This guide explains where an AI editor looks when it writes, fixes, or checks an article, what it puts first, and how it balances solid evidence with a bit of fun, following the current Article Quality v5 rules.
It contains no personally identifiable information. Private-life details, workplaces, addresses, and account information have all been generalized.
The 5-second answer: eight layers to check
An AI editor looks at roughly eight layers:
- The request at hand — what to make, and what not to make
- Past memory and past sessions — writing style, quality rules, exceptions decided earlier
- The latest main branch on GitHub — the contracts, quality epoch, editing rules, and QA that are in force right now
- The original article and source material — the "ground you must not break": numbers, dates, quotes, proper names, and uncertainty
- Primary, official, and research sources on the web — what's current, plus rules, specs, studies, and counter-evidence
- The article itself — reader needs, angle, structure, claims and evidence, humor, and natural flow
- The 12 language versions — localize the ideas, not just the words, and keep exact names intact
- The live site, real data, and QA evidence — phones, zoom, long text, links, tables, even error states
So this is not "read Google's guidelines and write something." If you only look at Google, you end up channeling a Google employee's ghost. If you only look at Microsoft, the manual starts giving birth to more manuals. What an article factory needs is editorial judgment that connects several strong sources to the purpose of that particular article.
1. First, look at "this order"
The top priority is what the user is asking for right now.
There are conditions that apply only this time: "be thorough," "keep it short," "lots of jokes," "research-based," "no images," "12 languages," "leave out personal information." However impressive an old template is, if the order says ramen and you serve curry, that's just an accident.
Mainly, five things get decided here:
- Who the article is for
- What question the reader has
- What the reader should understand or be able to do afterward
- How much research is needed
- Which bans, formats, and tone apply only this time
At this stage the article gets its "Reader Need" and its "Reader Outcome."
2. Next, look at past memory and past sessions
An article needs more than the instructions of the moment. It also needs the editing rules that have built up over time.
For example, today's standards include things like:
- Reads naturally to an ordinary adult, yet a middle schooler can still follow the meaning
- The conclusion in 5 seconds, the overview in 30
- The title alone tells you what the article is about
- Reading only the H2 headings (the section titles) shows the flow
- One topic per paragraph, as a rule
- Conclusion, then reason, then concrete example
- Explain what a technical term actually is before giving its name
- Use jokes to make the point easier to grasp
- Generalize anything that could identify a person
- Don't translate the 12 languages word for word
The important point here is don't treat past memory as a source of facts.
"We decided this earlier" is fine for production policy. But if you let "we said so before" settle prices, laws, specs, study results, or the current state of a system, the article turns into a time capsule.
Memory is for keeping editorial policy consistent. Facts go back to current evidence.
3. The latest main on GitHub is the "current instruction manual"
For long-term operation, the latest main branch on GitHub is the source of truth for the article factory.
The current quality system doesn't add a new 29th quality axis. Instead, it folds specialist editing know-how into the existing 28 parent axes as sub-gates.
The main sources of truth the AI looks at are:
- The current quality epoch
- The 28 quality axes
- Reader experience and cognitive accessibility
- Content quality
- Extended Microsoft and Google standards
- The professional editing workflow
- Japanese editing rules
- Natural writing style
- Multilingual and localization standards
- Information passport and freshness
- The definitions of PASS, SAMPLE_PASS, and COMPLETE_100
- The contract for formal adoption, publication, and verification
What matters here is that the latest rules outrank old success records.
Even an article that scored 100 yesterday becomes "yesterday's 100" once the quality rules change. It's like winning every match under last season's regulations: you can't walk into a tournament under the new rules and say, "I won yesterday, so I'm the champion today."
4. The original article and source material are "ground you must never break"
An editor can change how the text is presented. An editor cannot rewrite facts to suit themselves.
The things that get special protection are:
- Numbers
- Dates
- Amounts of money
- Units
- Head counts
- Names of systems and programs
- Product specs
- URLs
- Quotations
- Sources
- Proper names
- Study design
- Uncertainty
- The state of "we don't know"
If you turn "about 30%" into "around half" to make it read better, it's not more readable, it's a different timeline.
The basic job of an AI editor is to fix the order, the explanations, the examples, the headings, and the wording without changing the meaning.
5. On the web, start with the "strongest evidence"
When an article needs up-to-date information or outside facts, you search the web. But you don't bow down to whatever tops the search results.
Sources are weighed roughly in this order of strength:
- Standards, laws, and primary sources
- Guides and specs from the official provider
- Primary research and peer-reviewed studies
- Professional editorial and journalism standards
- Reliable secondary sources
- Communities, social media, and personal accounts
Of course, it varies by article type.
If the question is "how did users actually feel about it," Reddit and social media can matter. But if the question is "what does WCAG (the web accessibility standard) require," settling it with a forum post that says "probably 24px lol" would make the standard cry.
Also, for analysis, comparison, research, recommendation, and high-impact articles, you look for sources that could break your own conclusion.
"I gathered five sources that support this theory!" isn't research. It's a fan club.
