Most people type "summarize this" and then wonder why the result is useless in a meeting. The model did summarize. It just optimized for a school-book abstract: topic sentences, balanced tone, no decisions, soft numbers. That format is fine for proving you read the PDF. It is a bad format for work.
A useful summary has a job. Until you name the job, you will keep getting elegant mush.
Four jobs a summary can have (pick one)
| Job | Question it answers | Output shape that works |
|---|---|---|
| Decide | What should we choose? | Options, tradeoffs, recommendation, open risks |
| Act | What do we do next? | Owners, deadlines, blockers, exact next steps |
| Remember | What must not be lost? | Facts, numbers, quotes, constraints: high fidelity |
| Brief | What does [role] need in 90 seconds? | Audience-specific one-pager: context, stakes, ask |
"Key points" is not a job. It is how models stall when you under-specify.
The five inputs that stop point-loss
- Source type: transcript, contract, paper, Slack thread, earnings call.
- Reader: you tomorrow, your VP, a client, a new hire.
- Job: decide / act / remember / brief.
- Hard constraints: max length; must keep all $ and %; must flag unknowns.
- Do-not-list: "no fluff openers," "no inventing owners," "no softening bad news."
Prompt pattern:
"Summarize the text below for [reader] whose job is to [decide/act/remember/brief]. Length: [N words / bullets]. Structure: [sections]. Preserve every number and proper name. If something is ambiguous, write UNCLEAR: do not guess. Do not add advice that is not in the source."
The UNCLEAR rule is the difference between a summary and a hallucination with good posture.
Worked example: meeting transcript
Bad: "Summarize this meeting."
Typical output: Three paragraphs about "alignment," "next steps," and "productive discussion" with zero owners.
Good:
"Audience: eng manager who missed the meeting. Job: act. Output:
1) Decisions (only if explicitly decided)
2) Action items as table: task | owner | due date or NO DATE
3) Open questions
4) Risks mentioned
Ignore banter. Do not invent owners."
You will get something you can paste into the project channel without rewriting from scratch.
Worked example: long PDF / report
Bad: "Give me the key takeaways."
Good:
"Job: remember + brief a CFO. 200 words max. Lead with the three numbers that change a budget decision. Then: what the authors recommend, what evidence is weak, what is out of scope. Quote one sentence if it is load-bearing. No marketing tone."
Notice you asked for weak evidence. Generic summaries bury uncertainty because uncertainty sounds less "confident."
Compression without losing the point
| Technique | Use when | Risk |
|---|---|---|
| Hierarchical: 5 bullets → 1 paragraph | You need both skimmable and pasteable | Over-condensing numbers |
| Extractive first, then rewrite | Legal/compliance-ish text | Still need a human check |
| Section-by-section then rollup | 50+ page docs | Cost/latency; still need global thread |
| Claim / evidence pairs | Research, strategy decks | Model may pair loosely |
For very long sources, summarize in chunks with the same job and schema, then summarize the summaries with: "Merge these partial summaries. Resolve contradictions as CONFLICT: …"
Common ways summaries quietly lie
- Averaging opinions: two people disagreed; the summary says "the team felt."
- Dropping negations: "not approved" becomes "discussed approval."
- Inventing structure: adding a roadmap that was only a wish.
- Rounding away the point: "$47,200" becomes "about $50k" when the variance matters.
- Politeness compression: serious customer anger becomes "some concern."
If the source is emotional, legal, or financial, require a fidelity pass: "List every sentence you wrote that is not entailed by the source."
A 60-second QA before you send the summary
- Can you point to a source line for every number?
- Did any decision get upgraded from "discussed" to "decided"?
- Would the original author say you missed the actual point?
- Does the reader know what to do (if job = act)?
- Is the length right for the channel (Slack vs. board memo)?
If #3 fails, your prompt asked for the wrong job: not just "shorter."
Tool path on aihowto.pro
The Summarizer is built around length and focus so you are not stuck with a generic gist. For turning a summary into durable study material (definitions, quizable facts), chain into Study Notes instead of asking one prompt to be both executive brief and flashcards.
When not to summarize with AI
- You need a privileged legal interpretation (use counsel; AI can outline issues only)
- The document is adversarial and every clause matters (diff tools + human reading)
- You do not have rights to paste the content into the tool you are using
Summarization is a productivity feature, not a substitute for duty of care.
Bottom line
"Summarize this" produces abstracts. Useful summaries start with a job: decide, act, remember, or brief. Lock numbers, ban guessing, pick a structure, and QA for silent lies. The point survives when you tell the model which point you are trying to keep.