26 Sep 2026 · 7 min readWriting

How to Write Summaries That Actually Work

Most bad summaries are bad in one of two ways: too long to be a summary, or so compressed that the caveats and the numbers disappear. This guide covers the four types of summary and when each is right, what can and cannot be cut, a reliability check for lost meaning, and how to get a genuinely useful summary out of an AI tool without taking its word for it.

Four kinds of summary, and when each is right

Executive — the decision, the recommendation, the cost, the risk, and what happens next. For someone who will act on it and not read more.
Abstract — the problem, the approach, the key result. For someone deciding whether to read the whole thing.
Descriptive — what the document covers, section by section, without judging it. When the reader needs orientation, not conclusions.
Critical — the above plus an assessment of quality, method and limitations. The only honest version for anything academic or technical.

💡 Pro Tip: Most 'summaries' that disappoint are the wrong type. A paper's abstract cannot double as an executive summary, and a critical summary cannot double as a neutral one.

What to cut and what never to cut

Compressing means choosing. The failure is cutting at random rather than by decision.

Cut: repetition, throat-clearing, background everyone in the audience knows, method narration, and hedging.
Keep: the conclusion, the specific numbers, the caveats and limitations, the conditions under which it holds, and anything that would change the reader's decision.
Keep: disagreements and unresolved questions. A summary that presents a contested topic as settled is actively misleading.
Keep: who did the work, if it affects credibility.

⚠️ Note: The single most damaging thing you can cut is a limitation. It is the least interesting sentence in a paper and the reason two studies that look identical aren't.

Length that fits the purpose

One sentence: only for a headline claim. Fine for 'X reduces Y by Z%' and nothing more.
Three to five sentences: a proper abstract. The most useful single format.
A paragraph: for a report someone will skim before deciding to read it.
Bullet points: when the source is a list anyway — a spec, a meeting, a set of findings. Never convert prose into bullets; the structure is the information.

Check for lost meaning

Compression fails quietly. A summary can be accurate sentence by sentence and still misrepresent the source, because the caveats travelled separately from the claims.

Read the summary, then the conclusion of the source. Do they agree, including the hedging?
Check every number in the summary against the source. Numbers are where compression goes wrong most often.
Ask: could this summary be used to mislead someone without contradicting any sentence in it? If yes, it's too compressed.
Check that qualifiers travelled with their claims: 'in some cases', 'up to', 'in the tested range'.

Getting a summary out of an AI tool

AI summarisation is fast and usually good on structure, and it reliably flattens uncertainty — 'may', 'is associated with' and 'in this sample' come out as flat assertions. So specify the type and check the qualifiers.

These commands produce the summary and, just as importantly, the material to check it against.

/Summarize

Summarize Anything

Compress any text into a faithful summary at the length and reading level you choose.