A physics professor with only 1,196 followers, one post drove 90 stars to a figure-drawing skill: what schematic-figure can learn, and what not to follow

schematic-figure is a Claude Code / Codex skill released by a Yale applied physics lab, letting AI draw paper schematics with code as editable SVG/PDF. It gained 171 stars in 23 days, more than half from one X post by its advisor. The star count is modest, and not because it was first: Sakana had a similar one the same day, and a 56k-star collection sits nearby. What's worth learning is how it was built as a skill and how it was launched. Trying to chase stars in this same direction isn't worth it.

A post from Nov 2

On Nov 2 2026 in the afternoon (UTC 16:04), Logan Wright, assistant professor of applied physics at Yale, posted on X:

Would you like to make it 100x easier to make publication-level schematic figures to explain your research? @JinchenZhaoo in my lab developed a skill for Claude that has been *life-changing* for this. Not full automation. But legitimately amazing too.

He works on optical computing and physical neural networks, with over nine thousand citations, but has only 1,196 followers on X and posted only this one original post in 30 days. The post ended with 889 likes, 84 reposts, and 46,754 views, plus 1,236 bookmarks, more than likes. More bookmarks than likes means people wanted to keep it for later use.

Someone in the replies asked whether "100x" was an exaggeration. He replied: "100x or 10x, depends how good you were at drawing before. For someone like me who can't draw but has seen lots of figures and knows what they want, that's understating it."

That line is this project's user profile.

What it actually does

The first figure in a paper (Fig. 1) is often a schematic: the experimental setup, a molecular structure, a method flow. Readers look at it first, but most researchers were never taught to draw one. The skill is a folder: a set of instructions for the AI (SKILL.md), a drawing library, a 3D mini-engine, optical-path and molecule tools, a Blender rendering kit, and automatic checks. Install it in your Claude Code or Codex, say "use the schematic-figure skill to draw Fig. 1 for this paper," and paste in the paper text.

Compared with existing options, it differs in two ways:

• It outputs vectors and source, not raster images. You can move labels in Illustrator / Inkscape, or have the AI change just one part. A commenter reminded the author that "you need an editable version so labels are easy to fix before publication," and this skill does that by default.
• It goes beyond boxes and arrows. It covers physics, chemistry, biology, materials, optics, and patent figures, and can output 3D and photorealistic renders. The blog says that without the skill, Opus 5.5 mostly drew flat boxes and arrows held up by text; with it, it draws scenes and does 3D. The cost is about 4 times longer per figure, averaging under 20 minutes.

How the stars came

Data: daily star counts from trendshift.io; daily GitHub star records (11-06 is 169, and GitHub showed 171 that day). The two sites differ by one day because they use different time zones.

Each of the two rises matches one post, with a flat line in between. It never made GitHub Trending, and I couldn't find discussion on Hacker News or Reddit. In the days after the post, 5 new forks appeared, which lines up with the timing.

So this isn't organic growth from "more and more people using it." It's two pulses from two launches.

Why one post was enough

The person speaking is the user. A physics professor saying "I can't draw and this is life-changing for me" is far more believable than a developer saying "we built a powerful tool." He doesn't have many followers, but researchers forward things like this to colleagues.

A comparison picture shows the difference. The top of the blog has a three-way comparison of original / without skill / with skill. The 5 papers chosen are all under CC BY licenses, and 4 of them were published after the model's training cutoff, which heads off the "did it memorize them" objection. Someone in the replies (Jess Riedel) asked how the comparison was made. The author said the model never saw the original figure or caption, only body text with image-related wording removed.

The blog reads like a paper. 100 papers, 7 fields, training and test sets kept separate. Scoring used a Turing-test-like "looks published" metric: an AI judge looks at a figure and estimates the probability it is a formally published Fig. 1. They first verified that the score separates real figures from AI ones, then revised the skill based on the judge's feedback on the training set, iterating about 15 versions with Claude Code's dynamic workflows, stopping when the score stopped improving, and running the test set only once. The limitations are stated candidly: the author is not an expert in most fields, and results are clearly worse for new concepts no one has drawn before. Researchers in this community take that approach well.