6. The 28 quality axes are not "28 gods"
The current system keeps 28 parent axes. At a glance, that's 17 reader-experience axes and 11 content-quality axes.
The 17 reader-experience axes
- Attention
- Memory
- Decision-making
- Processing speed
- Understanding
- Information scent (clues that tell readers where to find what they want)
- Visual clutter
- Hierarchy
- Layout
- Text
- Color and vision
- Interaction
- Multilingual support
- Assistive technology
- Motion
- State changes
- Resilience to real data
The 11 content-quality axes
- Reader, purpose, and outcome
- Original value
- Trust, authorship, and how it was made
- Risk, numbers, and decisions
- Meaning of tables, figures, and images
- Steps, actions, and memory load
- Inclusion and culture
- Whether the reader's task actually succeeds
- Completeness per region
- Performance and focused reading
- Text as sound (how it reads aloud)
You don't have to chant all 28 in the text every time.
The point isn't to fill in the axes. It's to think ahead about where the reader will trip, and use the axes that relate to that obstacle.
If your article factory starts saying, "Tonight's offering to the 28 gods is still short by three axes," that's not quality control. That's a religious organization.
7. Look at the 14 professional-editing sub-gates
Under the 28 parent axes sit 14 sub-gates that come from professional editing.
- Reader need and outcome
- The article's angle
- The hook at the start
- What this article is about and why it matters
- Putting the key information first
- The role of each paragraph
- Matching claims to evidence
- Counter-evidence check
- Strength of sources
- Editing in the order structure, then facts, then prose
- Delivering on the title's promise in the body
- Signposts that let readers predict what comes next
- Consistency of notation and terms
- Freshness after publication and content debt
That said, you don't apply all of them mechanically to every article.
A food-experience piece doesn't need to wear Reuters' news structure wholesale, and an API spec doesn't need to open with "Suddenly, the noodles laughed."
Use only the techniques that fit the article type. That's the point.
8. The editing order is "structure, then facts, then prose"
Order is a quietly powerful part of professional editing.
Pass 1: structure
- Does it answer the reader's question?
- Does it have an angle?
- Is the conclusion too late?
- Does the flow come through from the H2 headings alone?
- Are any sections duplicated?
- Are any sections unnecessary?
Pass 2: facts and evidence
- Does each claim have evidence?
- Does the evidence really support that claim?
- Is the sample size or the comparison baseline for a number needed?
- Is there evidence against it?
- Is anything out of date?
Pass 3: prose
- Are any sentences overstuffed?
- Is technical jargon explained?
- Are there pronouns like "this" or "that" with nothing clear to point to?
- Are the jokes working?
- Has it turned into AI-style decorative bold text or a barrage of "Note:" asides?
Reverse this order and you get an edit that polishes the wallpaper in a house that's being torn down next week.
9. Split number rules into three kinds
When numbers show up, an AI sometimes suddenly wants to invent a threshold. This is where you stop it.
Always sort numbers into three kinds.
Numbers set by an official source or standard
For example, WCAG's reflow at the equivalent of 320 CSS px, 200% text enlargement, and target size.
These can be hard gates, within the scope and exceptions that the standard defines.
Numbers from research
If the subjects, language, conditions, or sample differ, a number can't simply be made into a universal standard.
You can't take "X words in English" from a study and alchemize it with a calculator into "X characters in Japanese."
Internal heuristics of the site
For example, "if there are 4 or more H2s, check whether a table of contents is needed."
These are handy for finding candidates for review, but a rule of thumb must not get promoted to police officer.
Google itself doesn't give a magic number like "the length Google likes." What to look at is not length but whether the article delivers what its purpose needs.
10. Localize the 12 languages; don't just translate them
The supported languages are:
- Japanese
- English
- Korean
- Chinese (Simplified)
- Chinese (Traditional)
- Spanish
- Portuguese (Brazil)
- Indonesian
- Thai
- Vietnamese
- French
- German
When going multilingual, words are sorted into three kinds.
LOCALIZE_CONCEPT
General ideas get replaced with words that people in that language naturally understand.
KEEP_EXACT_NAME
Things where the exact name matters, such as product names, standard names, APIs, URLs, code, file names, and paper identifiers, are kept as they are.
KEEP_EXACT_NAME_WITH_LOCAL_DESCRIPTOR
If the name is kept but doesn't tell you what the thing is on its own, a short explanation is added in that language.
Copying Japanese word order, line breaks, and puns into 12 languages is not translation.
That's forcing a Japanese plug into a foreign wall socket by sheer willpower.
Jokes get localized for meaning too. Ones that don't travel are swapped for a different, natural laugh.
11. PASS is a gate system, not an average score
The current basic approach is gate-first-score-second.
In other words, there is no such thing as:
"The facts are wrong, but the design is 95 and the writing is 96, so the average is over 90. Pass!"
A serious FAIL can't be erased by an average.
SAMPLE_PASS
No serious problems were found in the samples checked.
PASS
The scope of the current target is defined exactly, the required evidence is tied to the current content identity, quality epoch, and rule version, and no serious UNKNOWN remains on any required item.
