Summary Is Decided by Whether You Give a "Template"
Send only "please summarize" and AI returns a bland short text. For the same material, specifying target, use, reader, and granularity produces a summary you can actually use, like a different thing. This article explains concretely what "template" to ask for in three typical cases: long text, PDF, and minutes.
Long-Text Summary: Shrink in 2 Stages
For long text over ~10,000 characters, 2 stages is more stable than summarizing all at once.
- Summarize each chapter/section in 200–500 chars
- Gather those and have it produce a final 800–1,500-char summary
As of 2026 the main AIs handle ~1M tokens at once, so a single pass is enough for short materials. Still, 2 stages lowers the risk of "missing the first half" in long materials and makes chapter granularity easier to align.
Summarize the following chapter in 300 chars.
- Keep important numbers/proper nouns
- State the author's claim and evidence
- Don't add your own interpretation
[body]PDF: Accuracy Changes with How You Hand It Over
The main 2026 AIs are multimodal, so attaching the PDF directly is the easiest. Only when scan quality is poor or layout is complex does pre-converting to text with a dedicated extraction tool (text extraction like pdftotext, Adobe's extraction, an OCR service) stabilize things.
When you'll use tables/items downstream, extracting as JSON while keeping structure makes it easier to handle.
Extract this PDF in the following structure.
{
"title": string,
"sections": [{ "heading": string, "content": string }],
"tables": [{ "caption": string, "rows": [[string]] }],
"key_findings": [string]
}Minutes: Separate Decisions from Homework
For minutes, "decomposition" matters more than "summary." From a meeting transcription or a Slack thread, have it format into this structure.
Organize the following conversation log in this structure.
## Decisions (concretely decided things)
## Actions (owner / content / deadline)
## Issues (discussed but undecided)
## Key statements (speaker / point)
## To prepare by next time
If owner/deadline is vague, explicitly write "needs confirmation."Quick Table of "Granularity" by Use
| Use | Granularity guide |
|---|---|
| One-line summary | 30–50 chars, for a sharing message |
| Executive summary | 200–300 chars, 5 points |
| Article teaser | 500–800 chars, makes you want to read on |
| Work-report summary | 1,500–3,000 chars, with evidence |
| Verbatim summary | ~1/3 of original, keeps logical structure |
Summary Changes by "For Whom"
Change the content by reader even for the same document. Just one line, "summarize for [role]," in the prompt shifts the perspective.
- For executives: conclusion first, impact and decision points
- For engineers: implementation points, technical trade-offs
- For marketing: appeal points, competitor comparison
- For customers: benefits, price, how to apply
5 Instructions to Raise Quality
- Explicitly state "don't add your opinion"
- Specify "keep primary-source numbers"
- Narrow with "only the top N by importance"
- After summarizing, ask "what key points are missing?" (self-check)
- Check whether the original issues can be reconstructed from the summary (reverse test)
Common Misunderstandings / Failure Patterns
- Too long, ultimately unread: decide granularity first
- Key numbers dropped: explicitly say "keep original numbers"
- AI adds interpretation not in the original: say "no interpretation" and cross-check output against the source
- Subjective words like "wonderful," "groundbreaking" mixed in: instruct a ban on modifiers
- Confidential handling: check terms for internal-only/customer info and anonymize if needed before passing
Summary
Specify target × use × reader × granularity and you use AI's power directly. Long text in 2 stages, PDF via multimodal or an extraction tool, minutes via structuring—remember this split and it rides into practice without drift every time. The next article covers polishing your writing with AI: proofreading, revision, and tone adjustment.
