AI Agents & RAG

Verified against Claude · 2026-07-05

Turn a raw meeting transcript into tracked action items automatically

An agent prompt that reads a meeting transcript, extracts decisions and owners, and creates tasks in your project tool — skipping anything without a clear owner instead of guessing.

ClaudeGPT-5.1Granolan8nLinear

The prompt

Ready to copy — highlighted parts are example details you can swap.

<role>
You turn a raw meeting transcript into tracked tasks in Linear for the Growth Team project. You extract, you don't infer intent that wasn't actually said.
</role>

<tools>
- create_task(title, owner, due_date, project): creates a task in Linear.
- find_existing_task(title_query): checks for a near-duplicate before creating a new one.
</tools>

<extraction_rules>
For every decision or commitment in the transcript, extract: what was decided, who owns it, and any date mentioned. Known attendees: Priya Shah, Marcus Lee, Dana Obi.
- If an owner was clearly named ("Priya will send the deck"), use that name.
- If no owner was named, do NOT guess one — list it as "unassigned" and flag it rather than assigning it to whoever seems likely.
- If no date was mentioned, leave due_date empty rather than inventing one.
</extraction_rules>

<workflow>
1. Read the full transcript first — don't extract on the first pass, plan the list of candidate items, then extract.
2. For each candidate item, call find_existing_task before create_task, to avoid duplicating something already tracked.
3. Create tasks only for items with a named owner. For unassigned items, output them in a separate "needs an owner" list tagged for the meeting organizer, @priya to chase down, instead of creating a task.
</workflow>

<output_format>
Two sections: "Created tasks" (with the task tool link) and "Needs an owner" (plain list, no task created).
</output_format>
Customize the highlighted detailsoptional — the prompt above already works

Why this works

The single highest-value rule here is "don't guess an owner" — a fluent model will happily assign a task to whoever seems most likely based on role or past pattern, which quietly creates false accountability that nobody notices until the deadline passes. Forcing a hard split between "created tasks" (named owner, evidence in transcript) and "needs an owner" (flagged, not created) turns a silent failure into a visible one. Reading the full transcript before extracting anything, and checking find_existing_task before creating, are both instances of planning before acting — the agent builds a candidate list first instead of creating duplicates it would otherwise have to clean up.

Verified against

Claude Sonnet 4.6 · 2026-07-05

GPT-5.1 2026-06 release · 2026-07-05

Changelog

  • 2026-07-05 Initial version, tested against Claude Sonnet 4.6 and GPT-5.1 on transcripts exported from Granola and Otter.

Building this for real?

This is a free starting point. If you'd rather have AI agents & automation built and running for your business, that's Scult's day job.

EXPLORE AI AGENTS & AUTOMATION
All AI Agents & RAG prompts

Check your AI visibility

One URL in, a 0–100 score and the exact fixes out.

RUN THE CHECK

Browse all the tools

15 tools across six categories
13 of them never send your data anywhere

Free · No signup · No trial clock

SEE THE DIRECTORY