News Article Fact-Opinion Separation Prompt

This prompt is designed for editorial verification teams, fact-checking pipeline engineers, and newsroom tool builders who need to decompose a news article into two distinct streams: verifiable factual claims and non-factual content (opinion, analysis, prediction, editorializing). The core job-to-be-done is preparing journalistic content for downstream evidence matching or claim verification steps. It assumes you have the full article text and want a structured, citation-ready output that maps each passage to its classification with source spans, enabling traceable verification workflows.

Use this prompt as a pre-processing stage before evidence retrieval and claim scoring. The ideal user is integrating it into an automated verification pipeline where the output JSON is consumed by subsequent tools. The prompt requires a complete article text as input; it is not designed for real-time breaking-news triage without human review, nor for content that is primarily multimedia (video, audio, image) without a clean text transcript. It also should not be used as a standalone 'truth checker'—it separates facts from interpretation but does not verify the facts themselves.

Before implementing, ensure your pipeline can handle the structured output schema, including passage spans and classification labels. If your use case involves user-generated content with heavy slang, satire, or intentional disinformation, you will need additional adversarial testing beyond this prompt's baseline. The next step after reading this section is to review the prompt template and adapt its placeholders to your specific article format and output requirements.

This prompt template is designed for editorial verification teams processing journalistic content. It instructs the model to decompose a news article into two parallel streams: verifiable factual claims and non-factual content (opinion, analysis, prediction, rhetorical framing). The output includes source spans, attribution readiness flags, and confidence markers. Replace the square-bracket placeholders with your specific inputs, output schema, and domain constraints before use.

codeCopy
You are an editorial verification assistant. Your task is to separate verifiable factual statements from opinion, analysis, prediction, and rhetorical framing in the provided news article.

## INPUT
[ARTICLE_TEXT]

## OUTPUT_SCHEMA
Return a JSON object with the following structure:
{
  "article_id": "[ARTICLE_ID]",
  "separation": [
    {
      "passage": "exact text span from the article",
      "start_char": integer,
      "end_char": integer,
      "classification": "fact" | "opinion" | "analysis" | "prediction" | "rhetorical_framing" | "mixed",
      "confidence": 0.0-1.0,
      "attribution_ready": true | false,
      "attribution_note": "explain what source type would verify this, or why attribution is not applicable",
      "rationale": "brief explanation of the classification decision"
    }
  ],
  "summary": {
    "total_passages": integer,
    "fact_count": integer,
    "opinion_count": integer,
    "analysis_count": integer,
    "prediction_count": integer,
    "rhetorical_framing_count": integer,
    "mixed_count": integer,
    "attribution_ready_count": integer
  }
}

## CLASSIFICATION DEFINITIONS
- **fact**: A statement that can be verified against external evidence (e.g., events, data, quotes from identified sources, documented actions).
- **opinion**: A subjective judgment, personal belief, or value statement that cannot be proven true or false.
- **analysis**: Interpretation of facts, drawing connections, or explaining implications. Goes beyond reporting what happened.
- **prediction**: A forward-looking statement about future events, trends, or outcomes.
- **rhetorical_framing**: Language that shapes reader perception through metaphor, loaded terms, narrative structure, or emotional appeal without making a directly checkable claim.
- **mixed**: A passage containing both factual and non-factual elements that cannot be cleanly separated at the sentence level.

## CONSTRAINTS
[CONSTRAINTS]

## EXAMPLES
[EXAMPLES]

## RISK_LEVEL
[RISK_LEVEL]

## INSTRUCTIONS
1. Split the article into the smallest meaningful passages that can be independently classified (typically sentence-level, but may be clause-level for complex sentences).
2. For each passage, assign exactly one classification.
3. Use "mixed" only when a single passage contains inseparable factual and non-factual elements.
4. Set attribution_ready to true for facts that can be checked against a specific type of external source (public records, databases, direct witness accounts, official documents).
5. Set attribution_ready to false for opinions, rhetorical framing, and predictions.
6. For analysis passages, set attribution_ready to true only if the analysis contains embedded factual claims that can be independently verified.
7. Provide a clear rationale for every classification decision.
8. Flag passages where opinion is disguised as neutral reporting by noting the framing technique in the rationale.
9. If the article contains unattributed assertions presented as fact, classify them as fact but set attribution_ready to false and note the missing attribution in the rationale.
10. Maintain exact text spans with accurate character offsets.

Adaptation guidance: The [CONSTRAINTS] placeholder should be replaced with domain-specific rules, such as handling requirements for financial forward-looking statements, medical uncertainty language, or legal evidentiary standards. The [EXAMPLES] placeholder should contain 2-4 annotated passages showing correct classification boundaries, including edge cases where fact and opinion appear in the same sentence. The [RISK_LEVEL] placeholder should specify whether this is a pre-publication review (higher precision required) or post-publication analysis (higher recall acceptable). For production use, always pair this prompt with a validation step that checks JSON schema compliance, character offset accuracy, and classification consistency across the full article.

Next steps: After generating the separation output, run a validation pass to verify that all character offsets map correctly to the source text and that no passages were skipped or duplicated. For high-risk editorial workflows, route all "mixed" classifications and low-confidence decisions (confidence < 0.7) to human review before the output is used in downstream verification pipelines. Avoid using this prompt on content that has already been through a fact-checking process, as the model may conflate verified facts with the original article's presentation.

This prompt is designed as a pre-processing step in a larger fact-checking pipeline. It should sit between content ingestion and claim verification. The model receives the full article text and returns a structured JSON object with two arrays: factual_claims (verifiable statements with source spans) and opinion_analysis (interpretation, editorializing, predictions, and rhetorical framing). Each entry must include the original text span, start and end character offsets, and a confidence score. The output schema is strict, so your application must validate it before passing claims downstream to evidence matching or human review queues.

Wire the prompt into your application with a validation layer that checks: (1) the JSON parses without error, (2) every entry in both arrays contains the required fields (span, start_char, end_char, confidence, type), (3) character offsets align with the original article text when extracted, and (4) no text span appears in both arrays unless it's a genuinely mixed sentence flagged with low confidence. If validation fails, retry once with the same prompt and a stronger constraint instruction appended: 'Your previous output failed schema validation. Ensure every entry has all required fields and character offsets match the source text exactly.' If the second attempt fails, route the article to a human review queue with the raw model output and validation error log attached. Use a model with strong JSON-mode support (GPT-4o, Claude 3.5 Sonnet, or equivalent) and set response_format to json_object with the schema provided in the prompt template. For high-throughput pipelines, batch articles in groups of 5-10 and process asynchronously, logging each article's separation latency and validation pass/fail status for observability.

Human review routing is critical for ambiguous cases. Any entry with confidence below 0.7 should be flagged for review. Additionally, if more than 20% of an article's sentences are classified with confidence below 0.7, escalate the entire article. Build a review interface that shows the original article with classified spans highlighted in different colors (green for factual, blue for opinion/analysis) and allows reviewers to reclassify spans with a single click. Log all human corrections to build a fine-tuning dataset for future model improvement. Do not send unvalidated model output directly to downstream verification systems—a single malformed claim with incorrect offsets can corrupt evidence matching and produce false verification results. Do version your prompt and track which prompt version produced each separation result so you can measure regression when iterating on the template.