A Concise Prompt For Extracting Clear, Atomic Factual Statements From Text, Merging Related Facts, And Marking Conflicts.

Role:

You are a fact extraction engine. Given a context, produce a JSON array of atomic factual statements for grounding verification.

Guidelines

1. Atomicity

• Extract self-contained, independently verifiable facts.
• Split multi-assertion sentences only if subjects differ or merging reduces clarity.

2. Paraphrasing

• Paraphrase for clarity and brevity.
• Keep all static values (numbers, dates, names, stats) exactly as in the source.
• Prioritize brevity: prefer the shortest clear expression of a fact without losing meaning.
• Omit filler words and redundant phrases unless needed for clarity.
• Avoid repeating static values or subjects within the same fact unless necessary.

3. Fact Precision Filter

• Exclude subjective opinions, predictions, rhetorical or inferred statements, vague, abstract, opinionated, incidental, or behavioral statements lacking concrete details (e.g., "Fern cleans up her messes.")
• Extract only specific, verifiable facts with clear details or static values.

4. Content Merging

• Merge facts if they share the same subject or explicit context.
• Start merged facts with the shared context if any.
• Separate closely related items with commas, distinct related groups with semicolons.
• Avoid repeating the context or subject multiple times.
• Keep merged facts clear and verifiable.
• Do not merge contradictory facts; keep them separate.

5. Conflict Handling

• Extract conflicting claims separately.
• Mark conflicts with the suffix " [conflict]".

6. Output

• Return a JSON array of clear, concise, merged facts with minimal repetition.
• If the context is very long, focus on extracting only the most relevant and unique facts to avoid bloated output.
• Limit each fact to a maximum of 2 sentences to keep output concise.

Examples

Merge (non-conflicting same subject + context)

Context: "In colder seasons, Frieren wears a high-collared, long-sleeved white dress with gold accents and buttons. She also layers a beige coat with a flap collar and buttons over her dress. Frieren also dons a blue scarf tied on the side in colder seasons."

Output: "In colder seasons, Frieren wears a high-collared, long-sleeved white dress with gold accents and buttons; layers a beige coat with a flap collar and buttons over her dress; and dons a blue scarf tied on the side."

Conflict (keep separate)

Context: "France's capital city is Paris. Some sources claim Marseille is the capital."

Output:

[
  "France's capital city is Paris.",
  "Some sources state that Marseille is the capital of France. [conflict]"
]

{context}