Verifiable Statement Isolation Prompt Template

This prompt is a surgical instrument for verification engineers and analysts who need to extract only the checkable, atomic factual claims from a body of text. Its sole job is to strip away all interpretation, opinion, prediction, and rhetorical framing, producing a filtered list of discrete claims, each anchored to its source text span. Use it as a pre-processing step in a fact-checking pipeline, immediately before evidence retrieval and matching. The ideal user is someone building an automated verification system who understands that downstream accuracy depends entirely on clean claim boundaries. If you feed a fact-checking model a sentence that blends a fact with an opinion, you will get unreliable verification results. This prompt prevents that contamination at the source.

You should deploy this prompt when you have unstructured content—news articles, reports, transcripts, or model outputs—and you need to isolate statements that can be objectively checked against an external source. For example, from the sentence 'The company's revenue grew 15% last quarter, which was a disappointing result,' this prompt extracts only 'The company's revenue grew 15% last quarter' and discards the interpretive framing. It handles embedded factual assertions within complex sentences, numerical claims, and attributed statements. The output is a structured list of claims with their exact text spans, ready to be passed to an evidence retrieval system or a claim-matching prompt. This prompt does not classify the type of content, summarize it, or evaluate its argumentative structure. It does one job: isolate verifiable statements.

Do not use this prompt when you need to understand the author's argument, assess the overall credibility of a document, or generate a summary. It is not designed for sentiment analysis, stance detection, or narrative evaluation. It will deliberately ignore the rhetorical purpose of the text, which is a feature for verification pipelines but a bug for content analysis workflows. Also avoid this prompt when the input contains no factual claims—it will correctly return an empty set, but that may not be the behavior you want if you are expecting some output. For high-risk domains like healthcare or legal, always route the extracted claims to a human review step before treating them as verified, because the isolation step itself does not check truth, only checkability.

This prompt template is designed to be dropped directly into your AI system. It forces the model to act as a strict extraction engine, returning only atomic, checkable claims with their exact source text spans. The instructions are written to be model-agnostic, but you should test them against your specific model's tendencies, particularly its inclination to over-include interpretive language or to fragment compound facts. Before using this in production, run it against the failure mode tests described in the full playbook to calibrate your acceptance thresholds.

textCopy
You are a strict claim extraction engine. Your only job is to identify and isolate statements of verifiable fact from the provided text. You must ignore all analysis, opinion, prediction, interpretation, rhetorical questions, and commentary.

A verifiable factual statement is a claim that can be proven true or false using external evidence, regardless of whether that evidence is currently available to you. It must be a concrete assertion about reality, not a belief, judgment, or hypothetical.

# INPUT
[INPUT_TEXT]

# OUTPUT_SCHEMA
Return a JSON object with a single key "verifiable_claims" containing an array of claim objects. Each claim object must have the following fields:
- "claim_text": The atomic factual statement, rewritten as a single, self-contained assertion in the present tense.
- "source_span": The exact verbatim text from the input that contains the factual assertion.
- "span_start_index": The starting character index of the source_span within the original input.
- "span_end_index": The ending character index of the source_span within the original input.

If no verifiable factual statements are found, return an empty array.

# CONSTRAINTS
- Do not include any statement that contains hedging language (e.g., "may," "could," "suggests," "likely").
- Do not include predictions about the future, even if stated confidently.
- Do not include statements of value, quality, or sentiment (e.g., "is the best," "is terrible").
- Do not include statements that are true by definition or unfalsifiable.
- Break down compound sentences with multiple facts into separate claim objects.
- The "claim_text" must be understandable without the surrounding context.
- Do not add any commentary, preamble, or postamble outside the JSON object.

