A design engineer turned his interface craft into 14 AI skills, and in four months they reached 11k stars: the stars follow his posts, not the content
What it is A set of agent skills written in Markdown, split into UI, typography, colors, layout, accessibility, and copy. They get AI to refine an existing interface to a design engineer's standards, plus small tools like exploded-view diagrams and component stress tests.
Why it took off The author was already known for writing about interface details, and his X account, with 25k followers, posts a new skill almost every week along with before-and-after videos. Add a user showcase with 1.35M views, Perplexity listing /teardown, and Chinese weekly-chart roundups, and it all landed in the middle of the "teaching AI taste" wave.
CHow you could follow Stop building generic design skills. Pick a narrow area you truly know and have written about (charts, motion, forms, screenshots, Chinese typesetting), build a repo around it, ship one new skill every 1–2 weeks with a before-and-after video, and add the self-check scripts it lacks.
- Data (updated daily) · Repo created 2026-07-10 · 11,460 stars · 7 days +1,876
GitHub · stars Data 2026-07-30 → 2026-11-08 (102 days, 14 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
jakubkrehel/skills is a set of agent skills from Slovak design engineer Jakub Krehel. Once installed in Claude Code, Codex, or Cursor, they get AI to refine existing interfaces to his standards: corner radius, typography, colors, layout, accessibility, and copy. Built in July 10, by November it had 6 11k stars and 350k installs on skills.sh. The work is detailed, but several projects with 100k+ stars sit in the same category. It took off because of the author's years of accumulated reputation and an X account that posts every week. For open-source maintainers hoping to get stars, the thing to learn is the cadence: write articles first, turn the articles into skills, and ship one a week. Without the reputation, copying it won't get the same stars.
A CRM back office built in two days
In Sep 2026 14, designer Marcel Kargul (57k followers on X) posted:
"Built a whole CRM back office with Next.js in about two days 😍 Claude helped, plus the two best skills:"
The two links below are Jakub's better-ui and Emil Kowalski's emil-design-eng, plus a live demo you can open. The tweet ended up with 1.35M views, nearly 5,000 likes, and 12k bookmarks. Bookmarks were more than double the likes, so people were saving it as a tool.
Jakub didn't post this himself, but his repo got the most traffic from posts like this. By Nov 6, the repo had 10,988 stars (GitHub daily star count). Over the last 7 days it gained 1,848, and its single-day peak was 407 (11-02). It never once made GitHub Trending, so nearly all the stars came from outside the site.
What it is
An MIT-licensed Markdown repo, installed with one line: npx skills add jakubkrehel/skills. It has two groups:
- Refinement skills split by discipline: better-ui, better-typography, better-colors, better-layout, better-accessibility, and better-writing each cover one area, and better-interface ties them together, producing a prioritized change list for the page you have. It only edits existing code and uses the design tokens already in your project, without inventing a new theme. A Nov 3 comparison of 13 design skills (setproduct) called it "the broadest in discipline coverage."
- Small tools that make good videos: /break stress-tests a component against edge cases; /state-machine lays out all of a component's states on a temporary page; /variant generates several variants; /teardown draws an exploded technical diagram of any object, with editable colors and annotations, exported as SVG or PNG.
The author explained why he built it back in Sep 3:
"These were originally for myself. I know exactly how certain things should be done, but with every new project I had to explain them all over again. ... Skills have a bad reputation, and I understand why, since a lot of skills are just junk."
People have also pointed out the weak spot: it gives agents rules but no detection scripts or render checks, so whether an agent reads and follows them is left to chance. Several higher-starred projects in the same category (impeccable, hallmark) ship with checkers.
Where the stars came from
First, the author is himself the authority on this. Jakub has been known for writing about interface details for a while. In August 2025, his article on OKLCH colors hit the front page of Hacker News with 859 points. Starting in May this year he ran a web magazine called Interfaces, one issue on typography, one on accessibility, and in October he said subscriptions had passed his 500-person goal. He is now founding design engineer at Interfere (YC S25, which announced a $5.1M seed round in October), and before that he was at OpenSea. So the pitch of "putting a top design engineer's taste into your AI" is believable to people. The rules in the skills are the things he has already published in his articles.
Second, his own X account posts once a week. With 25k followers, from July to November he posted almost every week a "new skill + before-and-after video," and every post ended with the same install command (data for the author's posts comes from X, 11-06):
| Date | What was posted | views |
|---|---|---|
| 07-12 | Launch post: "encoding judgment into agents with skills," starting with 3 skills | 240k |
| 07-24 | the /better series, plus 4 more | 676k |
| 07-28 | before-and-after for last week's 4 | 588k |
| 08-20 | /explain-interface: "I used to take apart other people's pages with DevTools, now I can use this" | 41k |
| 09-03 | "Skill installed 160k times" | 220k |
| 10-11 | /state-machine demo | 47k |
| 10-25 | /teardown exploded view | 245k |
Between these posts he also shared new magazine issues, interface tips, and an "interface cheat sheet," so the account had content throughout, and skill posts were just one kind. By Nov 6, he said all the skills combined had passed 15k stars and 425k installs, up from 160k two months earlier.
