stack-diagram: a Japanese university student's architecture-diagram tool, 5 days, 1858 stars

On Nov 1, 2026, a Japanese developer published a library, and 51 minutes later posted a three-line tweet with one image on X that got 240k views. The library is an agent skill for Claude Code, Codex and Cursor: name your tech stack in one sentence and it draws an architecture diagram with official logos, in both light and dark versions. 5 days, 1858 stars. But a bigger project in the same niche already had 90k stars, and its buzz is halving day by day. What's worth learning isn't the feature, it's that picture.

One image, three lines

The repo was created on Nov 1 05:35 (UTC), and the tweet was posted on 06:26, 3:30 PM Japan time. The author かる (@caru_ini) lists "INIAD 8th 28 graduate" in his X bio, which points to the 8th cohort of Toyo University's Faculty of Information Networking for Innovation and Design, graduating in 2028. In other words, a current undergraduate (inferred from the bio). His two earlier repos were a Discord bot template with 79 stars and a ComfyUI deployment script with 53 stars, and his regular tweets got only single-digit likes.

The full text of that tweet:

Built a Skill that gets AI to draw architecture diagrams nicely! Uses the official logos for everything. Supports dark and light mode 👍

("Made a Skill that gets AI to draw nice architecture diagrams! All logos are official ones. Supports dark/light.") The image is the one on the README's first screen: a user reaches a Next.js app deployed on Vercel over HTTPS; Next.js calls a Hono API in the same deployment; Hono queries Neon's Postgres, asks Clerk for the login state, and uses Stripe to take payment; GitHub Actions handles deployment. Each box holds that company's own logo in its original colors, not redrawn. Beneath the image is just one line of prompt:

Diagram our stack: Next.js and a Hono API on Vercel. Postgres on Neon, auth with Clerk, payments with Stripe. GitHub Actions deploys.

As of Nov 6, the tweet had 3635 likes, 259 reposts, and about 240k views.

What it is

An open-source (MIT) agent skill, installed in one line with npx skills add caru-ini/stack-diagram. You describe the stack in one sentence, and the agent looks up logos in order from svgl, devicon, and gilbarbara/logos; AWS, Google Cloud, and Azure use the icon packs each vendor publishes. The default output is a single HTML file with fonts and logos inlined, so it opens offline. It can also export PNG, a dependency-free React component, or a Claude Artifact directly.

The three most carefully designed details in the README: connectors route around labels, so no arrow runs through text; the agent renders the diagram and checks it before handing it over; and every diagram comes in light and dark versions, with box backgrounds kept white so the logos keep their original colors.

Why this one

The idea is not new. Skills that draw architecture diagrams with AI already exist, and they are big:

In this table stack-diagram is the small one. It runs on one thing: the screenshot looks good and is recognized at a glance. The row of Next.js, Vercel, Neon, Clerk, and Stripe logos is something every developer scrolling X recognizes, and together they look like a proper cloud architecture diagram. Rivals either use abstract boxes, a hand-drawn look, or require writing code. This is the only one that is "one sentence + real brand logos". On a timeline, the image gets reposted without needing an explanation.

Two more tailwinds: the agent skill trend is hot, and one command works across Claude Code, Codex, and Cursor. The use case is concrete. It's the diagram you put in a README, a proposal, or a slide deck, and anyone who sees it can immediately picture where they'd use it.

Where users come from

A few things worth noting: it wasn't on Hacker News, Reddit showed nothing, and it never made GitHub's official trending page. All of it happened on X. It started with the author's own native-language post. Trendshift's #4 daily ranking let AI news sites and multilingual "cool GitHub finds" accounts pick it up. The biggest second wave came from a Spanish-language KOL with no connection to the author. The author kept up: v0.1.0 through v0.4.0 in the first three days, and the sharing gallery was up by day two, giving the second-wave reposts something new to share.

Where it stands now

Daily gains of +779, +413, +249, +123, roughly halving each day. Last commit on 11-03. The gap between stars and actual use is large: skills.sh shows 1.1K cumulative installs, while the author's public gallery has 11-06 images total as of 5. Of 87 forks, one, Chulinuwu/stack-diagram, added click-to-inspect nodes and automatic system-theme following. No standalone clone has appeared yet.

In other words, people starred the image, not the tool. That's good news for anyone trying to earn stars: stars came from the screenshot, not the feature.

If it were you

Grade B. Don't build a second stack-diagram, and don't build a more comprehensive diagramming skill. archify already has 90k stars. What's worth copying is its approach: pick a narrow slot in "AI draws a diagram from one sentence," use real assets everyone recognizes to make a finished image you can drop straight into a README, and post a tweet with just the image and one line of prompt. The bar is low; a student with no name recognition pulled it off. The weakness is that the heat lasts about a week, and a big account could add an "official logo mode" and fill the gap at any time.

Who it fits: front-end developers who write agent skills and care a lot about layout and visual polish; people with a scenario they use daily that can become a good-looking diagram; non-English speakers actually have an edge, since Japanese, Korean, Spanish, and Chinese repost accounts will pick it up.

Who it doesn't fit: anyone aiming to build an all-purpose diagramming tool that competes head-on with archify or diagram-design; anyone whose output isn't good-looking enough and needs a paragraph of explanation; anyone counting on stars to bring real users and long-term maintenance motivation. That group converts stars to use at a very low rate: 1858 stars for 5 gallery images.

How to build it: change the scenario, not the approach. Candidates: ER diagrams with real database and ORM logos; data-flow or integration diagrams with real SaaS logos (like Segment or Zapier); user-journey diagrams with official app icons; organization or investment-relationship charts with company logos; sequence diagrams. Three things are non-negotiable: use real brand assets and don't recolor them; have the agent render and check the output before delivery so no connector runs through text; ship both light and dark versions. The README's first screen should show a finished image plus the prompt that generated it, with no architecture talk.

How to get users: launch on your native-language X account with a three-line caption and one image, posting the day the repo is created. Aim for Trendshift's daily top spots, since the top few get picked up automatically by AI news sites and multilingual repost accounts. Ship something new the next day (a gallery, a new format, a new cloud provider) to give second-wave KOLs a reason to repost. Don't count on Hacker News.

Deadline: the next few months, while agent skills are still hot. Once skill directories are crowded with similar tools and big accounts like archify have absorbed every niche mode, a good-looking image alone won't be enough.

Ways this call could be wrong: (1) similar "one-sentence diagram" skills that appear in the next month, with good-looking sample images, fail to reach 1000 stars in a week, which would mean stack-diagram rode a one-time spike; (2) archify or diagram-design adds an official-logo tech-stack mode and stack-diagram's daily gains drop to zero; (3) the reverse: if someone uses the same approach on a different diagram and gets thousands of stars, the path is real, and you should move fast.

Data and sources

• GitHub: caru-ini/stack-diagram README and Releases; daily stars and forks (recorded from 11-02 to 11-06; the record for 11-02 is an earlier snapshot from that day).
• X: @caru_ini's launch tweet 2026-11-01 and the following posts (like and view counts read via 11-06); posts from @MoureDev 11-02, @0xbobaaa, @Ryrenz, @GitTrend0x, and @GitHub_Daily. Author identity from their X bio.
• Trendshift (trendshift.io/repositories/331355); reveneau.com coverage 11-02 (486 stars, daily rank #4); news sites such as mindpattern.ai.
• skills.sh install count and stacks.caru.moe gallery count: read 11-06.
• Competitor star counts: GitHub, 11-06.
• Instagram: Reel from @tunestechtips 11-04.