Summaries: Too Long vs Just Right (Prompt Comparison)

Sixty-eight pages and 1,240 public comments arrived as a single PDF on a Monday, and the planning committee meets on Friday morning. Dana Whitfield has four days to give a board that has not read a word of it the one thing they need: which objections could stall the Harrow Point wind farm, and how many people raised each. If her summary is the wrong length, the board either drowns or misses the decision.

This article compares two summarize prompts on that report. The first asks for a short summary and gets a long one. The second asks for a specific decision to be supported and gets a page the board can act on. You will see both prompts, why length alone is the wrong instruction, and a reusable summary prompt with a table of targets.

Why a 68-Page Report Still Isn’t Summarized

“Keep it short” is not a specification. Short compared to what, for whom, and carrying which facts? When the only instruction is brevity, the model plays safe and compresses evenly, which means it keeps the report’s own structure and its own headings. Dana gets nine hundred words that echo the table of contents and still does not tell the board which objection is the dangerous one.

The report also has a shape the summary must respect. Of the 1,240 comments, most repeat a handful of themes: turbine noise at the nearest farmhouse, bird and bat collision on the migration corridor, the visual impact on the ridge line, and shadow flicker in summer evenings. A flat summary treats a theme raised by 600 people the same as one raised by 4. That flattens the exact signal the committee is looking for.

Finally, a summary without an audience has no test for inclusion. The planning board needs risk and counts; a project engineer would need technical objections; the public needs plain language. Same report, three different summaries. The prompt has to say which one it is writing.

Two Prompts: Everything vs the Decision

Both prompts target the same 68-page document. Compare what each one tells the model to care about.

Summarize this 68-page report of public comments on the Harrow Point wind
farm. Keep it short and cover the main points. The report is attached.

That request names no reader, no length, and no purpose, so the model defaults to reproducing the report’s structure at reduced size. You get something long and even, and Dana still has to do the real work by hand. Now the second request.

You are an analyst briefing a planning committee that has not read the source.
Write a 250-word executive summary of the attached public-comment report on
the Harrow Point wind farm.

The committee must decide, on Friday, whether to approve. So the summary must
answer:
- Which objections could delay approval, ranked by how many of the 1,240
  comments raise them, with the count beside each.
- For each of the top three, the strongest single argument a resident made,
  quoted in twelve words or fewer.
- Which objection is supported by the fewest comments and is safest to
  dismiss.

Rules:
- Hard limit 250 words. Count them.
- Use no heading and no bullet longer than one line.
- Do not describe the report's sections; describe the decision.
- If a number is not in the source, write "not stated" instead of estimating.

The first prompt asks for a smaller copy of the report. The second asks for a briefing that supports a vote. Same source, and yet only one of them produces something Dana can hand to a board on Friday.

Compression Needs a Target, Not a Word Limit

• A name for the reader. “A planning committee that has not read the source” tells the model what can be assumed and what must be spelled out. Brevity with no reader is guesswork.
• A decision to support. Framing the summary around the Friday vote turns a survey into an argument, which is what actually shortens it. You drop everything that does not move the decision.
• Counts beside each theme. Ranking objections by how many of the 1,240 comments raise them is the single most useful line in the brief. It forces the model to weigh, not just list.
• A quoted argument per theme. A twelve-word resident quote carries more weight with a board than a paraphrase, and it is harder to dismiss.
• A hard word limit that is counted. Saying “count them” makes the model trim its own draft instead of overrunning by a third.
• A no-guessing rule. “Not stated” for any missing figure keeps a summary from quietly inventing a number that ends up in the minutes.

Notice that every one of these is a target, and none of them is the word “short.” A target is checkable; a vibe is not. That is the core lesson, and it holds whether you are briefing a board or condensing a long article, as our walkthrough on how to summarize an article with ChatGPT shows.

Set the Target Before You Summarize

Pick the row that matches your reader, then write the prompt to match. The length is a consequence of the target, not the other way around.

For Dana, the board row applies, and the brief above is built for it. If your reader is a peer in the same field, the trade-offs shift and the approach in our notes on research summary prompt settings is a closer fit.

The Summary Prompt I Reuse

Fill the blanks and keep the structure. It fits comment reports, meeting transcripts, and long studies because it always starts from the reader and the decision, not the word count.

You are an analyst briefing {{reader who has not read the source}}. Write a
{{length}} summary of the attached document on {{topic}}.

The reader must decide or do: {{the decision or next action}}.
So the summary must answer:
- {{question 1, with any count or ranking}}.
- {{question 2, with the format you want}}.
- {{question 3, naming the smallest or safest item}}.

Rules:
- Hard limit {{length}}. Count the words.
- Use no heading and no line longer than one line.
- Describe the {{decision}}, not the document's sections.
- If a figure is not in the source, write "not stated" instead of guessing.

The same target-first thinking is what separates a usable briefing from a transcript dump. If your input is a recorded call rather than a report, the pattern in meeting summary: wrong vs right applies the identical rule to spoken notes.

What the 250-Word Brief Actually Says

Run the second prompt and the shape of the output is predictable, because the brief forced it. The opening line names the decision and the volume of input: 1,240 comments, 68 pages, one approval pending. The body ranks the objections with their counts, so turbine noise at the nearest farmhouse leads on 612 comments, the migration-corridor concern follows on 400, and visual impact on the ridge line trails on 96.

Each of the top three carries one short resident quote, and the close names the objection safe to set aside, the four comments about shadow flicker on a summer evening. There is no description of the report’s chapters and no phrase like “various stakeholders raised concerns.” Every sentence either supports the vote or leaves.

That is what a target produces. Compare it with the first prompt’s output, which would have preserved the report’s own structure and handed the board nine hundred words that tell them, accurately and uselessly, that people had opinions.

Questions to Settle Before You Compress

What if the source is longer than the context window? Chunk it by section, summarize each chunk against the same target, then run a second pass over the chunk summaries. Compression in two stages beats one truncated pass.

Should I ask for quotes? Yes when a human will act on the summary. A short verbatim quote survives re-reading and is hard to argue with; a paraphrase invites disagreement about what the person meant.

How do I stop the model from averaging? Ask it to rank and count rather than list. Ranking forces it to separate the theme with 612 comments from the one with 4, which is the difference a board actually needs.

Sources & Notes

Dana Whitfield, the Harrow Point wind farm, and the 1,240 comments are illustrative; no real project or filing is described. Compression quality depends on the model reading the full source, so check that your document actually fits the context window before you trust a summary of it. Always keep the raw counts and quotes traceable to the source, because a summary that drops a figure is safer than one that invents it.