Condense a Long Report or PDF into an Executive Brief

Workflow area: Documents · Tier: Foundation · Hands-on time: ~25 minutes

Summarization is the most-used and most-misused AI skill in office work. Models are genuinely excellent at compression — and completely willing to smooth over a caveat, average away a contradiction, or state your report's "maybe" as a "yes." The recipe below gets the brief fast, then spends five deliberate minutes making sure the brief still tells the truth.

When to use — and when not to

• Use it for reading you must act onconsultant reports, audit findings, market studies, policy documents, lengthy proposals where you need the "so what" and the load-bearing numbers.
• Use it to build the pre-reada one-page brief circulated before the meeting beats forty pages attached to the invite.
• Use it for triageten documents in, one question: "which three of these do I actually need to read in full?"
• Skip it when you are the accountable expertif you will sign off on the content, the summary is a map, not a substitute. Contracts and compliance documents get read in full, by you.
• Skip unauthorized documentsa client's confidential report does not go into a consumer tool because you were short on time. Tenant tools or nothing.

The recipe

1. 1.

   Get the document in whole. Upload the PDF directly to Claude or ChatGPT (both read PDFs), open it in Copilot, or add it as a NotebookLM source. Avoid pasting fragments — a summary of chapters one to three silently presented as the whole document is a classic self-inflicted error.

2. 2.

   Demand structure, numbers and locations. "Summarize this" produces mush. Ask for a brief with an audit trail:

   Copy

   Read the attached report and produce a one-page executive brief
   for [AUDIENCE, e.g. "our leadership team deciding whether to
   renew this vendor"].
   
   Format:
   1. BOTTOM LINE — 2 sentences: what this report concludes.
   2. KEY FINDINGS — max 5 bullets, each with its supporting number
      and the page/section it comes from.
   3. NUMBERS THAT MATTER — the 5 figures a decision-maker will ask
      about, as a table: Figure | Value | Where in the report.
   4. RECOMMENDED ACTIONS — what the report itself recommends. Mark
      anything YOU infer beyond the text as "Inferred".
   5. CAVEATS — limitations, assumptions and disclaimers the report
      states. Do not soften them.
   
   Rules: use only this document. Preserve hedged language — if the
   report says "may" or "preliminary", the brief says so too.
   

3. 3.

   Spot-check the load-bearing claims. Good output cites its sources:

   Copy

   KEY FINDINGS
   - Vendor missed SLA in 4 of 12 months, worst in Q3 (94.1% vs
     99% target) — Section 3.2, p.14
   - Support costs rose 18% YoY, driven by ticket volume — p.19
   CAVEATS
   - Uptime data self-reported by vendor; not independently
     verified (p.6)
   

   Open the document at three of those citations — the two biggest numbers and one caveat — and confirm the text says what the brief says. This is the five minutes that catches an inverted conclusion before your boss repeats it to their boss.

4. 4.

   Interrogate the document. The brief answers the report's questions; now ask yours. "What does the report say about pricing at renewal?" "Does it compare against any alternative vendor?" "What did it NOT examine?" That last question is the sleeper — a summary can only compress what is there, and the gap is often the story.

5. 5.

   Rewrite the bottom line in your own words. The two-sentence conclusion is what gets forwarded, quoted and remembered. Write that part yourself, from your reading of the findings — it forces you to actually hold the conclusion, and it puts your judgment (which is what you are paid for) on top of the machine's compression.

Make it yours

• Claude: strongest default for long PDFs — large context means a 100+ page report fits whole; ask it to quote exact sentences for each key finding so verification is copy-paste.
• ChatGPT: same recipe; for data-heavy reports ask it to recompute one derived figure ("does the 18% YoY claim match the table on p.19?") — a useful cross-check on the report itself.
• NotebookLM: best when the report joins a pile — add it with related documents, and every answer comes with clickable citations into the sources. The right tool when you will keep asking questions all quarter; pb-23 and pb-24 build this out.
• M365 Copilot in Word: open the document and ask for the brief in the chat pane — output lands as a draft you can circulate directly; keep the step-2 structure or you will get generic prose.

Quality check

• Three citations verified by handthe biggest two numbers and one caveat, checked against the page they claim.
• Hedges survived compressionthe report's "preliminary" and "may" did not become the brief's "is."
• Report's recommendations vs. yours are separatedanything inferred is labeled, not laundered into the document's voice.
• The caveats section existsa brief with no limitations means the prompt failed or the report is unusual; find out which.
• Whole document processedpage count in, page count acknowledged; no silent truncation.