You Asked AI for a Draft. You Still Had to Write It.
You briefed it, corrected it, and still rewrote it. How we turn recurring voice corrections into a separate review before the next draft reaches you.
In this article
You sit down to approve a post and end up teaching the AI how you talk. Again.
Maybe the facts are right. The structure is sensible. You gave it examples of your writing, and it still sounds like somebody else.
But the opening sounds like someone introducing a panel. The sentence you would actually say is buried halfway down. And the ending congratulates marketing leaders for embracing the future. You cannot imagine sending it to a customer who knows you.
So you move the useful sentence up. Delete the little sermon. Ask for something more conversational.
Back comes the same argument, wearing contractions.
By then, you have your hands on the keyboard. It will be quicker to write it yourself. The draft was supposed to take work off your plate; now you have written the brief, explained the correction, and finished the post.
What we need here is a voice check: a separate review that compares the draft with writing you have approved, explains where it misses how you think and communicate, and sends it back for revision before it reaches you.
That is how we approach voice inside our Marketing OS content workflow. The writer drafts. A separate agent checks the work against a rubric, including examples of the author’s voice. If the draft fails, the writer gets specific findings to repair. A person still makes the final approval.
Here is what that check needs to recognize, how the loop works, and how to test it against your own edits.
A brand voice check needs to recognize how you explain ideas.
Think about a piece of your writing that sounds unmistakably like you. The words are part of it. So is the problem you noticed, the question you asked, the comparison that made the idea understandable, and the moment you gave the reader something useful to do.
That is what we are trying to carry into AI-written content: your way of reasoning and communicating an idea to another person.
In the post this article grew from, I explained independent review by asking readers to imagine grading their own test. It gives them a familiar reason to question self-grading before I introduce a separate agent. Then I describe the rubric, the rewrite loop, and the need to stop when retries run out.
The useful pattern is the teaching progression: familiar problem, clear solution, explanation, next action. A voice check should recognize that progression. It should not demand that every article repeat the test analogy.
For a marketing leader, the question becomes: would I explain this buyer’s problem, the solution, and the next step this way? An article can use your favorite phrases and still leave the reader wondering what you mean.
Give the reviewer three approved samples of your writing.
Start a file of pieces you would be happy for the AI to learn from. Use work you have actually approved. Keep each piece’s author, channel, intended reader, and original text with it. Add a short note about why the explanation works.
For example, beside the grading-rubric post, the note could say: “The familiar test analogy explains why a separate reviewer is useful. The practical example shows what the reviewer checks. The retry limit keeps the advice usable.”
That tells the AI what to learn from the sample. It has a way to compare the thinking, rather than simply borrowing the wording.
For each assignment, have the workflow randomly select three relevant approved pieces. Relevant matters: the personal voice in a founder’s article may not be the right model for a company product page. Random selection should happen within the appropriate pool.
Keep those same three throughout the draft’s review and revisions. If you change the references every time, a better grade might mean the comparison changed. With only three approved pieces, use all three. With fewer, finish the sample set before treating the voice check as ready.
Define what a passing voice check looks like.
We start with tempo, style, and likeness, grounded in the selected writing samples. Here is what those checks mean for an article like this one:
| Voice dimension | What the reviewer needs to assess |
|---|---|
| Tempo | Does the explanation move at a pace the reader can follow? Does a short sentence emphasize a useful point, and does a longer one connect the reasoning? |
| Style | Are the language, register, comparisons, and humor consistent with the author’s approved writing and appropriate to this reader? |
| Likeness | Does the author notice and frame the problem in a recognizable way, explain why the solution helps, and guide the reader toward a useful next action? |
Give each dimension a score from 1 to 5, with examples of what each level means. A 3 should identify recurring edits still needed. A 4 should mean the writing fits with minor repairs. A 5 needs passage-level evidence of a consistent match, including the reasoning.
Then set the passing rule. A reasonable starting point is a mean of at least 4.5, with no dimension below 4. Those are settings to calibrate against your own judgment. They are not evidence that a reader will find the article valuable.
Have a separate agent finish the whole content review.
The agent writing the article should not also be its sole judge. We give a separate agent the draft, brief, selected voice samples, rubric, and approved factual sources. We leave out the writer’s self-score and argument for why the draft is good.
Its voice check is one part of the assignment. It also needs to finish the rest of the content rubric:
- Does the opening describe a problem this buyer recognizes?
- Is the solution clear, and does the explanation show why it helps?
- Can someone read the headings alone and understand the argument and payoff?
- Do the examples teach the promised idea, and can the reader take the next step?
- Are the claims supported, the visuals understandable, and the downloads working?
A draft can sound like you and still make a weak argument. A good argument can still contain an unsupported claim. The reviewer needs to report both, rather than stopping when the voice score passes.
This is how the check fits into Marketing OS: brand knowledge and approved evidence guide the work, review standards travel with it, and the person responsible for the brand keeps the approval decision.
Send failed drafts back for a bounded rewrite.
A failing review goes back to the writer with the passage, the criterion it missed, and the change needed. The writer revises; the reviewer checks the new draft against the same samples and full rubric.
The diagram below shows that handoff and the decision to return the work to a person.
Set the limit before the loop begins. Two rewrites and three reviews in total is a sensible budget. When that budget is spent, stop and show the unresolved finding to a person. If factual evidence is missing, ask for it; rewriting cannot supply the proof.
A separate agent can still miss what you would notice. OpenAI’s evaluation guidance emphasizes task-specific criteria and calibration against human judgment. The loop needs your judgment to become useful, and a passing grade is still a request for your approval.
For the first test, choose an article you have already edited. Give the reviewer the original draft and keep your edits out of its packet. Compare its findings with your corrections. Did it notice the confusing explanation? Did it understand why you moved the solution up? Did its requested repair help?
If it missed the issue, sharpen the sample annotations or rubric before relying on it for the next article. Do not lower the bar to get a pass.
Measure success by how much rewriting still reaches you.
The result we want is fewer drafts that turn approval into another writing assignment. Count the articles you review and how many still need substantial explanation or rewriting from you. Record what the reviewer missed so those corrections can improve the standard.
Compare that with your work before the check. Keep the comparison honest: similar kinds of articles, your actual edits, and the same definition of substantial rewriting. A higher agent score by itself does not tell you the process is helping.
You will still make judgment calls. The useful shift is that the writer has already received a specific review and repaired what it can before the next draft lands in front of you.
