learn
intermediate
Claude
ChatGPT
Gemini
Multi-platform
How to Use AI to Summarise Meetings, Emails, and Documents Without Missing What Matters
Practical prompts and workflows for summarising meetings, email threads, and long documents so you actually retain the important bits.
7 min read21 February 2026
How to get genuinely useful summaries from meetings, email chains, and documents using AI
Specific prompt structures that extract decisions and action items, not just generic overviews
Workflows for different scenarios: live meetings, recorded calls, email threads, and long PDFs
You'll walk away with copy-paste prompts that work immediately
Why This Matters
You sat through a 45-minute meeting. Three things were actually decided. Two action items were assigned. The rest was preamble, repetition, and someone's screen share not working for five minutes. Now you need to tell your team what happened.
Most people either write nothing (and forget the decisions by Thursday) or spend 20 minutes writing notes that nobody reads. AI can do this in about 30 seconds - but only if you prompt it correctly. The default "summarise this" instruction gives you a useless overview that captures everything and emphasises nothing.
This is a bigger problem than it sounds, especially in organisations spread across time zones. In most Asian companies I work with, there's a meeting layer happening in English, a secondary discussion in the local language, and decisions that get communicated through WhatsApp or LINE after the meeting ends. Capturing and distributing what was actually decided - clearly, quickly, in a format people can scan - is the real productivity gain here.
---
Most people either write nothing (and forget the decisions by Thursday) or spend 20 minutes writing notes that nobody reads. AI can do this in about 30 seconds - but only if you prompt it correctly. The default "summarise this" instruction gives you a useless overview that captures everything and emphasises nothing.
This is a bigger problem than it sounds, especially in organisations spread across time zones. In most Asian companies I work with, there's a meeting layer happening in English, a secondary discussion in the local language, and decisions that get communicated through WhatsApp or LINE after the meeting ends. Capturing and distributing what was actually decided - clearly, quickly, in a format people can scan - is the real productivity gain here.
---
How to Do It
1
1. Choose your input method.
The approach depends on what you're starting with. Live meeting with no recording? You need real-time notes (even rough bullet points) to feed AI afterwards. Recorded Zoom/Teams/Google Meet call? You need the transcript. Email thread? Copy-paste the entire chain. Long document or PDF? Upload it directly.
The quality of your summary is directly proportional to the quality of your input. A messy transcript with speaker labels will give you a far better summary than your half-remembered notes from a call you were multitasking through.
2. Use a structured summary prompt, not "summarise this."
The word "summarise" is doing zero work for you. It tells AI to compress everything equally, which means the important decisions get the same weight as the five minutes someone spent talking about their weekend. Instead, tell AI exactly what you need extracted.
The magic structure is: decisions made, action items with owners, open questions, and key context. That's it. Four categories. Everything that matters in a meeting falls into one of them.
3. Specify your output format.
Do you need a Slack message for your team? A formal set of minutes for a client? A quick bullet list for your own reference? Tell AI the format and audience. "Summarise this for a Slack message to my team of six - they weren't on the call and need to know what was decided and what they need to do" produces dramatically different output than "summarise this meeting."
4. Handle multi-language and cross-cultural context.
If your meeting included discussion in multiple languages (common in pan-Asian teams), note this in your prompt. "The transcript includes some discussion in Mandarin that wasn't translated - flag any sections where the English transcript seems incomplete or where decisions may have been made in a non-English segment." This prevents AI from silently skipping untranslated sections.
5. Create a recurring template for regular meetings.
If you have a weekly standup, monthly review, or quarterly planning session, build a summary template once and reuse it. Include the standing agenda items, the format your team expects, and any specific sections (e.g. "always include: revenue numbers mentioned, hiring updates, blockers raised"). Save this prompt and paste it at the start of every summary session.
---
The approach depends on what you're starting with. Live meeting with no recording? You need real-time notes (even rough bullet points) to feed AI afterwards. Recorded Zoom/Teams/Google Meet call? You need the transcript. Email thread? Copy-paste the entire chain. Long document or PDF? Upload it directly.
The quality of your summary is directly proportional to the quality of your input. A messy transcript with speaker labels will give you a far better summary than your half-remembered notes from a call you were multitasking through.
2. Use a structured summary prompt, not "summarise this."
The word "summarise" is doing zero work for you. It tells AI to compress everything equally, which means the important decisions get the same weight as the five minutes someone spent talking about their weekend. Instead, tell AI exactly what you need extracted.
The magic structure is: decisions made, action items with owners, open questions, and key context. That's it. Four categories. Everything that matters in a meeting falls into one of them.
3. Specify your output format.
Do you need a Slack message for your team? A formal set of minutes for a client? A quick bullet list for your own reference? Tell AI the format and audience. "Summarise this for a Slack message to my team of six - they weren't on the call and need to know what was decided and what they need to do" produces dramatically different output than "summarise this meeting."
