596 Agent Commits, All Approved by Humans: Inside a One-Person AI Editorial

Ask an AI content team how they work and you get a shopping list of models and tools. AICU media, a Japanese editorial run by the creative-AI community AICU, answers differently. Their pipeline is a daily blog, a weekly digest, a monthly magazine, and a string of commercial books, all produced from one Markdown repository on GitHub. The models inside it changed four times in nine months. The process changed zero times.

That mix deserves a closer look, because most AI content operations get the ordering backwards: they optimize the model first and the workflow second. AICU media did the opposite, and their published numbers are concrete enough to copy from.

What kind of editorial runs on a daily blog

The operation started on October 19, 2023 with a simple commitment: publish one blog post every day. AICU's own counter reads 1,017 posts across 1,015 days, an average of 1.002 posts per day. Plain calendar arithmetic from that start date to the end of August 2026 comes out longer, about 1,047 days, so the outlet is clearly counting on its own convention; on either count the cadence lands at or just under one post per day. That near-daily rhythm matters more than it looks like it should, because almost-daily is the cadence a real media habit produces, and it is the raw material for everything above it.

AICU media describes itself as a three-story building:

• The daily blog is the ground floor. Fast information, published the same day.
• A weekly email newsletter digests the week.
• A monthly magazine, 27 issues in 29 months, re-edits the blog posts into themed features.

The key rule is that each floor is built from the one below it. The magazine is not written from scratch. The editorial team takes posts that already exist, changes the order, rewrites the headings, and regroups them by theme. The source post compares it to a DJ set: the tracks already exist, and the order decides whether the night works. By the end of the year the accumulated material becomes a commercial book, which is the fifth exit and the slowest one.

That alone is a publishing model worth stealing for anyone who generates content faster than they can organize it. But the interesting machinery is one level down.

One manuscript, four formats

Every issue of the magazine starts as a single Markdown manuscript stored in GitHub. A build step turns it into four simultaneous outputs:

Markdown manuscript (managed in GitHub)
        | pandoc + LuaLaTeX (pinned Ubuntu 22.04 container)
  ┌─────┬─────┬─────┬─────┐
 EPUB   Print PDF   Dist PDF   Web
(e-book) (press)  (direct sale) (HTML)

The payoff of a single source is the obvious one: fix an article once, and all four formats update at the same time. Corrections after publication run through the same path. When an editor adds or edits a post on note.com, GitHub detects the change, the editorial pipeline rebuilds all four forms, and the finished files land in distribution storage and get posted to Slack and Google Drive. The human eye is reserved for exactly three things: the cover, the print halftones, and the bleed.

The production economics are the impressive part. The first issues took about 3.5 person-months of work each. The current setup, including writing, runs at 0.5 to 1 person per month at eight hours a day, producing a 150-page issue every month. That is a magazine maintained as a side project, and it is only possible because layout, typesetting, and format conversion stopped being human tasks.

596 agent commits, all approved by a human

The editorial uses LLM-based editing agents for routine work: normalizing article formatting, adjusting headings, placing figures, unifying terminology. Every agent commit is reviewed as a diff and approved by a human before it merges. The latest issue, published in August 2026, integrated 29 agent commits, and all 29 were human-approved.

Over nine months the totals look like this:

• December 2025: 17 commits with Claude Opus 4.5
• January to March 2026: 231 commits with Claude Opus 4.5 and 4.6
• April to June 2026: 183 commits with Claude Opus 4.6 and 4.7 plus Sonnet 4.6
• July to August 2026: 165 commits with the Claude 5 family

The number that got my attention is not 596. It is the row labels. Over those nine months the team swapped model generations four times, and the editorial process never changed once. The tools turned over every quarter and the workflow stayed identical, which is the opposite of how most teams run. When a team falls in love with a specific model, every upgrade becomes a rework of prompts, integrations, and expectations. When a team builds process and state management instead, the model is just a swappable worker.

There is a second detail in the same commit log: large flagship and security-oriented models are not the ones that edit best. The team ended up using different model tiers for different editorial tasks, and the commit history shows who fixed what and how, which makes the log itself a documentation layer. Every correction stays traceable, like a red pencil that never runs out of ink.

The part they refuse to automate

Here is the design choice I keep coming back to: the cover is deliberately not automated. The editorial describes it as outside the scope of automatic typesetting, separated out as a human design task.

