What A Summary Is

The mechanics of cutting things down

I used to think summaries were just shorter versions of longer texts. That assumption cost me a few too many client meetings early on. The actual job is more like translation between compression formats. You are taking information from one structure and mapping it into another while losing as little signal as possible. Most people get that wrong and end up producing thin recaps that feel empty. What A Summary Is, really, is a fidelity-preserving reduction operation. You strip away the scaffolding — the narrative setup, the tangential evidence, the repetitive elaboration — and leave only the load-bearing claims intact. If someone reads your summary, they should be able to reconstruct the original argument's skeleton without needing to go back to the source material.

The three-layer method I actually use

When I get a document that needs condensing, I work through it in three passes. First pass is extraction. I read through and pull out every claim, finding, or conclusion that stands on its own. This usually takes longer than people expect on dense material. A forty-page technical report might yield anywhere from twelve to twenty standalone propositions depending on how much of it is decorative. I write them down in a flat list without worrying about order or phrasing. Second pass is pruning. I look at that list and ask which claims depend on other claims for their meaning. If proposition B is just a restatement of proposition A with a different example attached, A survives and B gets cut. This is where most summaries bloat — people keep the supporting examples instead of keeping the claims the examples support. The example is scaffolding. You remove it. Third pass is reconstruction. I arrange the surviving claims into a logical sequence that mirrors the original's argumentative structure but uses my own sentences. Not paraphrased originals. My own sentences, because paraphrasing keeps the rhythm and idiom of the source text, which defeats the purpose. A summary should sound like it was written by someone who absorbed the material and is now explaining it to a colleague.

I typically compress a 3,000-word piece down to around 400 words using this method. The ratio varies. Dense policy documents compress better than narrative-driven pieces because there is less embedded context to carry along.

A problem that tripped me up and the workaround

Here is a specific edge case that caught me off guard. I was summarizing a compliance audit report for a financial services firm. The original document had seventeen findings, but nine of them were framed as conditional — "if X condition is met, then Y risk exists." I initially wrote all seventeen as flat factual statements. When I sent the summary to the client's legal team, they flagged three of the findings as misleading because the conditional qualifiers got stripped out in the extraction pass. Those conditions were not minor details. They were the entire basis for whether the finding was actionable. The workaround was simple once I spotted it. In the extraction pass, I started tagging every claim with its dependency type. Factual, conditional, hypothetical, or disputed. That tag stayed attached through the pruning phase and dictated how I phrased it in the reconstruction. Conditional claims needed "if" clauses intact. Hypothetical claims needed hedging language preserved. It added maybe ten percent to my extraction time but eliminated the revision loop that would have taken hours.

Counter-intuitive truths beginners miss

One thing that always surprises people is that a good summary can be longer than you think it should be. There is a prevailing myth that brevity is the sole virtue of summarization. It is not. Fidelity is. If the original text contains five independent conclusions, each requiring two sentences of clarification to be intelligible on its own, your summary is four paragraphs, not two. Trying to force it shorter will either drop conclusions or produce vague statements that misrepresent what the source actually said. A client once asked me to compress a six-page risk assessment into a single paragraph. I told him that was impossible without losing material he'd specifically asked me to include. He insisted. The resulting paragraph was technically accurate but useless. He brought it back three days later and asked me to do it properly. Another counter-intuitive point: the best summaries often include one specific detail from the original that no analysis would predict should survive. Usually it is a number, a date, or a named entity that anchors the abstract claims in concrete reality. Remove it and the summary becomes abstractions all the way down, which makes it harder rather than easier for a reader to evaluate whether the source material warrants their attention.

Where this approach breaks down

This method does not work well with text that is fundamentally experiential rather than informational. Personal narratives, creative nonfiction, opinion essays where the value lives in the voice rather than the conclusions — summarizing those produces something technically accurate and emotionally hollow. I learned that the hard way summarizing a memoir excerpt for a publishing client. I applied the three-layer method and came out with a paragraph that captured every plot point but read like an autopsy report. The client said it was correct and worthless. I stopped trying to force that approach onto voice-dependent material and instead produce what I call thematic distillations, which preserve the emotional arc without attempting to compress the itself. Different tool for a different job. Also, this approach assumes the source text has a coherent argument structure. Texts that are deliberately fragmentary, or collections of loosely related observations, resist this method because there is no skeleton to preserve. You end up making interpretive choices about what counts as a "claim" that introduce bias the original did not contain.

Tools that help and ones that don't

I use a basic outline view in whatever word processor I am working in. No special software required. Some people reach for AI summarization tools at this stage. They can handle the first pass reasonably well on straightforward material but they consistently fail on the conditional-tagging problem I described above. They also have a tendency to hallucinate connections between claims that do not exist in the source, which introduces errors that are very hard to spot because the output sounds fluent. I've reviewed AI-generated summaries where the prose quality was high but three of the five conclusions had been subtly altered to sound more confident than the source justified. That is a real risk you need to factor in if you outsource the reconstruction pass to a model. The manual method takes longer initially but the error rate is near zero because you are reading the source at each step. For high-stakes summaries — legal, medical, regulatory — the extra time is non-negotiable.

Quick reference for compression ratios

Technical documentation with heavy procedural content compresses at roughly 10 to 1. Narrative reports with supporting evidence compress at 6 to 1. Policy briefs with conditional recommendations compress at 5 to 1. Executive summaries of academic papers often end up being 8 to 10 pages when the original is 30 pages, which feels long until you realize that every sentence in the source contains dependent claims that cannot be removed without changing the meaning. The ratio is not the goal. The goal is the thing the reader needs to know after finishing your summary, minus the thing they only needed to know to follow the original's reasoning. Strip the reasoning scaffolding. Keep the load-bearing claims. Tag the conditions. Write it in your own voice. Verify the numbers against the source before sending it out.