4. Handle multi-language and cross-cultural context.
If your meeting included discussion in multiple languages (common in pan-Asian teams), note this in your prompt. "The transcript includes some discussion in Mandarin that wasn't translated - flag any sections where the English transcript seems incomplete or where decisions may have been made in a non-English segment." This prevents AI from silently skipping untranslated sections.
5. Create a recurring template for regular meetings.
If you have a weekly standup, monthly review, or quarterly planning session, build a summary template once and reuse it. Include the standing agenda items, the format your team expects, and any specific sections (e.g. "always include: revenue numbers mentioned, hiring updates, blockers raised"). Save this prompt and paste it at the start of every summary session.
---
What This Actually Looks Like
The Prompt
Here's a transcript from a 40-minute team meeting (sales and operations, 6 people, Thursday weekly sync). Extract and organise into these exact sections: DECISIONS MADE: List every decision that was agreed, even informally. Include who proposed it and whether anyone disagreed. ACTION ITEMS: List every task assigned or volunteered for. Format: [Person] will [action] by [deadline if mentioned]. If no deadline was stated, note "no deadline set." OPEN QUESTIONS: Anything raised but not resolved. Include who raised it. KEY CONTEXT: Important information shared that isn't a decision or action item but that the team should know (numbers, updates, changes, risks mentioned). Format this as a Slack message I can paste into our team channel. Keep it scannable - people will read this on their phones. Total length: under 300 words. [PASTE TRANSCRIPT]
Example output — your results will vary based on your inputs
Example output (your results will vary based on your inputs):
Thursday Sync - 20 Feb Summary
Decisions:
- Pushing the Singtel pilot launch to 12 March (was 1 March). Chris confirmed with the client.
- Dropping the self-serve dashboard from Phase 1 scope. Will revisit in Phase 2 if pilot metrics justify it.
- New reporting cadence: weekly during pilot, switching to fortnightly once stable.
Action items:
- @Adrian: Update the pilot timeline doc and share with Rene by Monday
- @Chris: Send revised SOW to Singtel by Tuesday
- @Joey: Prep the data room for compliance review - no deadline set
- @Michelle: Schedule dry-run presentation with the team before client meeting
Open questions:
- Integration timeline with their existing CRM still unclear. Chris waiting on their tech team's response.
- Budget approval for the additional data processing - Rene to confirm with finance.
Context:
- Singtel's internal restructure might change our day-to-day contact. Chris flagged we should build relationship with the new commercial lead as backup.
- Competitor mentioned in the pitch review last week has since pulled out of the APAC market.
Thursday Sync - 20 Feb Summary
Decisions:
- Pushing the Singtel pilot launch to 12 March (was 1 March). Chris confirmed with the client.
- Dropping the self-serve dashboard from Phase 1 scope. Will revisit in Phase 2 if pilot metrics justify it.
- New reporting cadence: weekly during pilot, switching to fortnightly once stable.
Action items:
- @Adrian: Update the pilot timeline doc and share with Rene by Monday
- @Chris: Send revised SOW to Singtel by Tuesday
- @Joey: Prep the data room for compliance review - no deadline set
- @Michelle: Schedule dry-run presentation with the team before client meeting
Open questions:
- Integration timeline with their existing CRM still unclear. Chris waiting on their tech team's response.
- Budget approval for the additional data processing - Rene to confirm with finance.
Context:
- Singtel's internal restructure might change our day-to-day contact. Chris flagged we should build relationship with the new commercial lead as backup.
- Competitor mentioned in the pitch review last week has since pulled out of the APAC market.
How to Edit This
That's actually close to ready. I'd remove the "@" symbols if my Slack channel would actually ping people (use names instead), and I'd add a one-line note at the top: "Key thing: pilot moved to 12 March. Revised SOW going to client Tuesday." That way people who only read the first line still get the most important update.
The whole process - paste transcript, run prompt, light edit - takes about 2 minutes. The meeting was 40 minutes. That's a good ratio.
---
The whole process - paste transcript, run prompt, light edit - takes about 2 minutes. The meeting was 40 minutes. That's a good ratio.
---
Prompts to Try
Meeting Summary (Structured)
Summarise this meeting transcript into four sections: DECISIONS MADE (include who proposed each), ACTION ITEMS (format: [Person] will [action] by [deadline or "no deadline set"]), OPEN QUESTIONS (who raised each), KEY CONTEXT (important info shared). Keep total summary under 300 words. Format for [Slack/email/document]. [PASTE TRANSCRIPT]
What to expect: A clean, scannable summary organised by the four categories. If the transcript is messy or speakers aren't labelled, AI will do its best to attribute but may get names wrong - double-check attribution.
Email Thread Digest
Here's an email thread with [X] messages. Extract: the final decision or current status, any outstanding asks that haven't been answered, and the key points of disagreement if any. Ignore pleasantries and email signatures. Present as a 3-5 bullet summary I can forward to someone who needs to get up to speed in 30 seconds. [PASTE EMAIL THREAD]
What to expect: A tight bullet summary that cuts through the reply-all noise. Particularly useful for those 15-message threads where the actual decision is buried in message 11.
