Decisions, Actions, Blockers: The Three-Column Prompt That Replaces Your Status Update

Lesson 7 of a 10-part series on AI for product managers argues that status updates are a mechanical transform on notes you already have. Feed a meeting transcript into one prompt, ask for decisions made, actions with owners, and blockers, then re-cut that same structured output for engineering, executives, and your weekly email. The slide on screen shows the target format: three columns labeled DECISIONS, ACTIONS, BLOCKERS, each with a one-line claim and a one-line qualifier ("We're shipping without social login. Unblocks launch. Can revisit in Q2." / "Eng updates auth docs by Friday. Owner: Alex. No review needed." / "Legal hasn't signed off on ToS. Need response by EOD Thursday."). The workflow holds up. The parts the video skips are where it breaks.

The prompt works because it asks for a schema, not a summary

"Summarize this meeting" returns prose that tracks the order people spoke in. Asking for decisions, actions with owners, and blockers returns three buckets that a reader can scan for the one thing they need. The structure also makes omissions visible: an empty blockers column is a claim you can check, whereas a summary that simply never mentions blockers looks complete.

This is a RAID log with the risks column dropped, which is what most PM status reporting already was before anyone typed it into a model. The ACM study of an LLM-powered meeting recap system at Microsoft found the same split usefulness, with a short highlights view and a structured hierarchical minutes view serving different needs rather than one format winning (Asthana et al., 2023). The video's three columns are the highlights view. Keep the raw transcript reachable, because someone will ask what actually got said.

The qualifier lines in the on-screen example carry more weight than the headlines. "Owner: Alex. No review needed." tells a reader they can stop thinking about it. "Need response by EOD Thursday" converts a blocker into a request with a deadline attached. If your generated output is three bare bullets with no owner and no date, the prompt returned a summary and you shipped it anyway.

One source, three outputs is where narratives quietly diverge

The video's framing is that engineering wants technical blockers and the executive team wants business risk, so you generate both from the same base. That part is fine. The failure mode is that the two documents stop agreeing.

A model asked to write the exec version will compress "legal hasn't signed off on ToS" into something about launch timing risk. Asked for the engineering version, it keeps the ToS detail and drops the launch framing. Both are accurate. Neither one alone tells a VP that the engineering team is blocked on legal, or tells an engineer that leadership has already absorbed a slip. Six weeks later, two groups have two different mental models of the same project and neither has a reason to suspect it.

The cheap fix is to keep the base artifact as the source of record, link both derived versions to it, and never edit a derived version without editing the base. The video treats the base as a scratch intermediate. Treat it as the document.

The transcript is the weakest link in the chain

Every output here inherits the errors of the input. Automatic transcription is not a clean pipe. Research on OpenAI's Whisper found that around 1 percent of transcribed audio segments contained entire hallucinated phrases that appeared nowhere in the source audio, and that hallucination rates were higher for speakers with aphasia (Koenecke et al., FAccT 2024). Speaker attribution is a separate error source, and a mislabeled speaker in a meeting recording turns directly into a wrong owner in the ACTIONS column.

So the two fields most worth checking by hand are the two the video never mentions: names and dates. A hallucinated adjective in a summary costs you nothing. "Owner: Alex" when Alex never agreed to it costs you Alex.

The read of the room is the part the video gets right

The strongest line in the video is that AI writes the update and you decide what it means. The subtext of what did not get said is not recoverable from a transcript, because it is an absence. A director who went quiet when the date slipped, a decision that was technically made but that nobody in the room believes, an action item accepted with no intention of doing it: none of that is in the text.

The video presents this as a thing you withhold. It is closer to the thing you act on. The generated update is what you send. The read of the room is what tells you which one-on-one to schedule before you send it.

What the video leaves out: the update was doing work

Writing a status update forces a weekly pass over your own project. You notice the item that has been "in progress" for three weeks because you had to type it again. Automate the writing and you keep the artifact and lose the pass.

The cheap replacement is to read the generated update against last week's before sending. Anything that appears in both columns unchanged is either stalled or badly worded, and both are worth five minutes.

One more caveat on the video's own numbers. The claim that the average PM writes the same status update 50 times a year is not sourced anywhere in the video or its description, and it reads as a stand-in for "weekly." It works as a hook. Do not repeat it as a finding. The on-screen slide also labels this Lesson 06 of 10 while the narration says lesson 7 of 10, so the series numbering is worth checking before you go looking for the next one.

Key Takeaways

• Ask for a schema (decisions, actions with owners, blockers) instead of a summary. The buckets make gaps visible.
• Keep the structured base as the source of record and derive the engineering and exec versions from it, so the two audiences never drift onto different stories.
• Hand-verify owners and dates. Transcription errors and speaker misattribution land directly in the fields that assign work.
• Include the qualifier line, not just the claim. "Owner: Alex, no review needed" is what lets a reader stop tracking an item.
• Read this week's generated update against last week's. Unchanged items are stalled items.
• The 50-updates-a-year figure in the hook is unsourced. Treat it as rhetoric.

Resources

• Summaries, Highlights, and Action items: Design, implementation and evaluation of an LLM-powered meeting recap system. Asthana, Hilleli, He, and Halfaker on how Microsoft employees actually used LLM meeting recaps, and why a highlights view and a structured minutes view serve different needs.
• Careless Whisper: Speech-to-Text Hallucination Harms. Koenecke et al. at ACM FAccT 2024, measuring fabricated phrases in Whisper transcriptions. The reason to verify names and dates by hand.
• Careless Whisper preprint (arXiv). Open-access version of the same paper.
• vibe.academy. The publisher behind this lesson series, positioned as AI training for non-technical professionals.