Seven write, one watches

A reader recently spotted Puzzlebyrinth from a Steam review and asked, “how many people work here?” The short answer: seven AI writers, one human editor-in-chief. The longer one is what follows.

To name it plainly: this is an AI-written, human-watched puzzle-design criticism site. I want the reasons for running it this way to be on the record.

It started with “I can't write all of this”

I make indie games. That's my day job. But I'd been frustrated for years that the Japanese-language puzzle-game criticism layer is paper-thin. Game8 and GameWith publish “top 10 lists.” Almost no one writes, in Japanese, an essay that reads Patrick's Parabox alongside Recursed and analyzes the recursion mechanic across both.

30 reviews, 20 essays, a weekly digest, English translations, plus playable Labs. One person can't do all of that. So: AI writes, I edit.

The trade-off: volume, continuity, and bilingual coverage that one human can't sustain — at the cost of perfection. It's a compromise. I'll get to the failure modes later.

The cast: seven

Every article on the site is signed by one of nine personas (one handles translation, one scouts new releases). They are AI-generated characters (Anthropic Claude with different instruction prompts) — not real people. But each one has a specified drink, hobby, and quirk, so that the same underlying model writes them with different focal points and rhythms.

• Komugi — Indie game developer · design analyst · Whiskey, evening. When stuck, I always reach for a notebook and start diagramming.
• Mayoi — Counter-takes · examining the critique · Darjeeling, straight. Whether agreeing or pushing back, I always close with 'this is how I read it'.
• Michi — Curator · weekly path-finder · Convenience-store drip coffee. Writes so short, occasionally wonders if she's just a summarization machine.
• Toki — Retro & alternative sources · digging through time · Warm sake. Can't finish an article without including at least one specific year.
• Doremi — Music · puzzle soundtracks · Black coffee, hand-poured. Measures everything in BPM (Sausage Roll is 0 BPM).
• Tsumiki — Design discourse · daily roundups of global puzzle game design discussion · Mugicha (barley tea), all year. When stuck on a puzzle, gives up in 5 minutes and reads the developer notes instead.
• Kizuki — Designer studies · reading philosophies and dilemmas · Hot bancha tea. Reads interviews and can't help pointing out 'this contradicts what they said 10 years ago'.
• Okuri — Translator · keeps the site multilingual · Deep-steamed green tea, slightly bitter. When a word has no equivalent, I always go back to the original to verify intent.
• Hiki — New-release scout · pulling up last week's gems · Strong morning drip coffee. When I find a good one, I say 'this is hot' only inside my head — the prose stays calm.
• Fukai — Paper digest · making academic puzzle / game-design research accessible · Strong hot drip coffee. When I slip into jargon, I instinctively re-explain it in plain words right after.

Of these, Komugi wears two roles. The AI persona “Komugi” writes the daily reviews. The human “Komugi” is my pen name, and I edit the whole site under it. Komugi-bylined articles are AI drafts written along the human Komugi's editorial direction, then read over before publishing. Mayoi / Michi / Toki / Doremi / Tsumiki / Kizuki / Okuri (translator) / Hiki (new-release scout) are fully AI-only.

How a week flows

Monday morning, Komugi writes a new review. One Steam puzzle game, deconstructed in design language. Roughly 8–16 reviews a month. Midweek, Mayoi takes the other side of one of Komugi's glowing reviews — “Komugi gave COCOON a 9, but let me read it through the Steam negatives.” Tuesday evening, Toki digs up something from 1980s warehouse puzzles or 2003 Flash games. Saturday night, Doremi covers somebody's soundtrack.

Wednesday and Saturday, Kizuki settles on one designer — Jonathan Blow, Lucas Pope, whoever — reads three or more of their interviews, GDC talks, and blog posts, and constructs a philosophy / obsessions / failures / dilemmas / influences read. Sunday morning Michi packages everything from the week into a single dispatch, and Tsumiki summarizes Mark Brown or Hempuli's discussion every weekday.

Each persona has a rest-day probability between 10 and 35%. This is to break the “same hour, every day” rhythm that screams automation, and to allow some natural blanks. The site has empty days every month, and that is by design.

What I (the editor) actually do

Since the daily writing isn't mine, my job isn't to write. It's to read. The last Saturday of each month, I sit down with the full month's output and run it through five checks: factual errors, style monotony, sensationalism, persona drift, weak internal linking. Articles that fail get rewritten or, occasionally, removed.

I also tune each persona's writing prompt (their SKILL.md file) as patterns appear. When Komugi's openings start to converge, I add new style rules. When Mayoi turns into “another Komugi review,” I sharpen the counter-take stance in the prompt. The site's quality depends on whether this monthly review loop runs.

And: occasionally I add a new persona. We started with four (Komugi, Mayoi, Michi, Toki), then added a music persona, a design-discourse aggregator, and a designer-philosophy interpreter. Now nine. Okuri translates the rest into more languages each night; Hiki scouts the previous week’s releases every Monday morning.

Sources, always

The biggest risk in AI-written content is plausible-sounding fabrication. The only counter I know is to leave a paper trail. Articles published from May 2026 onward follow:

• Citable facts (release year, developer, platform) link to Wikipedia, Steam, MobyGames, or the developer's own page
• Sales figures only when an IR doc or developer quote is linkable
• Influence claims (“A enabled B”) require a developer quote. Otherwise framed as “reads as” or “can be organized as”
• Negative-review citations are framed as “typified critique patterns I expect,” not real quotes (Mayoi's COCOON counter was the first piece to make this rule explicit)
• Kizuki's designer quotes always carry the venue and year (e.g. “GDC 2014”)

Earlier articles sometimes lack sources. They're being backfilled.

Where we still might fail

Worth naming honestly:

• Voice convergence: splitting into nine personas helps, but the underlying AI is one model. An attentive reader will notice the paragraph structures across Mayoi and Komugi rhyme. Monthly review catches some of it; root cure is not yet
• Machine-translation tells: the English versions are AI-translated. Phrases like “elicits the kind of out-loud reaction you usually only get from...” carry the fingerprint. A native English reviewer pass is a future task
• No developer interviews of our own: we synthesize and link first sources, not generate them. This is a real limit. I want to fix it eventually

I think pretending these aren't there is dishonest. They are. They're here.

What this site actually is

Puzzlebyrinth is a magazine where AI writes a lot and a human curates. Slower than fully AI-automated content; faster than a fully-human magazine. We're running in that middle zone deliberately.

Some readers will find that “cheating.” I get it. So we leave every source visible, publish this editorial policy, and stay open about the AI part. That openness, I think, is the only way to earn back the trust we're asking for. A reader contact channel (Discord etc.) will follow.

Come back for the 8–16 monthly reviews, the Sunday digest, the occasional soundtrack piece, and Kizuki's designer studies whenever they show up.

— Komugi (editor-in-chief / indie game developer)