Use when a long page gets retrieved in pieces and the pieces do not stand up alone.
retrieval-chunk-boundaries.md
You are splitting a page the way a retrieval system would, then judging each piece on its own. You are not editing for style.
Page text: {{PAGE_TEXT}}
Heading structure as published: {{HEADING_STRUCTURE}}
Questions the page is meant to answer: {{TARGET_QUESTIONS}}
Approximate words per chunk: {{CHUNK_SIZE}}
Split {{PAGE_TEXT}} into chunks of roughly {{CHUNK_SIZE}} words, breaking at the nearest paragraph or heading boundary. Do not adjust the split to make a chunk look better.
Output a table: | Chunk | First line, verbatim | Which of {{TARGET_QUESTIONS}} it could answer alone | What it depends on from earlier chunks (pronoun, defined term, table header, list stem) | Verdict (standalone / partial / orphan) |
Then give:
- The five orphans that matter most, each with the one sentence to add so the chunk survives alone.
- Any place where {{HEADING_STRUCTURE}} splits a single answer across two chunks, and where the heading should move instead.
- Chunks that answer none of the target questions, listed as candidates for cutting.
Constraints:
- Quote first lines verbatim. Do not tidy them.
- If a chunk answers a question the page never intended, note it rather than forcing it onto the nearest target question.
- Do not merge chunks. Report the split as it falls.
Replace each placeholder with your own detail. The more specific you are, the less the model invents.
When should I check chunk boundaries rather than edit sections?
When a long page keeps being answered from in part and the part quoted makes no sense alone. Section editing assumes the reader arrives at the top. This assumes they arrive in the middle, which is what happens when a retrieval system hands over one passage instead of the page.
What do I need before running it?
The page text, the heading structure as actually published, the questions the page is meant to answer, and a chunk size. Use the size your retrieval setup uses if you know it. If you do not, run it twice at different sizes, since the orphans move as the boundary moves.
What comes back, and which column do I work from?
A table giving each chunk's first line verbatim, which target questions it could answer alone, what it depends on from earlier chunks, and a standalone, partial or orphan verdict. The dependency column is the actionable one. It names the pronoun, defined term or stranded table header that breaks the passage.
What is the mistake that costs me?
Repairing orphans by repeating the full context in every chunk. The fix is one added sentence per orphan, which is exactly what the prompt asks for. Restating definitions everywhere produces a page nobody reads. Define a term once per section and keep a table header in the same section as its table.