The editor agent just handed back a revised draft. Before you approve it, you have two versions of the same post sitting side by side. That is the moment where content quality either holds or slips a notch, and most teams either skip it or scan the diff too fast to catch what changed.
This is a walkthrough of how to read that diff. Which edits are real improvements, which are the editor rearranging furniture, and which are quiet voice regressions that will make the final post sound less like your brand and more like everything else.
Why the diff is where quality is won or lost
The writer agent gives you a first draft in your brand voice. The editor agent takes your revision requests in plain language and rewrites accordingly. Both agents are doing what they were told. Neither of them knows that the sentence you asked to tighten happened to be the one carrying your strongest opinion, or that a request to make a paragraph friendlier flattened a line that was doing real work.
The diff is where you catch that. Not by rereading the whole post from the top, but by looking specifically at what moved.
A skimmed diff review is how AI slop sneaks back into a post you thought you had under control. A serious one is how you keep the brand voice consistent across dozens of posts a month.
Set your baseline before you open the diff
Before you compare anything, know what you asked for. Your revision requests to the editor are the reference frame. If you asked for three things and the diff shows fifteen, the editor took liberties, and each one needs to be judged on its own.
Keep the writing rules and brand voice open in another tab. That is the tiebreaker when you are not sure whether an edit is an improvement or a drift.
Diffs that signal a real improvement
Some changes are almost always wins. Look for these first, because if the editor did a good job here, you can move through the rest of the review faster.
Tightened openings. The writer draft often warms up in the first two or three sentences. If the editor cut a hedging intro and dropped the reader straight into the point, that is a real improvement. Compare word counts on the first paragraph. A 30 to 50 percent cut with the same meaning intact is a good sign.
Concrete replacing abstract. If a vague noun phrase got replaced with a specific example, a number, or a named artifact, that is an upgrade. "Improves your workflow" becoming "cuts an hour off the weekly editorial meeting" is the kind of edit worth keeping.
Removed hedging. Look for deleted instances of "can help," "may be useful," "often considered," "one of the most." The writer sometimes leans on these to sound safe. If the editor cut them and the sentence still stands, keep the cut.
Reordered claims that build. If the editor moved a paragraph so a claim gets set up before it lands, that is structural editing doing what it should. Read the two versions in order and see whether the argument follows better in the new one.
Diffs that are cosmetic reshuffling
These are the changes that look substantive in the diff view but are not really doing anything for the reader. Left unchallenged, they eat your review time and sometimes cost you a good line.
Synonym swaps. "Users" to "customers" to "readers" and back. Unless the writing rules specify one term, these edits are noise. Do not spend time on them, and if the writer's original word was more specific, revert.
Sentence splits that add nothing. A long sentence broken into two short ones can be a real improvement or a pointless fragmentation. Read the split version out loud. If the second sentence starts with "This" or "It" and the referent is now three words away, the split hurt more than it helped.
Reordering inside a bulleted list. Unless the list was in a deliberate order (chronological, priority, cause to effect), item order rarely matters. Do not litigate it.
Added transitional phrases. "That said," "with that in mind," "on top of that." Some transitions earn their place. Most are padding the editor added to smooth over a cut. If you can delete the transition and the paragraph still flows, delete it.
Voice regressions to catch
This is the part most teams miss. The editor fixed the thing you asked about, and while it was in the neighborhood, it softened three other lines you liked. The diff will show these as small edits, easy to approve in a hurry.
Opinion getting sanded down. "This is a bad habit" becoming "this can sometimes be counterproductive." "Most tools get this wrong" becoming "many tools approach this differently." When your brand voice is willing to be opinionated, edits like these are regressions even if the sentence is still readable.
Cliches sneaking back in. The editor tries to make prose smoother, and smooth often means generic. Watch for phrases like "in today's landscape," "leverage," "seamlessly," "unlock," "revolutionize." If any of these appeared in the edit and were not in the draft, revert.
Second person turning into third. "You" becoming "teams" or "marketers" or "one." Practitioner-to-practitioner voice lives in the second person. If the editor generalized a direct address, put it back.
Confident verbs turning weak. "Kills the tab clutter" becoming "helps reduce tab clutter." "Refuses to ship generic drafts" becoming "aims to avoid generic drafts." The verb tells you whether the sentence has a spine.
Callouts and specifics getting genericized. If the writer draft named a specific tool, platform, or workflow, and the editor swapped it for a category label, ask why. Sometimes the swap is right, because the specific example was wrong. Often it is not.
The three-column mental model
When you open the diff, you actually have three columns, not two. The writer draft, the editor pass, and the writing rules. Every disputed change gets adjudicated by the third column.
If the editor changed a line and the writing rules do not prefer one version over the other, default to the tighter and more specific one. If the writing rules explicitly prefer one, that is the answer. If the change violates a rule outright, a banned word appeared or a cliche crept back, it is not a debate. Revert.
The three-column habit makes the review faster because most edits stop being subjective the moment you check them against a rule you already wrote down.
What to do when the editor pass is worse
Sometimes the whole revision missed. The editor addressed the letter of your request and lost the spirit. When that happens, do not try to salvage line by line. Go back to the editor with a sharper request.
Two moves that help:
- Quote the specific line you wanted preserved and say "keep this exact phrasing." The editor can protect specific text if it is told to.
- Name the voice regression pattern, not the sentence. "Do not soften opinions in this pass" gives the editor a rule to hold across the whole post, not just one line.
If you are still not getting there after two revision passes, the request is probably ambiguous. Rewrite the ask in one sentence, be explicit about what should not change, and send it once more.
A short checklist for the diff review
Every review comes down to a handful of comparisons:
- First paragraph: tighter and more specific, or padded and generic
- Hedging words: fewer, or more
- Direct address: preserved, or generalized
- Named specifics: preserved, or replaced with category labels
- Opinionated verbs: preserved, or softened
- Banned words and cliches: absent, or crept back in
- Structural order: builds the argument, or scrambles it
- Writing rules: respected, or quietly violated
Run down the list. Approve when every line points the right way. Send it back when three or more point the wrong way. Rewrite the request when it happens twice in a row.
The reason this step exists
A human approval gate is only as sharp as the review that happens inside it. The point of the four-agent loop is that no post ships until someone reads it with intent. If that read is a scan, the loop collapses into any other AI content pipeline, and the output slides toward the same generic middle every other tool is fighting over.
The diff review is where the human is doing the work only the human can do. Twenty minutes here is worth more than any prompt tuning upstream.