Executive Summary

The February review was a snapshot taken on March 2. Set that moment beside today (July 18) and the whole five-month trajectory fits in a single table. Published articles grew 6.9-fold and bilingual pairs 9.7-fold, and in between two new axes appeared: skills and hubs.

1.1The Monthly Trend

Output over the five months was uneven. March and April carried the sprint's momentum, and May dropped sharply to 62 articles. That breather was followed by June's peak burst of 234. What stopped in May was publishing, not work. Most hands that month went into architecture cleanup and building the quality pipeline, and that investment came back as June's throughput.

1.2Category Distribution

The center of gravity across 883 articles is tech. Tech — covering data quality and Physical AI — accounts for 530 articles, 60% of the total, followed by business at 175, centered on company analysis. Story (101) and art (77) were filled steadily by DataClinic diagnostic stories and data art.

February's three days were one person running batches across a stack of open windows. Forty-seven posts were split into eight batches and turned bilingual, common modules were extracted, and three new skills were built. It was intense, but the speed was tied to human focus. You cannot produce 883 articles that way. The core shift over five months was moving publishing out of a human sprint and into the daily work of a system.

The turn showed up most clearly in May. That month, with publishing down to 62 articles, most of the effort went not into visible posts but into the structure beneath them. The conflicts that arose when several tasks touched the single articles.json file at once were cut off by sharding, and the AI-prose review and SEO checks people used to do by eye became gates in the pipeline. June's 234 articles could not have appeared without that overhaul.

How this shift works is easiest to see in the making of a single article. The story of seven agents building one blog post follows data collection, visual analysis, story planning, writing, quality review, translation, and publishing as they pass down a relay of distinct agents. Once the work one person used to do was split by role, the same quality could be held while several pieces ran at once.

As output grew, structural bottlenecks invisible in February came into view. Over five months, four architectural changes untied them one by one. Sharding addressed concurrent-edit conflicts; splitting the source of truth from its mirror addressed the tangle of code and content; orphan rescue addressed broken related-article links; and the automated publishing engine addressed the limits of human hands.

3.1Sharding — Scattering the Single Hotspot

Every article's metadata lived in one file, articles.json. When several tasks registered new posts at once, they edited near the same lines and collided — and once a bad merge made 82 entries disappear. The fix was one sidecar file per article. New posts register in articles.d/<id>.json, and CI gathers these fragments to build articles.json. Because separate files never collide, the bottleneck of concurrent work vanished at the source.

3.2Source and Mirror — Keeping Code and Content Apart

The assets that run the blog have different natures. Content like articles and images is naturally edited directly on the live site, while code like skills, agents, and build tools has to be managed in one place to stay consistent. So the source of truth for code assets lives in a separate repository, and the live site mirrors it automatically. When a change lands in the source, a sync PR is generated for the mirror, and a person reviews and merges it. Content on the mirror, code in the source — this split keeps the two kinds of change from overwriting each other.

3.3Related-Article Rescue — From 209 Orphans to 16

Once there were 883 articles, 209 of them sat orphaned, connected to nothing. That means a reader has no path from one article to the next, and it is a poor signal to search engines too. Related articles are computed by tag similarity, but we added a reverse-guarantee rule so that no article stays fully isolated — every one appears at least once in another article's related list. Running that computation at build time cut orphans from 209 to 16. Ninety-two percent were reconnected.

3.4The Publishing Engine — Throughput Beyond Human Hands

The last piece was automating publishing itself. A pipeline that discovers what to write, drafts it, passes it through quality gates, and carries it to publication can now complete a full loop without human intervention. Each stage leaves a run record, so you can trace which article came out along which path. This engine was the real muscle behind June's throughput.

Left as a plain list, 883 articles are just scattered dots. Over five months the hubs that gather these dots by topic grew from 3 to 16. A hub is a bilingual collection page that pulls related articles under a single topic keyword. For readers it is an entrance to dig deep into one subject; for search engines it is an anchor that signals topical authority.

The Physical AI Datasets hub, for example, brings together articles about the physical-world data that robots and self-driving systems learn from. As individual articles that had been scattered were re-arranged around this hub, readers could move naturally from one piece to another on the same subject.

4.1Two-Way Linking — Articles and Hubs Point at Each Other

A hub's power lies in direction. Early hubs only gathered articles. There was a path from hub to article, but no path back from an article to the hub it belonged to. These five months added a backlink badge so each article points to its hub. Once articles and hubs pointed at each other in a two-way structure, a mesh emerged in which you can explore an entire topic no matter where you enter.

When output rises, quality is the first thing to wobble. One of the five months' decisions was to make quality checks a condition of publishing rather than an appendix to it. A gate every article must pass before it ships came into being, and skills run that gate.

5.1AI-Prose Review and the SEO Gate

Machine-written prose carries a distinctive tell: saying the same thing twice around a dash, closing every sentence in nominalized form, leaning on the same worn metaphors. The ko-prose-humanizer skill diagnoses and scores 11 such stylistic signals and restores the texture of the prose to something human, without touching the facts or the figures. Alongside it, seo-check automatically audits the four SEO layers, so articles with broken titles, descriptions, structured data, or heading structure are caught before they publish.

5.2References and Image Reinforcement

In articles with an academic or journalistic tone, sources are the ground of trust. The bibliography skill manages references in a standard format and automatically renders a references section, citation download buttons, and search-engine meta tags into the body. The image-reinforce skill finds context-appropriate images, verifies them, and inserts them so an article does not look empty on text alone. Quality moved from something resting on human diligence to something the pipeline guarantees.

5.3The Local LLM Serving Experiment

The quality work did not stay with cloud models alone. On an in-house greenhouse setup running Ollama, we loaded open models such as Llama, Gemma, and Qwen and ran experiments serving them locally with real measurements. The point is to confirm with data which tasks a local model handles well and which it does not — the boundary between cost and quality. It gives us the grounds to decide where the next step in publishing automation should sit.

If the first half of the year was the time we moved publishing into a system, the second half's task is to widen that system's quality and reach. Three directions of expansion are planned on the skeleton built over five months.

• Multilingual expansion. The first half began an EN backfill and a Japanese (ja) experiment. Beyond 435 bilingual pairs, the next goal is to make multilingual publishing — tuned to each language's search intent — the pipeline's default.
• Putting local serving to work. We apply the boundaries found in the greenhouse Ollama experiment to the real publishing path, deciding which stages to move to local models to cut cost and latency, on measured evidence.
• Recovering editing ease. Published HTML is still the live standard, but it is heavy to edit for prose alone. Work continues to refine the HTML-neutral manuscript that accompanies each published article, making post-publication edits easier.

The February review was the record of "128 made in three days." This piece is the record of "883 made in five months." What changed between the two reviews was not the size of the number but the way the number is made. If the sprint was a single burst, the past five months were the work of turning that burst into a daily, repeatable routine. Whatever the second-half review revises 883 to, the more important question then will still be the same: how light were the hands that made that number.

This piece is the next installment in the Pebblous blog review series. Read it alongside the earlier reviews and the related articles that show the changes of these five months.

• 1.Pebblous Data Communication Team. (2026). "The Birth of an Agentic Blog: Pebblous Blog 2026 Status Report." Pebblous Blog.
• 2.Pebblous Data Communication Team. (2026). "Pebblous Blog February 2026 Review: Content and Code Growing Together." Pebblous Blog.
• 3.Pebblous Data Communication Team. (2026). "The Story of Seven Agents Building One Blog Post." Pebblous Blog.
• 4.Pebblous Data Communication Team. (2026). "Physical AI Datasets Hub." Pebblous Blog.