One open secret of product work: most PMs do not manage anyone. They manage a product. They work across design, eng, data, marketing, and support. The "mini-CEO" line is marketing. Nobody reports to that person.
Kevin Niparko, writing the official Grok Bot guide for PMs, used a different sentence: for the first time a PM can have a team that reports to them. The team is agents. Each one has a computer, a lane, and a memory that should not leak into the others.
This article is the Academy rewrite. The copy-paste briefs live on Grok Bot product prompts, chief of staff, engineering, and research.
An attention list, not a priority doc you stop updating
Weekly goals go stale by Tuesday. What you actually clicked, answered, and sat in is a better live view.
His Chief of Staff builds an attention list from Slack, mail, calendar, and meeting notes: the projects you are in, their state, what is next. Agents then use that list as a filter. Archive mail that does not map to it. Surface the incident channel you are already in. Compare the list to the priorities you wrote down on Monday. Drift is a finding, not a vibe.
To stand it up:
Create an attention list. Every hour, read email, Slack, meeting notes, and calendar. Write a short list of projects I am actually in, their current state, and what needs to happen next. Stay quiet if nothing changed. Do not send mail. Do not create calendar events unless I ask.
To use it: "Archive mail that is not on the list." "Where did I spend time this week that I can hand to an agent?"
You still pick the start time when it holds a slot. You still send the external note.
Research that sits in the tools you already have
Product teams are drowning in raw context: call recordings, customer Slack, interview databases, warehouse usage, tickets, Notion decisions. The hard part is making that pile legible to an agent without the agent inventing a quote.
Setup is unglamorous. Connect the plugins. Sign the computer into anything that has no plugin. Tell the research bot that a missing transcript is "missing," not a paraphrase from the title. Our research prompts and product prompts are written that way on purpose.
Shipping, at a higher altitude
Niparko wrote that Grok Bot already accounts for a double-digit share of internal merged PRs on their team. Treat that as his environment, not a promise about yours.
The useful design is the eng manager who does not code. That bot breaks work down, hands it to IC eng agents, and checks the output against the goal. The IC bots spin cloud agents, test, and pass work to each other. You talk to the manager in goals, not in six threads you cannot review.
If a bot can open your flag tool, it can click. Official product-performance use still says do not change production settings. These pages return write-ups. You flip the flag. That rule is in every brief on the product page.
A PM roster that stays referenceable
His names, as jobs:
- Chief of Staff. Calendar, Slack, inbox. The only generalist. Quiet if nothing changed. Owns the attention list.
- Eng manager. Does not code. Delegates. Checks the goal.
- Eng (several). Receive work from the manager. Use cloud agents. Validate.
- Data. Warehouse, charts, the morning dash. Tactical questions you used to leave unanswered because asking a person had too much friction. Hard, taste-heavy questions still go to humans.
- PM sidekick. RFCs, research, feedback, docs. In the work, not summarizing Slack like an intern.
- Recruiter. Loops and names. Recruiting prompts if you steal that seat.
Why not one omniscient agent?
A blank box that "does everything" is the default pitch. He cut the other way:
- Referenceability. You know who does what.
- Parallelism. Discrete work at the same time.
- Scoped memory. The inbox bot should not learn your eval harness. The eval bot should not archive mail.
It is also more pleasant. That is allowed to count.
Four lessons that survive a rewrite
- Named agents, separate memory. Broad mission, narrow memory.
- They learn on the job. "Read my Slack and write a style guide" was the highest-leverage instruction he named.
- They should be quiet. Bias to action is good. A ping on every thought is not. Stay silent unless you are needed.
- Human on the important stuff. He still sends most messages and edits most docs. People can tell when the ask was thought through.
Destructive actions, purchases, and external sends stay on your side. So do production changes.
Source
Field story: Grok Bot for PMs by Kevin Niparko on x.ai. We kept the attention list, the partitioned roster, and the "eng manager does not code" rule. His hardware side quests and internal PR share stay his.