: "We expect revenue to double."
Source: "Under optimistic scenarios, we expect revenue to double, though our base case shows 15% growth."
Output: omission_category: "Qualification Removal", distortion_severity_score: 4

**Example 3: Speaker Position Reversal**
Quoted: "This approach could work."
Source: "Some argue this approach could work, but I believe it's fundamentally flawed and dangerous."
Output: omission_category: "Speaker Position Reversal", distortion_severity_score: 5

To adapt this template, replace the square-bracket placeholders with your actual data. The [QUOTED_TEXT] should contain the exact quote as it appears in the content under review. The [SOURCE_EXCERPT] must include sufficient surrounding material—typically 2-3 paragraphs before and after the quote's location—to enable meaningful context comparison. The optional [ATTRIBUTION_CONTEXT] field can include speaker identity, date, venue, or other metadata that helps interpret the quote. For production use, wire the JSON output schema into a downstream validator that checks field types, required fields, and score ranges before the result enters your verification pipeline. If the distortion severity score is 3 or higher, route the result to human review with the full analysis attached.

The Context-Stripping Flagging Prompt is designed to operate as a single step within a larger quote-verification pipeline. It expects a matched pair of inputs: the quoted text as it appears in the content under review, and the surrounding original source passage from which the quote was extracted. The model's job is to compare these two inputs and flag whether the removal of surrounding context has distorted the quoted statement's meaning. This prompt is not a standalone fact-checker; it does not verify the accuracy of the quote itself or assess the credibility of the source. It answers one narrow question: has context-stripping changed what a reasonable reader would understand the speaker to mean?

To integrate this prompt into an application, wrap it in a harness that enforces a strict JSON output contract. The expected output schema includes a distortion_severity score (e.g., none, low, medium, high), a missing_context field containing the specific omitted material, and a rationale explaining the distortion mechanism. Your harness should validate this schema immediately after the model responds. If the output fails schema validation, retry once with an explicit error message appended to the prompt: 'Your previous output did not match the required JSON schema. Please correct the format and ensure all required fields are present.' If the second attempt also fails, log the failure and route the item to a human review queue with the original inputs and both raw model responses attached. Do not silently accept malformed outputs in a production verification pipeline.

Model choice matters for this task. The prompt requires careful comparative reasoning over two text passages, which smaller or older models may handle inconsistently. Use a capable instruction-following model such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Set temperature to 0 or a very low value (0.1) to maximize deterministic scoring. For high-throughput batch processing, implement a queue-based architecture that processes quote-source pairs in parallel with configurable concurrency limits and rate-limit backoff. Log every input pair, output, schema validation result, and retry attempt for auditability. If your use case involves regulated content (legal testimony, financial disclosures, medical statements), always route medium and high severity outputs to human review regardless of schema validity, and never auto-publish a 'no distortion' finding without a reviewer spot-check on a sample of low-severity outputs.

Evaluation should be built into the harness, not treated as an afterthought. Maintain a golden dataset of quote-source pairs with known distortion labels, including edge cases such as benign truncation for brevity, ellipsis usage that is defensible, and malicious context removal that flips meaning. Run this evaluation set against any prompt or model change before deployment. Track false-positive rates (flagging legitimate excerpting as distortion) and false-negative rates (missing genuine context-stripping). If your false-negative rate exceeds your risk tolerance, lower the threshold for human review routing. If false positives are overwhelming reviewers, adjust the prompt's few-shot examples to better distinguish routine editorial trimming from meaning-altering context removal. The prompt includes few-shot examples precisely for this calibration purpose—update them based on your observed failure patterns, not generic assumptions.