Selective Quoting and Cherry-Picking Detection Prompt

This prompt is designed for editorial QA teams, fact-checkers, and content verification engineers who need to detect when a partial quotation omits material that contradicts or materially qualifies the cited passage. The core job-to-be-done is not to flag every excerpt as problematic, but to identify deceptive selectivity: cases where the omitted text would change a reasonable reader's understanding of the source's position. The ideal user is someone integrating this prompt into a content pipeline—whether a CMS pre-publication check, a batch verification workflow, or a human review queue—who needs a structured, repeatable signal rather than a subjective opinion. The prompt expects two inputs: the quoted passage as it appears in the derivative work, and the full surrounding source context (typically several paragraphs before and after the quoted segment). Without sufficient source context, the model cannot reliably distinguish routine excerpting from cherry-picking.

When this prompt is the right tool: You have a specific quote and its original source document, and you need a risk score, the omitted qualifying text, and a fairness assessment before publication. When to reach for a different tool: If you are verifying whether a quote was actually said (use the Attributed Statement Verification Prompt), checking if a paraphrase distorts meaning (use the Paraphrase Fidelity Scoring Prompt), or evaluating whether ellipsis usage is deceptive (use the Quote Omission and Ellipsis Integrity Prompt). This prompt is also not designed for real-time chat moderation or for evaluating entire articles holistically—it performs one focused comparison per call. Required context: The model needs the full quoted passage and at least 2–3 paragraphs of surrounding source material. If you only have a sentence of context, the false-positive rate on routine excerpting will be unacceptably high. Risk profile: False positives—flagging legitimate editorial excerpting as cherry-picking—are the primary failure mode. This prompt includes calibration instructions to reduce over-flagging, but human review remains essential for any high-stakes publication decision.

What to do next: Copy the prompt template from the following section, populate the [QUOTED_PASSAGE] and [FULL_SOURCE_CONTEXT] placeholders with your material, and run it against a model that supports structured JSON output (GPT-4o, Claude 3.5 Sonnet, or equivalent). Start with a calibration set of 10–20 known examples—half legitimate excerpts, half genuine cherry-picking—to tune your risk-score threshold for routing to human review. If you are building this into a production pipeline, wire the output into a triage step: scores below your threshold auto-pass, scores above trigger human review with the omitted text and fairness assessment packaged for the reviewer. What to avoid: Do not use this prompt to evaluate quotes without access to the full source context. Do not treat the risk score as a binary publish/block decision without human calibration. And do not confuse cherry-picking detection with quote fabrication detection—this prompt assumes the quoted words were actually said, and evaluates whether the selection is fair.

This prompt template is designed to compare an attributed quote against its full original source passage and flag instances where material has been omitted in a way that contradicts, qualifies, or materially alters the meaning of the cited text. Use it as the core instruction block in a verification harness that supplies the quote, the source passage, and any relevant context. The template uses square-bracket placeholders that your application must populate before sending the request to the model.

textCopy
You are a quote-integrity analyst. Your task is to compare an attributed quote against its original source passage and detect selective quoting or cherry-picking.

## INPUT
- Attributed Quote: [QUOTE_TEXT]
- Source Passage: [SOURCE_PASSAGE]
- Additional Context (optional): [CONTEXT]

## INSTRUCTIONS
1. Identify every segment of the source passage that was omitted from the quote.
2. For each omission, determine whether the omitted material contradicts, qualifies, or materially alters the meaning of the quoted portion.
3. Classify each omission as one of: "benign_truncation", "qualifying_omission", "contradicting_omission", or "context_stripping".
4. Assess whether the quote, as presented, creates a misleading impression of the source's position.
5. Assign a cherry-picking risk score from 0 (no risk) to 100 (severe distortion).

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "cherry_picking_risk_score": number,
  "fairness_assessment": "fair" | "minor_concerns" | "misleading" | "severe_distortion",
  "omissions": [
    {
      "omitted_text": "string",
      "omission_type": "benign_truncation" | "qualifying_omission" | "contradicting_omission" | "context_stripping",
      "impact_description": "string",
      "would_omission_change_reader_understanding": boolean
    }
  ],
  "overall_assessment": "string",
  "recommendation": "publish_as_is" | "add_qualifying_context" | "revise_quote" | "escalate_for_review"
}

## CONSTRAINTS
[CONSTRAINTS]

## EXAMPLES
[EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [QUOTE_TEXT] with the attributed quote under review and [SOURCE_PASSAGE] with the full original passage from which it was drawn. Use [CONTEXT] to supply surrounding paragraphs or document metadata that helps assess whether the omission is defensible. Populate [CONSTRAINTS] with domain-specific rules such as style-guide allowances for routine excerpting, maximum acceptable omission length, or publication standards. Provide [EXAMPLES] as few-shot demonstrations showing both benign truncation and deceptive cherry-picking with correct output labels. Set [RISK_LEVEL] to guide the model's sensitivity—use "high" for legal or regulatory content where even minor omissions require escalation, or "standard" for editorial workflows that accept routine excerpting. After the model returns the JSON output, validate schema compliance before routing results to downstream systems or human reviewers.

This prompt is designed to sit inside an editorial pipeline, not as a standalone chat interaction. The typical integration point is after a writer or editor has drafted content containing attributed quotes and before that content is published. The application layer should extract each quote–source pair from the draft, call this prompt for each pair, and aggregate the results into a review queue. Because cherry-picking detection requires comparing the quoted excerpt against the full original passage, the harness must retrieve or receive the complete source context—not just the cited sentence. If the source material is behind a paywall, in a scanned PDF, or in a non-text format, you will need a document intelligence preprocessing step before this prompt can operate.

The implementation should enforce a strict input contract: each call must include the [QUOTED_TEXT] exactly as it appears in the draft, the [FULL_SOURCE_PASSAGE] containing at least the paragraph before and after the quoted section, and a [CITATION_LINE] with the document title, author, date, and section identifier. The output schema should be validated at the application layer before results are surfaced to reviewers. Key fields to validate include the cherry_picking_risk_score (must be a float between 0.0 and 1.0), the omitted_qualifying_text array (each entry must include the omitted text and its position relative to the quote), and the fairness_assessment (must be one of the predefined enum values: fair, borderline, or deceptive). If the model returns malformed JSON or missing required fields, implement a single retry with the validation error message injected into the prompt context. After two failures, route the item to a human review queue with the raw model output attached.

For model choice, use a model with strong instruction-following and structured output capabilities. This task benefits from models that can reason about textual omission and pragmatic implication, not just surface-level string matching. Set the temperature low (0.0–0.2) to reduce variance in risk scoring. If your pipeline processes high volumes, consider batching quote–source pairs and using a model router to send borderline cases (risk scores between 0.4 and 0.7) to a more capable model for a second pass. Log every call with the prompt version, model identifier, input hash, output, and reviewer disposition. This audit trail is essential for calibrating the risk score thresholds over time and defending editorial decisions if a cherry-picking accusation arises. The next step after implementation is to run a calibration set of known cherry-picked and fair quotes through the harness, compare model scores against human expert judgments, and adjust your routing thresholds before enabling auto-approval for low-risk items.