AI Editor in your Newsroom
Walk into many newsrooms today, and you’ll often find an editor with an AI chat window open on a second monitor. The article is pasted in, and the prompt reads: "Check grammar, flag any unclear sentences, and cut this to 700 words." Then the editor takes the output, tidies it up, and moves on to the next piece in the queue.
This is because the newsrooms are running lean, and the pace of publishing has always been increasing. So editors are doing the arithmetic: AI can handle a grammar pass in 20 seconds. That is 20 minutes returned to their afternoon. Of course, they are using it.
But here is the part no one is talking about: the way most newsrooms are using AI assistants right now is leaving the majority of the value on the table - and introducing risks that a blunt prompt-and-paste workflow cannot catch.
WHAT A RAW PROMPT CAN'T KNOW
When an editor pastes an article into a general-purpose AI chat window and asks for a grammar check, they are asking a tool with no memory, no context, and no institutional knowledge to perform a task that has always depended on all three.
A skilled copy editor at a newspaper does not just apply grammar rules. They carry years of accumulated knowledge about how that publication works: the house style guide, the legal lines the paper has drawn after consultations with counsel, the phrases their editor-in-chief has banned after one too many reader complaints, the sources the investigations desk no longer trusts, the beats where a particular journalist's reporting tends to need extra scrutiny. That knowledge is invisible and irreplaceable.
When you ask a general AI tool to review an article without any of this context, several things happen quietly:
The tool applies a generic grammar standard, not your house style. It has no idea whether you follow AP or your own style guide, whether you write "15%" or "15 per cent", or whether you hyphenate "decision-maker" or leave it open. It guesses, and it often guesses wrong in ways that look right to a rushed eye.
It has no visibility into publication risk. It cannot flag that a paragraph accuses a named individual of financial misconduct without sufficient attribution - or if it does flag it, it does so on general principles, without knowing your jurisdiction's defamation law or your publication's prior settlements. It cannot tell you that a source quoted in paragraph five has been the subject of complaints about reliability in the past. It does not know your publication exists, let alone its legal history.
It has no sense of editorial taste. Every publication has a voice - a set of judgements about what kinds of stories get told how, what registers are appropriate for which sections, what the readers expect. A general AI tool cannot calibrate for this. It optimises for legibility, not for belonging.
None of this is a criticism of AI tools. It is a description of how they work when given nothing to work with. The problem is not the tool. It is the workflow.
INSTITUTIONAL MEMORY YOU CAN ENCODE
Claude, the AI assistant developed by Anthropic, has a feature called Skills. A Skill is a set of structured instructions - a configuration file - that you write once and that shapes how Claude behaves every time it is used. Think of it as the briefing document you would hand to a new copy editor on their first day: here is what we do here, here is what we watch for, here is how we make decisions.
Skills live as simple Markdown files, which do not require a developer to update each time. A senior editor who can write in plain English and is comfortable creating a text file can build one. The Skill is loaded into Claude's context before it reviews any article, which means Claude is never starting from a blank slate. It knows what publication it is working for before it reads the first word of the piece.
Set up the Skill in Claude
Navigate to: https://github.com/Pratzz14/journalism-editorial-review
Download the code files:
Navigate to Claude: https://claude.ai/new
- Enter "/{name_of_the_skill}" and paste your draft article. Get the article review.
THE SIX REVIEW MODULES
In the above journalism review Skill, it covers six distinct editorial checks. Each is a separate file that Claude consults when running that specific module. Here is what each one does:
Claims audit: Factual assertions, statistics, dates, direct quotes, causal links. Prioritises claims by consequence and checkability.
Source credibility: Whether named sources can be independently verified. Flags conflicts of interest, credibility gaps, and unverifiable digital content.
Attribution audit: Unattributed quotes, unsourced statistics, vague collective references ("experts say", "critics argue") that leave readers unable to assess the claim.
Writing authenticity: AI-generated language patterns - specific words, phrases, and structural tells that signal machine-generated text rather than a journalist's voice.
Publication risk: Defamation exposure, privacy concerns, confidential source protection, pre-trial publicity risk, and copyright. Flags by severity level.
Copy editing: Grammar, spelling, punctuation, wordiness, passive voice, journalistic conventions - leads, headline style, quote handling, number usage.
Each module is non-blocking, and the output is advisory. Claude flags the journalist and editor to decide. No automated system should have the final word on what a newsroom publishes - that responsibility stays with the people in the room.
You can also customise the skill using Claude:
You can also add/update/delete files in Notepad/ VSCode IDE, and reupload the skill to Claude:
THE HONEST CAVEAT
A Skill is only as good as the knowledge you put into it. The first version of your Skill will be imperfect. It will miss things your best editors would catch, and it will flag things that do not need to be flagged. This is not a reason to dismiss it - it is the same description you would give of a new hire's first month. The difference is that a Skill gets better the moment you teach it something, and it never forgets.
The editors who will get the most from this approach are the ones who treat the Skill as a living document rather than a setup task. Every time Claude misses something important, that is a lesson to write into the module. Every time it flags something you would not, that is a scope boundary to sharpen. Over a few weeks of iteration, you end up with something that encodes the editorial standards of your publication more explicitly than most style guides do - and which applies them consistently, at speed, to every piece before it goes out.
The paste-and-prompt approach is understandable. But it is borrowing against a debt you will eventually have to repay - in corrections, in reader trust, in legal exposure, in articles that sound like nobody in particular wrote them. Spending the time to configure AI tools properly is the same investment good editors have always made: doing the work before publication so you do not have to do it after.
If you have questions, want to share how your newsroom is using it, or just want to talk through the setup, feel free to reach out at pratik.mahankal14@gmail.com.
