When to use this prompt
When you have a paragraph that is informative but written in a way that AI engines cannot easily extract. Symptoms: the paragraph requires the reader to know context from earlier in the page, makes its main claim halfway through rather than upfront, or wraps the claim in qualifying language that obscures it.
Run this prompt across every key paragraph on your priority pages: service definitions, comparison claims, FAQ answers, and the first paragraph under each H2. These are the passages most likely to be extracted by an LLM if you make them extractable.
The prompt
<role>Editor who specializes in restructuring paragraphs for AI retrieval and citation.</role>
<task>Rewrite the paragraph below so a retrieval system can extract a fully self-contained citation from it. Preserve voice and length. Change information order and specificity, not style.</task>
<inputs>
<page_topic>[TOPIC]</page_topic>
<brand>[BRAND]</brand>
<audience>[AUDIENCE]</audience>
<original>
[PASTE PARAGRAPH HERE]
</original>
</inputs>
<instructions>
1. The first sentence must contain the main claim and stand on its own. A reader who never saw the rest of the page must be able to understand and cite it.
2. The next two to three sentences must supply specific, factual, attributable evidence (numbers, named entities, dates, integrations).
3. Replace vague qualifiers ("many", "often", "various", "some") with specific quantities, examples, or named entities.
4. Do not expand the paragraph. Total length stays within ±15% of the original.
5. The rewrite must read naturally for a human. Do not turn it into a bulleted list.
6. Do not invent facts. If the original lacks a specific number, leave the claim qualitative rather than fabricating one.
</instructions>
<output_format>
Rewritten paragraph:
[the rewrite, no extra formatting]
Three biggest changes:
1. [change] — [why it improves extractability]
2. [change] — [why it improves extractability]
3. [change] — [why it improves extractability]
</output_format>How it works
LLMs cite passages, not pages. The retrieval step that feeds an answer engine usually pulls a 2 to 4 sentence chunk, evaluates whether it answers the user’s question, and either cites it or moves on. If the main claim is buried in sentence four, the chunk that gets pulled often does not contain it, so the citation goes to a competitor whose first sentence said the same thing.
The XML structure separates the original paragraph from the instructions, which prevents the model from interpreting parts of the original as commands. The “do not invent facts” line is a 2025 best practice: frontier models follow instructions literally, so explicit prohibitions outperform implicit ones.
The “three biggest changes” section is a forcing function. It catches cases where the model only made cosmetic edits and did not actually restructure the paragraph.
Example output
Original: “Our platform helps SaaS teams of various sizes work more efficiently. We offer a range of features including event tracking, AI-powered insights, and integrations with many popular systems. Many customers find that our solution helps them improve user retention.”
Rewritten: “Acme Analytics improves 30-day user retention by 18% for mid-market SaaS teams of 25 to 200. The platform combines event tracking, AI-assisted cohort analysis, and prebuilt integrations with Segment, Snowflake, and HubSpot. Two years of customer benchmarks across 180 SaaS teams support this number.”
Three biggest changes:
- Moved the main claim (18% retention lift) to sentence one with a specific audience.
- Replaced “various sizes” and “many systems” with named ranges and named integrations.
- Anchored the claim with a specific evidence base (180 customer benchmarks over 2 years).
Variations
- FAQ rewrite: Add a constraint that the answer must work as a standalone Q&A pair so it can be lifted directly into FAQ schema.
- Definition rewrite: Force the first sentence to start with “X is…” for definition-style passages that need to win definition queries.
- Comparison rewrite: When the original paragraph compares to a competitor, require the rewrite to lead with the named comparison and the differentiator.