A 20-minute editorial review for AI drafts, from top to publish

The draft is done. The writer agent produced 1,600 words in your brand voice, the editor agent already took two rounds of revisions, and now it sits in your queue with an approve button next to it. Your job is not to rewrite it. Your job is to decide, in under half an hour, whether it goes live, gets bounced back to the editor agent, or gets killed.

Here is the rubric to run on every draft before you approve. It is six passes, timed roughly, and it ends with a clear decision instead of a blank stare at the screen.

Before you start

Have the topic brief open in one tab and the draft in another. Keep your brand voice rules and writing rules visible somewhere you can glance at. If you export drafts as markdown, work in the markdown, not a rendered preview: the review is faster when you can see the raw structure of headings, lists, and links.

Set a 20-minute timer. It sounds gimmicky. It works. Editorial reviews expand to fill whatever time you give them, and the whole point of a supervised pipeline is that you approve, not rewrite.

Pass 1: the two smell tests (60 seconds)

Two quick reads catch generic filler before you commit to a real review.

Smell test one: the opening. Read the first two paragraphs out loud. If they could sit on top of any post in your industry with the topic swapped, the draft is generic. A good opening names the specific problem, the specific reader, or a specific number. A bad opening announces the topic and promises to cover it.

Smell test two: the so-what test. Pick three subheadings at random and read only the paragraph underneath each. If the paragraph tells you what a thing is but not why the reader should care or what to do differently, it is filler. Practitioner posts should answer "and then what" on every section.

If both smell tests fail, do not run the rest of the rubric. Send the whole draft back to the editor agent with a revision request: "The intro and the body sections are generic. Rewrite with a specific reader problem in the opening and add a concrete recommendation or example under every H2."

That is a 90-second decision that saves you 18 minutes.

Pass 2: factual claims (5 minutes)

Now assume the draft is worth saving. Read it top to bottom, but only look for facts. Highlight anything that is a number, a date, a name, a quote, a product capability, a stat, or a specific claim about how something works. Ignore everything else.

For each highlight:

• If the fact is common knowledge or trivially verifiable, leave it.
• If the fact is specific and material, click through and verify. If the writer agent sourced it from a live research pass, the citation trail should still be visible in the run log. Check it.
• If the fact is unverifiable, cut it. Do not soften it with "roughly" or "often." Cut the sentence.

The most common failure mode of an AI draft is a plausible number without a source. Practitioners smell those instantly. One made-up stat costs you more trust than ten paragraphs of generic prose.

This is inline work. You do not need to bounce a draft to the editor agent to delete a bad stat.

Pass 3: first-person voice (3 minutes)

Skim the whole draft for pronouns. Look for two things.

Consistency. If the intro uses "we" and the middle switches to "the marketer," pick one and fix the outliers. In practitioner content, "you" and "we" almost always beat third-person abstractions.

Authority. First-person plural should sound like the brand actually has a point of view. "We recommend" is stronger than "many experts recommend." "We have seen this fail on client accounts" is stronger than "some teams have reported issues."

If the draft is written entirely in the passive third person and the brand voice calls for direct first-person, that is an editor agent job, not an inline fix. Send it back with: "Rewrite in first-person plural throughout. The voice should feel like a practitioner talking to another practitioner, not an analyst summarizing a market."

Pass 4: banned phrases and cliches (2 minutes)

Run a find on your banned words list. Every project should have one. Common suspects include "delve into," "in today's fast-paced world," "unlock the power of," "it is important to note," "in conclusion," "revolutionize," "game-changer," "leverage," "seamlessly," and "navigate the landscape."

If the banned word list is configured on the writer and editor agents, these should be rare. When they slip through, delete inline. Do not rewrite the surrounding sentence unless the phrase is load-bearing. Nine times out of ten, "It is important to note that X" becomes "X."

Also check for em dashes if your style rules ban them, unmatched heading levels, and paragraphs longer than four lines. Those are all inline fixes.

Pass 5: link opportunities (3 minutes)

Read the draft one more time with your site map or cornerstone list open. Look for two kinds of link opportunities.

Internal. Every mention of a topic you have already covered is a link candidate. Do not force it: three or four internal links in a 1,500-word post is plenty. Prioritize links from a new post into a cornerstone, not the other way around.

External. If the post cites a stat, a study, or a third-party tool, that citation should be a link. If the draft names something and does not link it, add the link.

This is entirely inline work. If the writer agent regularly ships drafts without internal links and your site has an established cluster structure, that is a configuration problem, not a per-draft problem. Fix it in the writing rules, not in the editor agent.

Pass 6: CTA fit (2 minutes)

Read the last section and any inline CTAs. Ask three questions.

• Does the CTA match the search intent of the topic? A how-to post ends with "try this," not "book a demo." A comparison post can end with a demo. A top-of-funnel explainer usually ends with a subscribe or a next-read.
• Is the CTA specific? "Sign up" is weak. "Start a free run and let the research agent propose ten topics for your site" is strong.
• Does the CTA appear once, at most twice? Multiple CTAs cannibalize each other. Pick the one that fits the intent and cut the rest.

If the CTA is wrong for the intent, that is an editor agent job. If the CTA copy is weak, fix inline.

Inline fix versus editor agent: the rule of thumb

The pattern is the same across all six passes.

Fix inline when the change is local: a phrase, a stat, a link, a heading level. You are faster than the agent for surgical work.

Send it back to the editor agent when the change is structural: voice, intent, opening, section-level rewrites. The agent is faster than you for anything that touches multiple paragraphs.

The trap is trying to rewrite structural problems inline. That is how a 20-minute review turns into a 90-minute ghost-write. If you catch yourself rewriting a full section by hand, stop, revert the buffer, and send the draft back with a specific revision request.

The revision request format that works

When you send a draft back, be specific about what to change and why. Vague requests get vague results.

Weak: "Make the intro better."

Strong: "The intro is generic and could apply to any tool in the category. Rewrite it to open on the specific problem of an editor spending 90 minutes on a review that should take 20. Keep it to two paragraphs, first-person plural, no throat-clearing."

The editor agent takes conversational revisions, so you can talk to it the way you would talk to a copy editor you trust. The clearer the ask, the fewer rounds you burn.

Approve, bounce, or kill

At minute 20, you have three choices.

Approve. The draft is factually clean, in voice, free of banned phrases, has its links, and the CTA fits. Ship it.

Bounce. One or two structural issues, clear revision request, back to the editor agent. Expect one more round, then re-review from Pass 1.

Kill. The topic was wrong, or the draft failed both smell tests and the underlying angle is bad. Killing a draft is not a failure of the pipeline. It is the pipeline working: the human approval gate exists so bad posts do not ship.

The point of a supervised agent loop is that you spend your time judging, not typing. Twenty minutes of structured judgment beats two hours of half-hearted rewriting, and the draft that comes out the other end is the one you would have written yourself, if you had the time.