Mixed-Content Decomposition Prompt for Analysts

This prompt is designed for intelligence analysts, researchers, and verification pipeline engineers who need to decompose a complete report or briefing into a structured, machine-readable schema. Its primary job is to tag every passage by its rhetorical type—fact, assessment, forecast, or rhetorical framing—and explicitly link each passage to its stated or implied evidence basis. This is a critical pre-processing step for automated evidence matching, as feeding a model a paragraph that mixes a verifiable statistic with an analyst's forward-looking opinion will produce unreliable verification results. The ideal user is building or operating a system where downstream components require clean separation of checkable claims from interpretive content before any fact-checking logic is applied.

Use this prompt when you have a complete, multi-paragraph document and need a holistic decomposition that preserves the document's structure and evidence chain. It is appropriate for formal reports, threat assessments, market briefings, and other long-form analytical products. Do not use this prompt for real-time chat moderation, single-sentence classification, or social media content where the concept of a 'report' structure does not apply. For sentence-level classification without the evidence-linking requirement, use the Fact-Interpretation Separation System Prompt instead. This prompt also assumes the input is in a final, readable text format; it is not designed to handle raw, unstructured data dumps, highly redacted text, or content where the primary evidence is embedded in non-textual formats like complex charts without descriptions.

Before integrating this prompt into a production pipeline, you must define clear evaluation criteria for boundary precision—specifically how the model should handle sentences that contain both a fact and an interpretation. The included eval harness tests for this exact failure mode. If your workflow involves regulated or high-stakes domains, such as intelligence analysis or legal review, the output of this decomposition must be treated as a draft requiring human review before it becomes the basis for any automated action or verified claim. Proceed by copying the prompt template into your harness, substituting your domain-specific examples, and running the provided test cases to establish a baseline for acceptable decomposition accuracy in your context.

This template is designed to be copied directly into your prompt management system, IDE, or orchestration layer. It accepts a raw intelligence report and returns a structured JSON decomposition. Every placeholder is enclosed in square brackets and must be replaced with your specific inputs before use. The prompt enforces strict output typing, requires evidence basis for each passage, and explicitly marks sentences where fact and interpretation are blended—a critical detail for downstream verification accuracy.

textCopy
You are an expert intelligence analyst specializing in content decomposition. Your task is to process a report that blends facts, assessments, and forecasts, and produce a structured decomposition.

## INPUT
Report Content:
[REPORT_TEXT]

## OUTPUT SCHEMA
Return ONLY valid JSON matching this schema:
{
  "passages": [
    {
      "passage_id": "string",
      "original_text": "string (exact text span from the report)",
      "content_type": "fact | interpretation | opinion | prediction | rhetorical_framing | blended",
      "evidence_basis": "string | null (source, data, or reasoning cited; null if none)",
      "blended_sentences": [
        {
          "sentence_text": "string (the full sentence containing both fact and interpretation)",
          "factual_part": "string (the verifiable claim within the sentence)",
          "interpretive_part": "string (the analysis, judgment, or framing within the sentence)"
        }
      ],
      "confidence": "high | medium | low (your confidence in the content_type classification)"
    }
  ]
}

## CONSTRAINTS
- Preserve the original text exactly in `original_text`. Do not paraphrase.
- If a sentence contains both a verifiable fact and an interpretation, mark the entire passage as `blended` and populate the `blended_sentences` array.
- For `blended` passages, the `content_type` must be `blended`, and you must still provide an `evidence_basis` for the factual part if one is detectable.
- If no evidence basis is provided in the text, set `evidence_basis` to null. Do not invent sources.
- Classify each passage independently. Do not assume the type of one passage based on surrounding passages.
- If confidence is `low`, briefly explain why in a `classification_rationale` field added to the passage object.

## EXAMPLES
Input: "The GDP grew by 2.1% last quarter, which signals a strong recovery."
Output passage:
{
  "passage_id": "p1",
  "original_text": "The GDP grew by 2.1% last quarter, which signals a strong recovery.",
  "content_type": "blended",
  "evidence_basis": "Implied economic data; no explicit source cited in text.",
  "blended_sentences": [
    {
      "sentence_text": "The GDP grew by 2.1% last quarter, which signals a strong recovery.",
      "factual_part": "The GDP grew by 2.1% last quarter.",
      "interpretive_part": "which signals a strong recovery."
    }
  ],
  "confidence": "high"
}

Input: "In my view, the policy will fail within six months."
Output passage:
{
  "passage_id": "p2",
  "original_text": "In my view, the policy will fail within six months.",
  "content_type": "prediction",
  "evidence_basis": null,
  "blended_sentences": [],
  "confidence": "high"
}

## RISK LEVEL
[HIGH_RISK_DOMAINS: legal, medical, financial, intelligence]
If the report content falls into a high-risk domain, flag any passage where classification confidence is `medium` or `low` with a `requires_human_review: true` field.

After copying the template, replace [REPORT_TEXT] with your source content and adjust [HIGH_RISK_DOMAINS] to match your operational context. The output schema is designed to be parsed by a downstream verification pipeline, so do not relax the JSON structure. If your model struggles with strict JSON adherence, add a post-generation validation step that checks for the presence of all required fields and retries with a stronger format instruction. For high-risk domains, always route blended passages and low-confidence classifications to a human review queue before any automated fact-checking begins.

This prompt is designed as a pre-processing step in a larger verification pipeline. It should sit immediately after content ingestion and before any evidence-matching or fact-checking stages. The primary integration point is a structured API call where the raw document text is passed as the [INPUT] and the model returns a JSON array of decomposed passages. The application layer must validate the output schema before allowing downstream consumers to act on the classification results.

Validation and Retry Logic: Implement a strict JSON schema validator that checks for required fields (passage_id, original_text, classification, evidence_basis, confidence) and enum compliance on the classification field (fact, interpretation, opinion, prediction, rhetorical_framing). If validation fails, retry up to two times with an error-specific repair prompt that includes the original input, the malformed output, and the exact validation error message. After two failed retries, route the entire document to a human review queue with the partial output and error context attached. Logging and Observability: Log every decomposition request with a trace_id, the model version used, token counts, validation status, and the count of passages by classification type. This data is critical for detecting drift in classification behavior over time or across document types.

Human Review Routing: Passages where the model assigns a confidence score below 0.7, or where the classification is prediction with an evidence_basis of none, should be automatically flagged for human review. Package these into a structured handoff that includes the ambiguous passage, the model's classification and rationale, and a specific question for the reviewer (e.g., 'Does this passage contain a verifiable factual claim about a past event, or is it a forward-looking assessment?'). Model Selection: Use a model with strong instruction-following and JSON mode capabilities. For high-throughput pipelines, consider a smaller, fine-tuned model that has been trained on your domain-specific fact-interpretation boundary examples to reduce latency and cost. Avoid using this prompt with models that have weak JSON output reliability without a repair loop in place.

Downstream Integration: The validated output should be split into two streams: the fact passages are routed to the evidence-matching and claim verification subsystem, while interpretation, opinion, and prediction passages are routed to a separate analysis or summarization pipeline. Never send passages classified as rhetorical_framing to a fact-checking module, as they will produce false positives and erode trust in the verification output. Before deploying, run a golden dataset of 50 manually labeled documents through the harness and measure boundary precision, classification recall, and the false positive rate for facts misclassified as interpretation. Gate deployment on these metrics meeting your agreed thresholds.