Document Key Points
Read this document and extract: the 5 most important points (ranked by significance, not by order of appearance), any recommendations or calls to action, any data points or statistics cited, and anything that contradicts or updates what I might already know about this topic. Be opinionated about what matters most - don't just list everything equally. [PASTE OR UPLOAD DOCUMENT]
What to expect: A prioritised summary that emphasises what's important rather than compressing everything equally. The "be opinionated" instruction is key - without it, AI treats every paragraph as equally worthy of inclusion.
Common Mistakes
Using "summarise this" with no structure.
This is the number one mistake. AI will compress everything equally, giving you a bland overview that misses the specific decisions and action items you actually need. Always specify what categories of information you want extracted.
Not specifying the audience and format.
A summary for your own reference, a Slack message for your team, and formal minutes for a client are three completely different documents. Tell AI who's reading it and how.
Trusting speaker attribution blindly.
Auto-generated transcripts often mislabel speakers, especially in calls with more than 4 people. Always verify that AI attributed the right decisions and action items to the right people before distributing.
Summarising without reading.
If you send out an AI summary of a meeting you attended without reviewing it first, you will eventually distribute something wrong. AI misses subtext, sarcasm, and the difference between someone suggesting something and someone agreeing to something. Skim the summary against your memory before sharing.
Over-summarising long documents.
For a 50-page report, don't ask for a one-paragraph summary. Ask for a one-page summary, then ask follow-up questions about the sections that matter most to you. Layered summarisation beats aggressive compression.
---
---
Tools That Work for This
Claude- Handles long transcripts and documents well thanks to its large context window. Good at following structured extraction instructions.
⚠ no native integration with meeting platforms, so you need to paste transcripts manually.
ChatGPT- Similar quality for summarisation. The mobile app is useful for quick email summaries on the go.
⚠ context window is smaller for non-Pro users, which can be a problem with very long transcripts.
Gemini in Google Workspace- If your organisation uses Google Meet and Google Docs, Gemini's native integration means summaries can be generated without copy-pasting transcripts.
⚠ summary quality is decent but less customisable than dedicated prompting in Claude or ChatGPT.
Otter.ai- Purpose-built for meeting transcription and summarisation. Integrates with Zoom, Meet, and Teams. Good for teams that want automated summaries without manual prompting.
⚠ the AI summaries are less customisable than what you get with your own prompts.
Fireflies.ai- Another strong meeting assistant, popular with sales teams across Southeast Asia. Captures and summarises calls automatically.
---
⚠ free tier is quite limited.
Frequently Asked Questions
Yes, but quality varies. Claude and ChatGPT handle Mandarin, Japanese, Korean, Bahasa, Thai, and Vietnamese with decent accuracy. For mixed-language meetings, specify in your prompt which languages were used and whether you want the summary in English or the original language. Summarisation quality for non-English content is improving rapidly but isn't yet as reliable as English.
Be thoughtful about what you paste into AI tools. Most cloud-based AI services process your data on external servers. For sensitive board discussions, HR matters, or confidential deal negotiations, check your organisation's AI usage policy first. Claude and ChatGPT both offer enterprise tiers with stronger data handling commitments. When in doubt, summarise confidential meetings manually.
Partially. Tools like Otter.ai and Fireflies.ai can auto-generate summaries for every meeting. For email and document summarisation, you'd need to build a workflow using Zapier or Make.com connecting your inbox to an AI API. It's doable but requires setup. For most people, the copy-paste-prompt approach is faster to start with.
For well-transcribed meetings with clear speaker labels, accuracy is high - around 90-95% for factual content. The main failure mode is misattributing who said what. AI also occasionally lists something as a "decision" that was actually just a suggestion being discussed. Always review before distributing.
---
---
Next Steps
Start with your next meeting. Record it (or take rough notes), run the structured summary prompt, and compare the output against what you remember. You'll immediately see where AI adds value (extracting action items from a rambling discussion) and where it needs checking (speaker attribution, distinguishing decisions from suggestions).
Once you've validated the approach for meetings, try it on the longest email thread in your inbox right now. The Email Thread Digest prompt will save you 10 minutes of scrolling and re-reading.
Related guides: [INTERNAL LINK: how to automate weekly reporting with AI] and [INTERNAL LINK: how to build an AI personal assistant that actually works] both build on the summarisation workflows covered here.
Once you've validated the approach for meetings, try it on the longest email thread in your inbox right now. The Email Thread Digest prompt will save you 10 minutes of scrolling and re-reading.
Related guides: [INTERNAL LINK: how to automate weekly reporting with AI] and [INTERNAL LINK: how to build an AI personal assistant that actually works] both build on the summarisation workflows covered here.