The process works through a monthly cover contest. Creators submit artwork around 2,000 by 3,000 pixels, rights get cleared, and the community plus judges curate the winners. A human then lays out the cover in Illustrator. The format follows the standard dimensions recommended by the digital distributor, with extra margin on the sides because the actual print size, JIS B5, is wider than the recommended frame. That kind of boring fine-tuning is exactly what the pipeline cannot judge, and it is why the cover stays outside it.

The back cover is different: it runs in Canva because advertisements swap frequently and several editorial members need to edit it together. Lightweight beats automatic when the content is ephemeral.

The reasoning generalizes nicely. Automation should eat the jobs with a stable shape. The cover is a design decision with a taste component and a legal component, so it gets humans. The back cover is a repeating template with fast-changing content, so it gets a drag-and-drop tool. The choice is per-task, not a blanket policy, and that is rarer than it should be.

Where an AI character gets its identity

The editorial staff now includes an AI character, Marsha Arancia, born from the same contest pipeline as the covers. Her designer won an AI character audition with an artwork called "The Way Home from Buying a Book," and the reasoning behind the design connects directly to how image models handle character consistency.

The designer described prompts as design blueprints for a character rather than command sentences, and tested the identity by generating versions where every attribute changed: boyish, girlish, youthful, middle-aged male, lady, elderly. Whatever stays constant across all those variants is the essence of the character. The attributes are settings. The constancy is the identity.

This is the same test that character-consistency workflows run on image models, and it transfers cleanly: if a character only works when every attribute is pinned, you are preserving a configuration, not a character. If the character survives a change of age, outfit, and context, you have found the thing worth keeping. When a character is generated with a full prompt each time, the parts of the prompt that encode identity are indistinguishable from the parts that describe the scene. AICU's approach separates them on purpose, which is the same impulse behind reference-based consistency systems in image generation (we have covered that angle separately).

The limitations they printed themselves

The paper the editorial submitted about its own process is refreshingly honest about what it does not measure. There is no quantitative evaluation of information freshness, no metric for the time between an article's first publication and its appearance in the magazine, no before-and-after comparison of editing workload, and no breakdown of how many agent proposals were accepted unchanged versus accepted after modification. The commit history is not public yet, so an outsider cannot independently verify the claims.

I like that paragraph because it is the opposite of the usual AI-blog bragging. The editorial states exactly what evidence does not exist, which makes the evidence that does exist easier to trust. The commit log is the raw data, and every claim about the pipeline is, in principle, checkable against it. That is more than most tooling vendors publish about their own workflows.

What transfers to your own pipeline

The specifics are Japanese-media shaped, but five rules transfer to almost any AI content or image-generation operation.

One source of truth beats five export buttons. Any time the same content needs to leave in multiple formats, generate once and transform. The moment you maintain parallel versions, corrections multiply in cost. Markdown to PDF, EPUB, and HTML is a solved problem, and the fixed container keeps the toolchain from rotting.

Gate agent output with review, not with hope. Every agent commit in AICU's pipeline is diffed by a human before it merges. That single rule is what makes the entire automation safe enough to run on a monthly production schedule. No silent auto-merge, no trust-the-model fast path. As workloads grow, this becomes the main thing preventing your archive from slowly filling with confident garbage.

Keep one demanding task manual on purpose. The cover is the publication's face, and it is the one step nobody automated. A process where every step is automated has no place for taste. Pick the output that is judged by taste and keep a human in that loop; it will also tell you what the AI parts are actually good for.

Design against model churn. AICU's pipeline outlived four model generations because it never depended on a particular model's quirks. Prompts and integrations layer on top of the process, never inside it. When your workflow survives a model swap without edits, you have stopped renting your process from a vendor.

Keep the history that lets you verify. Their commit log doubles as documentation and as evidence. You cannot improve a workflow you cannot replay, and you cannot defend a workflow you cannot show. Version control is a colophon: it records not just what was made, but who made it and what they changed.

The larger point is the one Marsha Arancia makes at the end of the source article: AICU media is not really an editorial office that produces articles. It is an editorial office that produces people who produce. The daily blog feeds the magazine, the magazine feeds the books, and the books feed the next round of creators, including AI ones. However much you automate, the loop stays human at the top. A process that survives four model generations without changing is not a betting on AI. It is a bet on the process itself.