COMPLETE_100
The hard gates that apply, the semantic review, and the required external and runtime evidence are all in place, and everything is closed across all current targets.
Something is not a FAIL just because no human has scored it. Currently, the official semantic reviewer is the source-grounded AI semantic editor.
But if the AI writes something itself and then says,
"After a rigorous review by AI-sensei, AI-sensei's text has been judged completely correct,"
that doesn't count as evidence.
The judge can be an AI. But the judge must not mold the exhibits out of clay.
12. After the text, look at the real screen
Even if the article text is correct, it can't be read if it breaks on screen.
On the live site, you check at least states like these:
- A 320px-class phone
- A 390px-class phone
- A phone in landscape
- A tablet
- A 1280px / 1440px-class PC
- 200% text enlargement
- WCAG Text Spacing
- Text growth under pseudo-localization (a test that stretches text the way translations do)
- forced-colors (a high-contrast display mode)
- reduced-motion (a setting that turns off animation)
Then you use the "landmines" in the real data.
- The longest title
- The longest unbreakable string
- The longest body
- The shortest body
- The most headings
- The most links
- The most tables
- The most definition-like elements
- The most code / pre blocks
- Articles that mix several writing systems
If you test only ordinary articles and say "looks fine!", that's as different in coverage as finishing a health checkup by doing squats in a library.
Errors, zero results, load failures, and missing translations are also part of the article experience.
13. The research rules run themselves, too
The quality rules themselves go out of date.
So research updates are built into the existing central audit.
- Regular refresh: every 7 days, as a rule
- Deep sweep: every 30 days, as a rule
- Major official changes: brought forward if needed
The sources checked include W3C/WCAG, ISO 24495, Microsoft, Google, Reuters, AP, GOV.UK, cognitive science, HCI (human-computer interaction), reading research, information seeking, accessibility, and multilingual and editing research.
New findings are checked in this order:
Does it overlap existing rules? → Is the source strong? → Which articles does it apply to? → What is the range of the numbers? → Is there counter-evidence?
and then they're sorted into one of:
- ADOPT
- CONDITIONAL
- TEST
- DEFER
- REJECT
- SUPERSEDED
In the first v5 deep sweep on August 28, 2026, no changes big enough to overturn the current PASS standard were found, and the decision was to keep v5.
14. Official guides are not oracles
Microsoft, Google, Reuters, AP, W3C, ISO. All of them are strong sources, but their scope of application differs.
For example, in technical documentation, holding back idioms and humor can help translatability and accuracy.
But apply that rule across the board to a food-experience piece or an entertainment article and you get this:
Consumed hamburger steak. Meat juices occurred. Satisfaction increased.
A sentence that left its feelings behind in the compliance office.
Official guides are about which context they're right in.
It's not "official, so ADOPT for every article." Where needed, it's CONDITIONAL.
15. Even if GitHub goes down, the article's brain shouldn't die
When GitHub can't be reached, you don't have to stop all article writing and semantic review.
As a recovery baseline, these are kept in a separate system as well:
- The 28 parent axes
- The AI editor policy
- The professional editing workflow
- Strength of sources
- Number standards
- How research updates are done
- PASS judgment
- The principle of not stopping the factory
But when GitHub isn't visible, you must not fill in the following by guessing:
- The current SHA (the commit ID)
- The latest receipt
- Whether it's been applied to the repo
- The current official progress
Those are UNKNOWN.
When GitHub recovers, you cross-check the latest main, the recovery baseline, and the latest primary and official information, then return to normal operation.
"Whoever writes last wins" (blind last-write-wins) isn't an editing policy. It's rock-paper-scissors.
16. What we don't do
This article factory avoids at least the following:
- Treating an AI's answer alone as evidence for a fact
- Writing "expert-verified" when no human review has been done
- Calling a number "Google's standard" when Google never said it
- Mass-producing low-value articles just for search rankings
- Using a clickbait title and then not answering in the body
- Changing numbers or uncertainty from the source just to make it easier to read
- Forcing every article into the same news-style or technical-document mold
- Mechanically copying Japanese headings, word order, and jokes into other languages
- Pounding the whole text in bold
- Letting "Note:" asides grow like weeds
- Halting unrelated factories because one article failed
- Claiming "100 points across the whole site" from sample checks alone
Summary: an AI editor is busiest before and after the writing
If you look only at the act of writing, an AI editor looks like a text generator.
But the real work is:
Read the request → read the past rules → read the GitHub source of truth → protect the source material → check primary, official, and research sources → try to refute yourself → build the structure → verify the facts → polish the prose → localize into 12 languages → try to break it on the real screen → close it out with QA → and keep researching the rules themselves on a schedule.
"Write an article" is the start button, not a job description.
The AI editor of an article factory patrols as writer, proofreader, researcher, translation editor, QA, and equipment maintenance, all in one.
And the most important thing is not producing text that follows the rules, but producing an article where readers can pick up the meaning right away, trace the evidence, and make the decisions they need to make.
The rules are tools for that. This is an article factory, not a place to worship the toolbox.
