The 5-Section Meeting Notes Template That AI Actually Fills Out Correctly

Published October 13, 2026

October 13, 2026 · 10 min read

If you have ever opened an AI-generated meeting summary and found a tidy paragraph of prose but no clear action items — or worse, action items assigned to the wrong person — you have run into the single hardest problem in AI note-taking. The fix is not a smarter model. It is a template: a five-section structure that forces the AI to populate specific slots instead of writing creative essays.

The template is Context, Decisions, Action Items, Open Questions, Notes. In 2026, this is the structure that practitioner playbooks have converged on, and it also happens to be the structure that modern LLMs fill in most reliably. Below we explain why the format matters, show the copy-paste template, break down the prompt engineering, and walk through the on-device workflow that keeps your transcripts off a vendor server.

Why AI meeting notes miss action items in the first place

Before we get to the template, it helps to understand what breaks. The analysis from Circleback's engineering team is unusually candid: AI meeting notes involve four pipeline stages — audio preprocessing, speech-to-text, speaker diarization, and language-model summarization — and errors at any stage compound through every stage after it. Leading transcription models hit under 3% word error rate on clean audio, but real meetings with overlapping speakers and background noise push error rates to 12% or higher. State-of-the-art speaker diarization still has 11–13% error rates, mostly from crosstalk.

Then comes the hardest part. Action-item extraction, as a 2026 analysis of 17 note-taking tools explains, requires three simultaneous inferences: intent (was something actually committed to?), agency (who owns it?), and temporality (by when?). Most tools treat action-item extraction as a secondary inference layered on top of summarization rather than a first-class structural requirement. The result is a summary that sounds authoritative but omits the exact elements that drive execution.

The five-section template fixes this by making each inference a slot the model has to fill. Instead of "summarize the meeting," the prompt becomes "populate these five fields," and the model's output format is validated against the schema.

The 5-section template, in full

This is the template. Copy it into a Notion page, a Markdown file, or your AI tool's custom-prompt field.

# Meeting: [Title]
Date: [YYYY-MM-DD]  |  Attendees: [names]

## 1. Context
[2–3 sentences: why we met, what prompted it, what we were trying to decide]

## 2. Decisions
- [Decision 1 — in bold, with the owner who proposed it]
- [Decision 2]

## 3. Action Items
- [Owner] to [verb] [object] by [YYYY-MM-DD]
- [Owner] to [verb] [object] by [YYYY-MM-DD]

## 4. Open Questions
- [Question we did not answer, parked for next time]

## 5. Notes
[Everything else: raw observations, references, side threads]

Each section does distinct work, and that distinction is what makes AI extraction reliable:

• Context grounds the model so it does not hallucinate background. It also helps absent colleagues get up to speed in two minutes.
• Decisions are stated outcomes, formatted in bold so they stand out. This matches the 2026 best-practices guidance to distinguish discussion, decisions, and action items with different formatting.
• Action Items use a non-negotiable schema: Owner, Verb, Object, Deadline. The Atlas practitioner playbook captures the format succinctly as "Owner Verb Object by Date," with a worked example: "Sarah to draft the onboarding spec by 2026-05-09."
• Open Questions catches items that were raised but not resolved. These are the things that most often fall through the cracks.
• Notes is the catch-all. The AI should keep the full transcript here as a reference anyone can search later — a practice Maestra's 2026 best-practices guide recommends so that any dispute about a detail can be settled against source material.

The Owner-Verb-Object-by-Date rule, explained

Every action item has exactly three fields: a single named owner (not a team), an action verb attached to a concrete deliverable, and an ISO-formatted deadline. The April 2026 Spinach guide estimates that 44% of action items fail without this structure, because "the team" cannot be held accountable and "soon" never arrives. Enforce the three-field discipline and your downstream integrations — Jira, Linear, Apple Reminders — have everything they need to create a real task.

