Effective Summaries

A summary's only job is to let the reader decide, accurately, whether they need the full source — or to stand in for it when they don't have time to read it. Every word that isn't in service of that job, and every claim not actually supported by the source, is a defect, not a stylistic choice.

Length targets

• Scale summary length to source length, not to a fixed word count: roughly 5–10% of the original for a short article (under ~2,000 words), 2–5% for a long report or paper, and closer to 1% or less for a book-length source where a page or two is the ceiling most readers will tolerate.
• Never pad a summary to hit a length target. A 3,000-word article that genuinely summarizes to 80 words is better served by 80 honest words than 300 words of restated filler.
• When a specific length is requested (e.g. "summarize in 3 sentences" or "under 200 words"), treat that as a hard constraint and prioritize the most load-bearing information within it — cut detail, never accuracy.

Fidelity rules

• Include only claims that are actually in the source. Never add outside knowledge, context, or interpretation the source doesn't state, even if it's true and you know it — a summary is a compression of the source, not an enrichment of it.
• Never resolve an ambiguity the source leaves unresolved by picking the interpretation that reads more cleanly. If the source is genuinely ambiguous or contradicts itself, say so rather than silently choosing a side.
• Preserve the source's own confidence level. If the source says "may cause" or "researchers suggest," don't upgrade it to "causes" or "researchers found" — hedged claims stay hedged.

Structure

• Lead with a TL;DR: one to three sentences that would let a reader who stops right there walk away with the core takeaway.
• Follow with key points as a scannable list, ordered by importance to the source's own argument, not by the order they appeared in the text.
• Close with caveats: what the source itself flags as a limitation, an open question, or a caveat to its own claims (sample size, scope, date of the data, stated assumptions) — these are frequently the part a reader most needs and most often gets cut.

Quoting vs. paraphrasing

• Paraphrase by default; it's more space-efficient and easier to scan than a wall of quotations.
• Quote directly only for language where the exact wording is the point — a specific commitment, a striking or contested claim, a legal or technical definition. Mark it clearly as a quote with attribution to the source.
• Never present a paraphrase as if it were a direct quote, and never lightly edit a quotation (even to "clean it up" grammatically) without marking the change.

Numbers and units

• Reproduce every number, unit, date, and percentage exactly as given in the source — don't round, don't convert units, and don't recompute a derived figure unless explicitly asked to, since a summarizer's arithmetic is not the source's data.
• If the source gives a range or a margin of error, keep it; collapsing "between 12% and 18%" down to "about 15%" discards real information and can misrepresent the source's own confidence.

Multi-document synthesis

• When summarizing several sources together, attribute each claim to which source it came from whenever sources disagree or only partially overlap — a synthesized summary that blends conflicting sources into one unattributed claim hides the disagreement instead of surfacing it.
• Note explicitly when sources agree (strengthens the claim) versus when only one source makes a claim the others don't address (weaker, standalone) — these are not equivalent and a synthesis should not flatten them into the same confidence level.
• Preserve chronology when sources represent an evolving situation (e.g. a story updated over several reports) — state what was known when, rather than merging all sources into a single present-tense account that erases what changed.