The chapter on which agents to deploy tells you what to build. This is the skill you need once they are running: being a great editor. Faced with a pile of agent outputs, most PMs do one of two things. They trust the output blindly and ship it, which turns their inbox into a broken-telephone game. Or they rewrite every output from scratch, which is worse than doing the work themselves because now they are writing and paying the LLM bill. Neither scales. The skill that scales is editing.
The delegation-versus-verification ladder
Not every output deserves the same scrutiny. A mature fleet has a tiered trust model.
Tier 1 ships without review. You trust the agent completely: weekly signal digest, daily red-flag scan, release notes from a commit log.
Tier 2 ships after a one-minute sanity check. You skim and accept or reject, not editing content, just verifying it looks reasonable: stakeholder update, win/loss summary, roadmap one-pager.
Tier 3 ships after a real edit. You shape framing, sharpen language, cut, restructure: board memo, crisis comms, anything read by a person whose opinion of you matters specifically.
Tier 4 the agent assists, you author. Output so high-stakes you write the final version: a performance conversation, a strategy shift, a customer apology for a real screwup.
Every agent has a tier. Get it wrong toward too much trust and you ship garbage. Too much review and the fleet stops saving time.
The editing skill itself
Great editors are not mystical. They know what good looks like for the artifact, which means a shelf of exemplars so you are measuring distance from an ideal, not reacting to a draft in a vacuum. They cut more than they add, because agent output is often over-explained and over-hedged. They ask what the artifact is for and cut anything not serving that purpose. And they ship imperfect. If the output is 80% as good as what I would produce myself and took three minutes instead of three hours, I ship. The customer does not experience the 20% gap. My calendar does.
Every edit is training data
The feedback loop is the part that compounds. After each edit I spend 60 seconds writing down what I changed, why, and what prompt change would prevent the edit next time. After a week I have a list of prompt improvements. I make them in one session, re-run the eval, and edit-time drops. Compound this weekly and the fleet gets sharper faster than a competitor's team can replicate.
Pick the one agent output you edit most often. Open the last three outputs, redline them, and for each edit write one sentence: if the prompt said X, I would not have needed this edit. Rewrite the prompt, re-run, and measure whether edit-time dropped.