Why this specific structure works for AI (and not others)

LLMs do not reliably invent structure from unstructured prompts. They do reliably fill structure when given one. The Brass Transcripts prompt breakdown makes the point clearly: a prompt succeeds because it provides explicit structure, telling the AI exactly what information to extract and how to organize it, so "each section heading becomes a search directive." Give the model a schema and it searches for schema-fitting evidence. Give it a vague "summarize this," and it will reach for the fluent-but-generic patterns in its training data.

The accuracy numbers back this up. The arXiv paper on action-item-driven meeting summarization fine-tuned a BERT classifier on 2,750 labeled dialogue statements and hit 95.4% classification accuracy for whether a statement contained an action item. Independent LLM pipelines using structured JSON output and Pydantic validation reach F1 scores of 0.96 for the combined task+owner+deadline extraction on clean synthetic transcripts, per a published evaluation on GitHub. The common thread: constrain the output format.

The copy-paste prompt that fills the template

Here is the prompt that pairs with the five-section template. Paste it with your transcript, or wire it into your AI tool's custom-template field:

You are a meeting notes assistant. Read the transcript below and
populate the following five sections. Do not invent content.

1. Context — 2–3 sentences on why the meeting happened.
2. Decisions — bullet list, bolded. Only include decisions that
   were explicitly agreed to, with the owner who proposed each.
3. Action Items — bullet list in the exact format:
   "[Owner] to [verb] [object] by [YYYY-MM-DD]"
   If owner is missing, write "[UNASSIGNED]".
   If deadline is missing, write "[NO DEADLINE]".
4. Open Questions — bullet list of things raised but not resolved.
5. Notes — anything else material that does not fit above.

Use dual-prompt chaining: first identify all explicit commitments,
conditional assignments, and unresolved decisions. Then populate
the five sections above, excluding inferred commitments.

Transcript:
[paste transcript here]

Two things in that prompt are worth highlighting. First, the [UNASSIGNED] and [NO DEADLINE] tags are explicit uncertainty markers, which the anti-hallucination playbook recommends to prevent the model from quietly inferring an owner it does not actually have evidence for. Second, the dual-prompt chaining — identify commitments first, summarize second — isolates the "what happens next" inference from the "what was discussed" inference, which prevents the two from bleeding into each other.

A worked example: raw transcript to filled template

Here is a short transcript and what the five-section template looks like when filled in correctly.

Transcript (fragment):

Maya: We will keep the current checkout for the October pilot.
Leon: I will send the revised onboarding checklist tomorrow.
Ravi: Someone should check whether the export includes
      cancelled orders.

Filled template:

# Meeting: Checkout Pilot Planning
Date: 2026-09-30  |  Attendees: Maya, Leon, Ravi

## 1. Context
Aligning on scope for the October checkout pilot and
confirming onboarding materials.

## 2. Decisions
- **Keep the current checkout for the October pilot** (Maya proposed)

## 3. Action Items
- Leon to send the revised onboarding checklist by 2026-10-01
- [UNASSIGNED] to verify export includes cancelled orders by [NO DEADLINE]

## 4. Open Questions
- Who owns the cancelled-orders export verification?

## 5. Notes
Ravi raised the cancelled-orders export question but no
owner was assigned in the meeting.

Notice what the model did and did not do. It did not guess that Ravi owned the export-verification task just because he mentioned it — a classic hallucination described in evaluations of "We need to..." phrasing. It left the owner unassigned and surfaced the ownership gap in Open Questions. That is the behavior you want.

Template vs. no-template: a comparison

Here is what changes when you move from freeform AI summaries to the five-section template.

Why the recorder you use matters as much as the template

A template is only as good as the transcript it is populated from. If the transcription stage drops words, misattributes speakers, or sends your audio to a vendor server you cannot audit, the quality of the final notes is capped by the weakest link in the chain. This is also where the privacy question enters: when every sentence of a client call, internal strategy session, or compensation discussion is uploaded to a cloud transcription service, you have created a copy of that conversation outside your organization's control.

The alternative is a bot-free, on-device recorder that captures system audio locally and runs transcription on the same device. On modern Apple hardware this is now genuinely practical because of the Neural Engine and new on-device models.

