Extract, then compress
Pull checkable facts with source locations before asking for any summary.
Summarizing is the most reliable thing AI does, but "summarize this" is close to the worst instruction you can give it. A summary without a purpose returns a shorter version of everything, weighted by what sounded important in the text rather than by what matters to you. Purposeful summarizing — naming the decision you are trying to make — turns the same tool into something closer to a research assistant. The pattern that works is extract, then compress. First ask for the specific elements you need pulled out: obligations, deadlines, dollar amounts, named parties, conditions, anything that changed. Ask for them in a structured shape with a location reference, so you can jump back to the source and confirm. Only after you have the extracted facts should you ask for a narrative summary, and it should be built from those facts rather than from a fresh read. Two cautions apply every time. A summary is a lossy compression, so anything you will act on financially or legally must be checked against the original passage — the extraction step exists precisely to make that check fast. And length is not fidelity: a five-line summary that names every deadline beats a page of fluent prose that mentions none. When a document is genuinely long, work in labeled chunks and summarize the summaries, keeping the extraction structure constant so nothing silently drops between passes. Extraction protects you because it is checkable. A narrative summary cannot be audited against the source quickly, but a table with page references can be verified in under a minute. Keep the extraction structure identical across documents of the same type. Consistency lets you compare two quotes or two contracts directly rather than re-reading both from scratch. A franchise owner reviewing three supplier agreements used the same extraction table for all three and immediately saw that only one included a price-escalation cap. A summary of each would have buried that difference.
Key takeaways
- The pattern that works is extract, then compress.
- Two cautions apply every time.
- Extraction with locations gives you checkable specifics.
Sign in to complete activities, save progress, and use the AI Coach.
Evergreen instructional material; no time-sensitive claims to source.
Evergreen curriculum written from KleinHub teaching material — no time-sensitive claims.
AI-generated lesson, reviewed by Education Review and approved through the KleinHub approval queue before publication.
Sign in to complete this lesson, track progress, and use the AI Coach.
Sign in to start lessonCreate account