Most professionals spend significant time processing information that contains two important points buried in forty minutes of context-setting. AI handles consolidation well. This guide shows you how to get a useful summary, not a padded one, and how to check it before you trust it.
Works for: strategy documents, meeting transcripts, long email threads, policy PDFs, research reports.
The tool, and how to set it up
Use Claude (claude.ai or the API). For most documents, the standard interface handles it fine. For a 90-minute meeting transcript or a 40-message email thread, switch to a model with a larger context window: Claude Sonnet or Opus holds the entire document in memory and reasons across it, rather than working in chunks.
Two ways to get the content in:
- Paste it. Copy the text directly into the chat. Works for email threads, transcripts exported from Otter or Teams, copied web articles. Quick and sufficient for most cases.
- Upload the file. Drag a PDF or Word document into the conversation. Use this for formatted documents where the structure matters, or where copying would lose key formatting (tables, section headers).
Do not use a consumer AI tool for anything containing client names, personal data, commercially sensitive figures, or legal content. More on this below.
How to do it
-
Open a new Claude conversation. Paste or upload your document.
-
Start with the structured summary, not a blob. Ask for specific categories rather than "summarise this." The difference in output quality is significant. At minimum: key decisions made, action items with owners, open questions, and the one thing that would change how you'd act if you missed it.
-
For a long email thread (30 or more messages), include an instruction to work chronologically and flag where the position or consensus changed. Email threads often contain reversals and contradictions; a summary that only reflects the last message misses the argument.
-
For a meeting transcript, ask Claude to ignore pleasantries and filler. Transcripts contain a lot of "great, yeah, absolutely" that pads the output. A good instruction is: "Work from this transcript. Ignore filler and off-topic chat. Summarise only what was decided, assigned or left unresolved."
-
Iterate down. Ask for a compressed version: "Now give me a 3-bullet version for my manager" or "Give me the one paragraph I'd send to someone who missed this meeting." The compressed version is where the AI occasionally drops something important, so compare it against the full version before treating it as final.
-
If a section needs more depth, ask a follow-up: "What was the reasoning behind the decision on budget?" Claude has the full document still in context, so it can pull specifics without you re-pasting anything.
The prompt
Copy and adapt this for most summarisation tasks:
I'm going to paste a [document / email thread / meeting transcript] below. When you've read it, produce a structured summary using these headings:
**Decisions made**: list each confirmed decision with any relevant context
**Action items**: who is doing what, and by when (if stated)
**Open questions**: things that were raised but not resolved
**Risks or concerns**: anything flagged as a potential problem
**The one thing**: if I only remembered one point from this, what should it be?
Be specific. Use names and figures from the source where they appear. Do not invent details that aren't present. If something is ambiguous, say so rather than resolving it.
[Paste document here]
Adjust the headings for the context. A strategy document might swap "Action items" for "Recommended next steps." A policy document might not have action items at all.
How to QA it
Read the summary against the original source before acting on it. This takes five minutes, not thirty, because you are checking rather than reading cold.
Check these specifically:
- Names and roles. AI occasionally swaps who said what, or attributes a decision to the wrong person. Verify any name that appears in an action item.
- Numbers. Figures, dates, percentages and deadlines are the most commonly mangled specifics. If the transcript says "£40,000 by Q3" and the summary says "£40,000 by end of year," that matters.
- Caveats and conditionals. The classic failure is a confident summary that drops the qualifier. "We agreed to proceed, subject to board sign-off" becomes "We agreed to proceed." The missing clause is the one that affects what you do next.
- What's absent. Skim the original for topics you remember being significant. If they're not in the summary, either they were less significant than you thought, or they got dropped. Either way, worth knowing.
For the compressed version (the three-bullet or one-paragraph version), compare it directly against the full summary rather than going back to the source each time.
How to stay safe
Do not paste client-identifiable information, personal data, financial figures tied to named individuals, or legally binding content into a consumer AI tool. That includes Claude.ai if your organisation has not set up an enterprise agreement. The text you paste into a consumer tool may be used for model improvement.
For sensitive material: use your organisation's enterprise AI deployment (Copilot via M365, Claude for Enterprise, or similar), work with a redacted version, or summarise it yourself.
A summary is a starting point for your judgement. If a contract, a redundancy decision, or a regulatory document is involved, you read the source. AI can help you locate the relevant section faster, but the reading is still yours to do.
Start with something low-stakes: your own meeting notes from last week, or a long article you've been meaning to read. Get a feel for what the output looks like before you rely on it for anything that matters.