What changed in 2025–2026: Apple's SpeechAnalyzer

At WWDC 2025, Apple introduced SpeechAnalyzer, a modern on-device speech recognition framework with a new proprietary Apple model. According to Argmax's benchmarks, Apple matches the speed and accuracy of mid-tier OpenAI Whisper models on long-form conversational speech transcription — all without the audio leaving the device. iOS 26 ships three modules on top of SpeechAnalyzer: SpeechTranscriber for long-form audio, DictationTranscriber for short utterances, and SpeechDetector for voice activity detection, as documented in a developer-focused breakdown of the new framework.

What this means in practice: a Mac or iPhone from the last few years can now transcribe a two-hour meeting locally, at or near real time, without any upload. The Yaps analysis of Apple Silicon captures the shift: the dedicated Neural Engine and unified memory finally make local speech recognition fast, private, and accurate at the same time — a combination that previously required picking two.

How Basil AI handles the five-section template

Basil AI is built around exactly this workflow: capture locally, structure locally, deliver locally. Three concrete technical differentiators:

1. Neural Engine transcription. Basil uses Apple's on-device Speech framework (now SpeechAnalyzer on iOS 26+) to transcribe audio directly on the device. No audio file is ever uploaded. See our technical deep dive on local processing for the architecture.
2. On-device summarization into the five-section template. After the transcript is produced, Basil's summarization step runs locally and populates the Context / Decisions / Action Items / Open Questions / Notes slots. Action items use the Owner-Verb-Object-by-Date format, and missing owners or deadlines are flagged rather than invented.
3. Zero egress. There is no vendor server holding your recording, no DPA to negotiate, no subpoena surface area, and no "we may use your data to improve the service" clause. Compare this to the cloud-first competitors documented in our AI meeting notetaker comparison guide.

This is not a compliance claim — compliance is always the customer's determination for their own regulatory context. It is an architectural fact: when transcription and summarization happen on your Mac or iPhone, no third party holds a copy of your meeting.

Common pitfalls when adopting a template

Even with the right template, teams trip over the same handful of issues. A few to watch for:

• Verbatim transcripts instead of structured minutes. As the Atlas playbook notes, the most effective approach is structured rather than verbatim — skip verbatim transcription unless you are a court reporter or compliance officer. Keep the full transcript in Notes as a fallback.
• Over-trusting LLM confidence scores. As published evaluations show, LLM confidence is overconfident (often 1.0). Treat confidence as a ranking signal, not a probability. Use validation rules (missing owner, unresolved date) to downgrade untrustworthy items.
• Shared ownership. "Engineering to investigate" is not an action item. One name per task, every time.
• Vague verbs. "Follow up" and "look into" are hiding places. Use "draft," "send," "schedule," "approve," "ship."

Share minutes within 24 hours — on-device makes this trivial

Multiple 2026 playbooks converge on the same distribution rule: get minutes out within 24 hours while details are fresh. KenzNote's 2026 best-practices guide is explicit that minutes should be distributed within 24 hours while discussions are still fresh in participants' minds.

With a cloud bot-based tool, that is bottlenecked by the vendor's processing queue. With an on-device pipeline, the summary is typically ready in under a minute of the meeting ending — because there is no upload step and no shared GPU to wait for. In practice, you can leave a one-hour meeting, open your note-taking app, and have a filled-out five-section template to review and share before you have finished your next coffee.

The bottom line

The template is simple. Context, Decisions, Action Items (Owner-Verb-Object-by-Date), Open Questions, Notes. Give that structure to any decent LLM as a prompt — or to a tool built around it — and the quality of your AI meeting notes jumps immediately. Pair it with on-device transcription, and you get the three things that used to be mutually exclusive: structured, accurate, and private.

The industry data on meeting load is sobering — 392 hours per employee per year is a lot of time to leave undocumented or badly documented. The template is a 15-minute change that compounds across every one of those hours.

Try the 5-section template on-device

Basil AI records meetings on your Mac or iPhone, transcribes locally, and generates a structured summary — Context, Decisions, Action Items, Open Questions, Notes — without uploading a single second of audio.