It caught a trend. The blog's first paragraph says the idea came from the viral Opus 5.5 video on X and links to @claudeai's official post. In those weeks everyone was talking about "models can draw now," and this offered a concrete scientific use.

The replies were mixed. Sanjay Tyagi ran it on a text description from one of his own papers from 2019 years ago and said "scientifically accurate, just not that cool." The author replied to add "in 3D" and "use Blender for realistic materials." An organic chemist pointed out that a few chemistry drawing conventions were wrong, and the author said he would add notes to the chemistry section. Others were skeptical: "one arrow pointing the wrong way changes the scientific conclusion, even if the figure looks publishable."

Who it's competing with

There are also several scattered small author projects (patent figures, grant application illustrations, vector reconstruction) with similar ideas, 0–1 stars each.

This space isn't short on builders. The big collections take most of the stars for "scientific AI skills," and PaperBanana holds CS method diagrams. A standalone figure skill alone gets only 123 stars, even for a name like Sakana. schematic-figure did better than that because of the post and the blog, not because the space was empty. So far I haven't seen anyone build on it: the 16 forks are mostly straight copies, and one just renamed it after forking.

If it were you

Verdict: Learn its method, don't chase its niche. "AI-drawn scientific schematics" already has a 56k-star collection and PaperBanana with 7 thousand stars. Another standalone figure skill would get a hundred or two at best.

Who it suits: People who are already in a specialist field (research, patents, circuits, architecture, medicine), who know what that field's figures should look like, and who can run Claude Code workflows. Ideally someone in that field with real standing will repost for you.

Who it doesn't suit: People who want to build a big, do-everything "AI drawing platform" to chase star counts; people outside any vertical community who can't get real users to endorse them; and people who expect one launch to keep stars growing. This project stopped committing code after it was created, and its stars stopped too.

How to do it: Copy its workflow, change the figure type. Pick a diagram that professionals draw every day and get wrong (patent drawings, circuit diagrams, textbook illustrations, clinical workflow charts, academic posters). Define a metric that can be scored automatically (e.g., "looks like it was formally published"), prepare separate training and test samples, have an agent revise the skill version by version, and stop when the score stops improving. Output must be vector or code, not raster. You can also fill gaps it left: chemistry and biology drawing conventions, or a "scientific correctness check" where one arrow's direction can't be wrong. That is where the replies pushed hardest.

How to get users: The launch kit has three items: a three-way comparison image of original / without skill / with skill, using samples from after the model's training cutoff; a long blog post covering method, evaluation, and limitations; and one real user saying "this helped me a lot" on X or LinkedIn. Then submit to skill directories. This project spent nothing on paid acquisition. Two posts and one blog got 171 stars.

When to do it: Not urgent, but don't wait too long. This category of schematic is already taken. Other professional figure types are still open, and once the big collections absorb them, a standalone skill will have even less room. Each major model capability upgrade (like the Opus 5.5 moment) is a good time to launch on the wave.

What would show I'm wrong: ① schematic-figure gained over 1,000 stars in a month with no new posts, showing demand for a single vertical skill is bigger than I think, so building a similar one is worth it directly; ② it shipped v2 and posted again and gained another hundred-plus stars, showing that "good content + real user endorsement" can be reused; ③ a big collection or Anthropic / OpenAI shipped a dedicated scientific-figure feature with similar quality, which would close the room for standalone skills.

Data and sources: GitHub repo page and daily star counts (10-17 to 11-06); trendshift.io/repositories/306131; blog lgwrightlab.github.io/schematic-figure; X post @LoganGWright1(11-02) and its replies, @JinchenZhaoo(10-15); two LinkedIn posts; skillsllm.com; competitor star counts from their respective GitHub pages (11-06); BioRender Leo see BusinessWire 2026-10-06.