Summarization

After a meeting is transcribed, click Summarize in the Meeting Page panel. Pensum sends the transcript (plus the meeting’s attendees and project, if set in frontmatter) to the AI and inserts a structured summary into the meeting note above your existing content.

The summary is reference material, not the meeting’s record of truth — your notes during the meeting and the extracted tasks are. The summary is for the times you (or someone else who attended) needs a quick recap without re-reading the full transcript.

How to run it

1. Record the meeting (Cmd+Shift+R) and transcribe it
2. Open the meeting note (or its Meeting Page sidebar panel)
3. Click Summarize

If you’ve set attendees: and/or project: in the meeting frontmatter, the AI uses them for context (it’ll name attendees correctly and frame the summary around the project’s goals if known).

What you get

A structured [!ai-generated] callout inserted just below the meeting’s H1 heading:

# Q3 pricing review

> [!ai-generated] Summary by claude-sonnet-4-6, 2026-05-16
>
> **Overview**
>
> The team reviewed the proposed Q3 pricing changes. The headline shift is a tiered structure for SMB customers, with enterprise pricing largely unchanged. Marketing has eight weeks of runway before the public announcement.
>
> **Key topics**
> - SMB tier structure (three tiers, $9/$29/$99 monthly)
> - Enterprise pricing — staying with custom quotes
> - Migration plan for existing customers on grandfathered plans
> - Marketing rollout timeline
>
> **Decisions**
> - Move to three-tier SMB pricing effective Aug 1
> - Grandfather existing customers for 12 months
> - Public announcement targeted for mid-July
>
> **Open questions**
> - How to handle annual subscribers mid-cycle (Sam to investigate)
> - Whether to launch with a promo or hold (Jamie to discuss with marketing)

The four sections (Overview / Key topics / Decisions / Open questions) are the template. Sections come up empty when the meeting didn’t produce them (e.g. brainstorming meetings often have lots of topics and few decisions).

The frontmatter field ai_summary_status: none flips to ai_summary_status: draft after summarization.

Why the callout style

• Visually distinguishable from your own writing — see provenance
• Survives in any markdown reader (Obsidian’s callouts render as styled blockquotes elsewhere, so the structure isn’t lost outside Pensum)
• Doesn’t touch your notes — the callout goes above your existing content, never replacing or rearranging what you wrote during the meeting
• Searchable — the callout text is part of the note, so full-text search finds it

Reviewing the draft

The summary is a draft. You decide what happens next:

• Keep as-is — leave the callout in place. It stays marked as AI-generated, which is the correct provenance.
• Edit inline — fix names, add context the AI missed, correct misframings. Once you edit the content, provenance flips to ai-suggested-edited in the index (the callout label itself doesn’t change automatically — feel free to update the label too if you want).
• Move to your own section — if a particular bullet (e.g. a decision) is something you want to track separately, copy it to your own ## Decisions section and delete the corresponding line from the callout.
• Re-run — click Summarize again to get a new draft. The previous draft is replaced.
• Remove — delete the callout entirely. ai_summary_status stays at draft until you re-run or manually set it back to none.

What the AI is told to do

The summarisation prompt is conservative:

• Only what’s in the transcript. No outside knowledge, no inferences beyond what was said. If someone said “we’ll figure it out next week,” that’s an open question, not a decision.
• Concise. Each section capped at a few bullets. Overview capped at ~150 words.
• Decisions are explicit only. A decision means someone said “let’s do X” and others agreed. Hedged statements (“we should probably do X eventually”) go to open questions, not decisions.
• No inventing attendees. When attendees are set in frontmatter, the AI uses those names. When they aren’t, the AI uses generic references (“the team”, “someone”) rather than guessing names.

The goal: summaries you can trust. Better to have a sparser summary than one that puts opinions in people’s mouths.

Speaker identification helps

If you ran Identify Speakers before summarizing, the transcript has named speakers (Jamie, Sam, Taylor) instead of Speaker 0/1/2. The summary will use those names where attribution matters (“Sam committed to following up with legal” rather than “someone will follow up with legal”).

For this to work:

1. Populate attendees: in the meeting frontmatter
2. Click Identify Speakers in the Meeting Page panel
3. Click Summarize

Model

Summarization defaults to a substantive model (Claude Sonnet 4.6 in the catalog) because the quality difference between flagship and mini models shows up most for synthesis tasks. Change which model under Settings → AI → Meeting Summary.

Cost for Pro BYO users: a few cents per meeting on Claude Sonnet — most meetings cost less than a coffee. Pro All-in-One users have this included.

Pairs with task extraction

Summarize answers “what happened?” — Extract Action Items answers “what’s now on someone’s plate?” They’re complementary. Most users run both after every meeting:

1. Transcribe
2. Identify speakers
3. Summarize (gives you the recap)
4. Extract action items (gives you the tasks)

See task extraction for the second half.

What summarization isn’t

• A meeting minutes generator. It doesn’t reproduce the full sequence of who said what — that’s what the transcript is for. The summary is the executive view.
• A decision-making tool. It surfaces what was decided, not what should be decided. If the meeting didn’t reach a decision, the summary won’t manufacture one.
• A replacement for paying attention. A summary of a meeting you didn’t attend gives you the shape; the nuance is in the transcript and in your own follow-ups with attendees.