Long-document summarization

The task

You need a summary of a long report, contract, transcript, research pack, or documentation set. The useful output depends on the reader’s next decision.

Why the old approach is outdated

“Summarize this document” treats every detail as equally useful. It often produces a neat abstract while missing the risk, decision, or action the reader needed.

Long-context models make it easier to include the whole source, but the prompt still needs an explicit reading purpose. The context window management concept explains why a large window is not the same as useful memory.

The current approach

Summarize this implementation plan for an engineering lead deciding whether to approve it.

Return:
- The proposed change in plain language.
- The strongest reason to approve it.
- The blocking risks.
- Missing facts that would change the decision.
- Source sections that support each risk.

Do not summarize background sections unless they affect the approval decision.

When to use a different model

Use ChatGPT when the summary feeds into code, spreadsheets, structured data, or a broader API workflow. Use Gemini when the source includes mixed media. Use Perplexity when you first need to find current sources before summarizing them.

What to avoid

• Asking for a “concise summary” without saying concise for whom.
• Letting the model ignore section names or source references.
• Combining extraction, judgment, and rewriting in one prompt when the document is large.
• Keeping stale conversation context beside the document unless it is still relevant.