Direct Quotation and Paraphrased Claim Extraction Prompt

This prompt is designed for editorial verification teams and AI pipeline builders who need to extract factual claims from source content and rigorously distinguish between verbatim direct quotations and paraphrased assertions. The core job-to-be-done is to produce a structured inventory of claims where each item is tagged with its quotation type (direct_quote or paraphrase), the exact original text, and the attributed speaker or source. This is not a general-purpose claim extraction prompt; it is specifically for workflows where misattributing a paraphrase as a direct quote, or vice versa, carries high editorial, legal, or reputational risk.

Use this prompt when you are processing interview transcripts, press releases, reported speech in articles, legal testimony, or any content where the precise wording and attribution of a statement must be preserved for downstream verification. The ideal user is a verification engineer integrating this into a fact-checking pipeline, or an editorial operations lead setting up a human-in-the-loop review queue. Do not use this prompt if you only need atomic claim decomposition without quotation fidelity, or if your source material does not contain attributed speech. In those cases, the Atomic Claim Extraction Prompt Template is a better fit. You must also avoid this prompt when the source text is heavily redacted or anonymized, as the attribution metadata will be unreliable.

Before wiring this into a production harness, you must define clear validation rules. The primary failure mode is misattribution: a paraphrased claim being incorrectly tagged as a direct quote and linked to a speaker who never said those exact words. A secondary failure mode is context-stripping, where a direct quote is extracted but the surrounding hedging or conditional language is dropped, changing its meaning. To mitigate this, your implementation should log the quotation_type and attributed_to fields for every extracted claim and route any claim where the extraction confidence is below a defined threshold for human review. The next section provides the exact prompt template you will need to implement this extraction logic.

This prompt template separates verbatim quoted claims from paraphrased assertions in source content, producing structured output with quotation type tags and attribution metadata. Use it when downstream verification workflows need to distinguish exact wording from interpretive restatements—a critical distinction for quote accuracy checks, editorial review, and source fidelity scoring. The template expects source content and optional context about speakers and publication metadata. Do not use this prompt for general claim extraction without quotation tracking; use the Atomic Claim Extraction Prompt Template instead.

codeCopy
You are an editorial verification assistant. Your task is to extract claims from the provided source content and classify each as a direct quotation or a paraphrased assertion. For each claim, preserve the original wording, identify the speaker or source, and attach quotation type metadata.

## INPUT
Source Content: [SOURCE_CONTENT]
Known Speakers and Roles: [SPEAKER_MAP]
Publication Context: [PUBLICATION_CONTEXT]

## OUTPUT SCHEMA
Return a JSON object with a "claims" array. Each claim object must include:
- "claim_id": string, unique identifier for this claim
- "claim_text": string, the exact claim text as it appears or is implied in the source
- "quotation_type": enum, one of ["direct_quote", "paraphrase", "mixed"]
- "original_wording": string or null, the verbatim source text for direct quotes; null for pure paraphrases
- "speaker": string or null, the attributed speaker or source entity
- "speaker_role": string or null, the speaker's role or affiliation if known
- "source_location": string, reference to where the claim appears (paragraph number, line, or section)
- "context_window": string, surrounding sentences that provide context for the claim
- "attribution_confidence": enum, one of ["explicit", "implied", "ambiguous"]
- "attribution_evidence": string, the text that establishes the speaker attribution

## CONSTRAINTS
- Do not merge claims from different speakers into a single claim object.
- When a direct quote is embedded within a paraphrase, classify as "mixed" and include both the quoted portion and the paraphrased portion.
- If speaker attribution is ambiguous or missing, set "attribution_confidence" to "ambiguous" and note the ambiguity in "attribution_evidence".
- Do not invent speaker attributions. If the source does not identify the speaker, set "speaker" to null.
- Preserve the original wording exactly for direct quotes, including any grammatical errors or stylistic choices.
- For paraphrased claims, the "claim_text" should capture the asserted fact without adding interpretation.
- Return only the JSON object. No additional commentary.

## EXAMPLES
[EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]

Adaptation guidance: Replace [SOURCE_CONTENT] with the article, transcript, or document text. Populate [SPEAKER_MAP] with known speaker names and roles to improve attribution accuracy—this is especially important for interview transcripts and multi-speaker content. Use [PUBLICATION_CONTEXT] to provide date, publication name, and document type, which helps the model resolve ambiguous references. The [EXAMPLES] placeholder should contain 2-3 few-shot examples showing correct extraction for direct quotes, paraphrases, and mixed cases. Set [RISK_LEVEL] to "high" when the output feeds into publication or compliance workflows; this signals downstream harnesses to enable human review gates. For production use, add a [CONSTRAINTS] section specifying maximum claims per document and any domain-specific attribution rules. Validate outputs against the schema before ingestion, and route any claim with attribution_confidence: "ambiguous" to human review.

This prompt is designed to sit at the front of a claim extraction pipeline where source content contains both verbatim quotes and paraphrased assertions. The implementation harness must handle three distinct concerns: input preparation (chunking source text with speaker metadata intact), output validation (ensuring quotation type tags and attribution fields are present and consistent), and downstream routing (sending verbatim quotes to quote-accuracy verification while paraphrased claims go to fact-checking). The prompt itself is stateless, so the harness owns all context management—tracking which speaker said what, preserving original text spans for later comparison, and maintaining document-level provenance across multiple extraction calls.

Wire the prompt into your application with a pre-processing layer that segments source content into speaker-attributed chunks. Each chunk should carry a speaker ID, speaker name, and the raw text span. If your source is a transcript, use diarization metadata. If it's a written article, parse attribution cues like 'according to' or 'said.' The prompt template's [SOURCE_TEXT] placeholder receives the chunk, while [SPEAKER_CONTEXT] receives a structured object with available speaker names and IDs. After the model responds, run a validation layer that checks: (1) every extracted claim has a quotation_type field set to either verbatim or paraphrased, (2) verbatim claims include an original_quote field that matches a substring of the source text (use fuzzy string matching with a configurable threshold, typically 0.85 similarity), (3) paraphrased claims do not contain verbatim text spans longer than 5 consecutive words from the source, and (4) the attributed_to field maps to a valid speaker ID from the input context. Claims failing validation should trigger a retry loop—resubmit the same chunk with the validation errors appended as [CORRECTION_NOTES] in a follow-up call, up to 2 retries before routing to human review.

For model choice, prefer models with strong instruction-following and structured output capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent). Set temperature to 0 or near-zero to minimize variation in quotation boundary detection. If your pipeline processes high-volume content, consider batching multiple speaker-attributed chunks into a single prompt call with clear delimiters, but cap batch size at 5-8 chunks to avoid attribution confusion. Logging should capture: the input chunk, speaker context, raw model output, validation results, retry count, and final accepted or rejected status. Store the original source text span alongside each extracted claim for downstream quote-accuracy verification. For high-risk editorial workflows (e.g., publishing corrections, legal review of attributed statements), always route claims where quotation_type is verbatim but the fuzzy match score falls below 0.95 to a human reviewer before the claim enters any published verification report. Avoid wiring this prompt directly into a publish action without a review gate when attribution errors could cause public misquotation harm.