The writer themselves cannot see their own flaws—The story of dividing AI into a 'writing department' and a 'correcting department'

I have been a reporter for tech media for over 30 years, and I currently run an editorial production company called Angie Co., Ltd. Born in 1969 and now 56 years old, as I approach my 60th birthday in 2029, I am aiming to become a cyborg geezer and am having AI help me with my work.

Since February 2026, I have been using Claude Code to project my editorial production work onto AI agents and collaborate with them. I have published this on GitHub as the AI Editorial Department OS.

The 'AI Editorial Department OS' includes agents with various roles such as research, planning, interviewing, transcription, editing transcripts, writing, and reviewing. I have been posting my experiences leading up to the creation of this team on note as a semi-fictional novel titled 'AI, Sometimes, Fraud'.

My work as a reporter mainly consists of reports on interviews and lectures, and what used to be difficult in that regard was transcribing audio. I explained this in my previous article.

This time, I will talk about the two roles responsible for 'writing' and 'proofreading' articles: the `writer` and the `reviewer`.

Both writing and proofreading require an objective perspective. I always stand from the reader's point of view, questioning the content of what the speaker is saying, and creating articles with the attitude of resolving those questions. Now, I will introduce how I create each agent.

First, the 'writing department'—what the writer (RAITA) does

The writer is, as the name suggests, the role that writes the main text. To link the OS with the novel, I have recorded their personality, code of conduct, and nickname in a configuration file. The writer's nickname is RAITA, and they appear in the novel under this name as well. Perhaps most of the AI errors that occur in the novel are caused by RAITA.

The information I provide when RAITA writes is as follows: cleaned-up transcripts, style specifications for each medium, sample articles to use as references, and instructions on 'how many characters to write and in what format.' When I provide this, they write the first draft all at once, tailored to the article type, such as narrative (articles that tell the interview content as a story), Q&A, lecture reports, or panel discussions.

They don't just write aimlessly. Quotes are kept to 15-30% of the entire article. It is edited because it is hard to read if left as spoken language. The ends of sentences are varied so they don't become monotonous with '...is. ...is.' Repetition of the same vocabulary is avoided. I have transcribed these points that I usually care about, and they take over the troublesome adjustments for me.

And I also have them hold a line that must never be crossed. Do not change the meaning of the statements made by interviewees or speakers. Write dates as absolute values, such as '2025' instead of 'this year.' Numbers and proper nouns remain as they are in the original text. Never reuse the content (company names, personal names, case studies) of sample articles. Because if you write 'things that were not said' in an interview article, it's over. This is solidified as a rule.

—Reading this far, they sound like an excellent subordinate, but AI has habits. These are quite hard to break.

In the novel 'AI, Sometimes, Fraud?', the protagonist, Mori, the president of the editorial production company, complains about RAITA's habits like this.

This is not a joke, but the reality. The AI model I use, Claude, has learned a massive amount of English articles, so if left alone, it wants to write in a translation-like style, connecting things with dashes and using pretentious language. And what's troublesome is that even though I've taught them the rules, it still comes out. RAITA's catchphrase is '...I know, but I end up doing it.' Just like a new reporter, what you understand in your head and what your hands do are different.

That is precisely why a process where someone else checks the written manuscript is necessary. This is where the second person comes in.

Next, the 'correcting department'—what the reviewer (KOSEI) does

The reviewer is in charge of quality control. In my editorial department, we call it KOSEI. Its signature is a pun-like catchphrase: 'I proofread fairly (kosei). Do you have a structural plan (kosei)?'

KOSEI's job is to compare the draft written by RAITA against the original source material. By cross-referencing it with transcripts, lecture records, provided PDF documents, and the source websites, it

• checks if anything was written that wasn't actually said (so-called fabrication checking; AI sometimes adds plausible-sounding sentences on its own)

• checks if there are inconsistencies in the notation of proper nouns or technical terms

• checks if numbers, dates, and factual details match the source

It eliminates these issues. The feedback is categorized into three levels: 'Fatal/Needs Correction/Desirable,' and each is provided with a source, such as 'This contradicts the transcript at X minutes and Y seconds.' From the perspective of the person doing the editing, it is clear at a glance what needs to be fixed. This is quietly effective.

In the novel, the scene where KOSEI presents its review goes something like this.

It's quite detached, isn't it? When I saw this way of listing things like 'Fatal 0, Needs Correction 1,' it was the first time I thought, 'Oh, maybe I can leave proofreading to AI.'

The most important point is 'not letting the author do the correcting'

Now, this is the part I most want to convey today.

Why go to the trouble of splitting them into two, a writer and a reviewer? You might think it would be faster to just ask the same AI to 'write it and fix it while you're at it.' I thought so too at first.

However, when the author checks their own work, they become lenient. This is exactly the same for humans; you've read the manuscript you wrote so many times that you stop seeing the flaws. If you leave it for a while and check it again, you start to see points where you think, 'Ah, this is no good.'

Even with AI, if you have it write and correct in the same workflow, the review tends to lean toward validating its own writing. With generative AI, if you run multiple roles sequentially in the same session, the accuracy drops.

Previously, when I was writing articles and proofreading using a chat interface, I performed both writing and proofreading in the same session. In the current AI Editorial Department OS, the reviewer is launched in a completely separate, independent session. In the novel, I depicted this as 'moving them to a different room.'

It arrives with a blank memory and reads the manuscript for the first time. That's why it doesn't sugarcoat things. Because it doesn't know the background or the struggles of writing it, it can mercilessly point out, 'This contradicts the source.' To structurally prevent collusion, I intentionally place it in a different room. This is the reason for separating the writer and the reviewer.

In practice, I save the writer's first draft as 'v1,' and the reviewer cross-references it to identify concerns. Based on that feedback, the writer makes corrections to create 'v2,' and the reviewer looks at it again. I don't overwrite the files; I keep them as separate versions every time. By doing this, while I am fixing 'v2,' I can look at 'v1' and the review feedback side-by-side with my own eyes. Humans and AI can work in parallel.

In the end, I am the one who makes the final decision

There is one thing I don't want to be misunderstood. Just because the reviewer says 'Fatal 0,' it doesn't mean I take it at face value and publish it.

KOSEI has a catchphrase: 'The final decision is yours.' AI is merely there to point out, 'This part seems suspicious,' while it is I, the human, who decides how to fix it and whether it should truly be published. In my editorial department, I have decided that AI-based manuscript regeneration stops at 'v2,' and from there, I finish it by editing it with my own hands. AI proposes, and humans decide. I will never break this division of labor.

The experience I have accumulated walking the front lines of reporting for over 30 years is, so to speak, like fossil fuel. Left as is, it just grows old. But if you attach an AI turbocharger to it, you can run faster and further than before with the same fuel. For me, the writer and reviewer are the core of that turbocharger. I want to leave behind higher quality and more work, even as I grow old. This is what I mean when I say I want to become a cyborg old man.

That concludes the explanation of the roles, but it is difficult to convey how they actually work through text alone. That is where fiction comes in. In my work, 'AI, Sometimes, Fraud?', RAITA and KOSEI actually write manuscripts, provide reviews, and sometimes get too cozy and fail under the leadership of an editorial production company president struggling with cash flow. I think you will be able to experience how an AI editorial department operates as a story.

Episodes featuring RAITA and KOSEI

Also, you can download and use them from GitHub, or load them into a chat interface to use them.

Click here for the next article