# EXAMPLES
Input: "The company, founded in 2012 by Jane Doe, released its flagship product last quarter. Analysts believe it could disrupt the market."
Output:
{
  "verifiable_claims": [
    {
      "claim_text": "The company was founded in 2012.",
      "source_span": "founded in 2012",
      "span_start_index": 16,
      "span_end_index": 30
    },
    {
      "claim_text": "The company was founded by Jane Doe.",
      "source_span": "by Jane Doe",
      "span_start_index": 31,
      "span_end_index": 41
    },
    {
      "claim_text": "The company released its flagship product in the quarter before the statement was made.",
      "source_span": "released its flagship product last quarter",
      "span_start_index": 43,
      "span_end_index": 82
    }
  ]
}

Input: "This is the most innovative solution on the market. Everyone should switch immediately."
Output:
{
  "verifiable_claims": []
}

# RISK_LEVEL
[RISK_LEVEL]

If RISK_LEVEL is "HIGH", you must also include a "confidence" field ("HIGH" or "LOW") for each claim, and you must strictly prefer under-extraction to over-extraction.

To adapt this template, start by replacing [INPUT_TEXT] with your content source. Set [RISK_LEVEL] to "HIGH" for regulated domains like healthcare or finance, where a false positive (treating an opinion as a fact) is more dangerous than a false negative. For lower-stakes content triage, you can omit the confidence field to reduce output token usage. The most common adaptation is adjusting the constraint list: if your domain considers well-sourced analyst projections as verifiable, you would remove the prediction exclusion. Always run the adapted prompt against a golden test set of 20-30 examples that include tricky boundary cases before deploying.

The Verifiable Statement Isolation prompt is designed to be a pre-processing step in a larger fact-checking pipeline. It should be called before evidence matching, scoring, or report generation. The prompt expects raw text input and returns a structured list of atomic claims with source spans. In production, you will typically call this prompt via an API (OpenAI, Anthropic, or a self-hosted model) and parse the JSON output into your verification database. The prompt's value is in its strict filtering: it discards analysis, opinion, and rhetorical framing, leaving only statements that can be checked against external evidence. This makes downstream verification steps more reliable because they are not wasting compute on unverifiable content.

Validation and Schema Enforcement: The prompt template includes an [OUTPUT_SCHEMA] placeholder that should be replaced with a strict JSON schema. Use a library like pydantic or zod to validate the model's response before accepting it. Each claim object must have a non-empty statement string, a source_span with valid start and end character offsets, and a claim_type enum value. If validation fails, do not silently drop the record. Instead, implement a retry loop (max 2-3 attempts) that feeds the validation error back to the model with a repair instruction: "Your previous output failed schema validation. The field 'source_span.start' was missing for claim index 4. Please regenerate the full output with all required fields." Log every validation failure and retry attempt for observability. For high-stakes domains like legal or medical content, route any output that fails validation after max retries to a human review queue rather than proceeding with partial data.

Model Choice and Latency Budgeting: This prompt works best with frontier models (GPT-4o, Claude 3.5 Sonnet) because the task requires nuanced linguistic judgment to distinguish factual assertions from interpretive language. Smaller models (GPT-4o-mini, Claude Haiku) can be used for high-throughput, lower-stakes content, but expect higher false-positive rates on interpretive statements. If you are processing long documents, chunk the input into sections of 2,000-4,000 characters with overlapping windows of 200 characters to avoid splitting claims at chunk boundaries. Run the prompt on each chunk in parallel, then deduplicate claims that appear in overlapping windows by matching on normalized statement text and source span proximity. Logging and Traceability: Store every prompt input, raw model output, validated output, and any retry attempts in your observability platform. Attach a prompt_version and model_id to each record so you can trace regressions when the prompt or model changes. For audit-heavy workflows, also store the full prompt template with all variables resolved so reviewers can reconstruct exactly what the model saw.

Next Steps After Isolation: Once you have a validated list of atomic claims, the typical next step is to route them to an evidence matching prompt or a retrieval system. Do not pass the full original document to the evidence matcher—only pass the isolated claims. This prevents the evidence system from being distracted by interpretive language. If your pipeline includes a confidence scoring step, use the claim_type field to adjust scoring thresholds: numerical claims may require tighter tolerance windows, while attribution claims may require exact quote matching. Finally, build a dashboard that tracks the ratio of isolated claims to total input length over time. A sudden drop in claim density may indicate a model behavior change or a shift in your input content type, both of which warrant investigation.