Third, others amplified it. One was the Marcel post above. Two was Perplexity: on October 29 and 30, Perplexity's Computer added /teardown to its skill marketplace, and CEO Aravind Srinivas reposted it personally, free to use until Nov 6 (whether Perplexity paid anything is unverified). Three was the Chinese and Japanese communities: from October, 陈发财 and the Japanese Skills Studio reposted it one after another. In November, on 1 and 2, the Chinese X account @slashui posted "This week's top ten GitHub repos, half teaching AI how to have taste," and @GeekCatX named "jakubkrehel/skills +1,272 this week." On 11-02 it added +407, its highest single day on record. The timing overlapped with the Chinese weekly chart and Perplexity's free window, and it's unclear how many stars each contributed.
Fourth, timing. By 2026, having AI write front-end code was already common, and "AI look" (purple-blue gradients, Inter everywhere, cards nested in cards) had become a recognized flaw. The "teaching AI taste" category was rising across the board. In the week of 10-25–11-01, four of the top ten on GitHub's weekly chart were skills like this.
Not found: paid promotion, Product Hunt, big Reddit threads, TikTok, and no coverage from mainstream tech media.
Who else is doing the same thing
| Project | Scale (early November) | Compared with alternatives |
|---|---|---|
| frontend-design in anthropics/skills | The whole repo, 188k stars | The official one, a general aesthetics skill |
| nextlevelbuilder/ui-ux-pro-max-skill | 146K stars | A searchable design knowledge base that only gives advice and doesn't edit code |
| pbakaus/impeccable | 104k stars, 8.8k added per week | Generates pages from scratch, with 62 detectors and CI support |
| emilkowalski/skills | 54K stars | The closest rival, Emil (author of Sonner and Vaul), focuses on motion |
| Nutlope/hallmark | 49k stars, 6.7k added per week | Builds landing pages that don't look AI-made, with themes and render checks |
| 花叔 huashu-design | 29K stars | One sentence produces a whole piece (web pages, PPT, animation) |
| ibelick/ui-skills | 10.7K stars | A directory site for design-engineer skills of similar scale to his |
| heydamir/unslop | 9.6k stars (one week after creation) | A single skill dedicated to removing the AI look |
(Star counts from GitHub, 11-06)
It's not the first, nor the biggest. ibelick did it in January, Emil in March, and the author's previous repo, make-interfaces-feel-better, was March too. Star counts in this space differ by an order of magnitude. The setproduct comparison had a line: "Star counts reflect marketing more than quality." The category is crowded: unslop was created in Oct 29 and hit 9.6k stars in a week; various "taste" and "anti-slop" skills reach tens of thousands of stars within a week; his make-interfaces-feel-better was copied wholesale into everything-claude-code, and dozens of skill directories are reposting it.
If it were you
One judgment: don't copy the "generic design skill" play. Learn the rhythm instead. Behind these 11k stars are a year-plus of the author's articles, a magazine, and an account that posts consistently, which followers can't get quickly. The category already has four or five projects with 100k+ stars, and new competitors keep appearing about once a week.
Who it suits: developers and designers who have already published work on one specific area and are recognized for it. For example, people who have written a lot about motion, charts, forms, email templates, App Store screenshots, or PPT layout already have ready-made "judgment" to encode.
Who it doesn't suit: people with no work or articles of their own who want to redo "making AI have taste" from scratch, or who want to port this skill set to another framework or language. Forks, collections, and directory sites that copy it wholesale already number in the dozens, so the market doesn't need one more.
How to do it:
- Pick a narrow topic where AI output is clearly bad and you genuinely know the subject. Generic UI and landing pages are saturated. Charts, motion details, forms, mobile screenshots, and Chinese typesetting still have room (inference).
- Articles first, skills second. Each rule should map to an article or example you've written, and the README's first screen should show the article and install-count badges, linking to each other.
- Keep one repo and ship new skills regularly, adding one every 1–2 weeks. Don't open a new repo each time, so stars and installs accumulate in one place.
- Build one or two small tools that look great as a video (like /teardown or /break) specifically for spreading and for getting picked up by platforms.
- Fill the gap it leaves: ship a check script that takes its own screenshots and measures alignment and contrast. That's where the higher-starred peers pull ahead.
How to get users: post a weekly "before-and-after video plus one install command" on your own X account; send your work to big accounts that share usage (people like Marcel, who posted right after finishing a project with it); watch the Chinese and Japanese "GitHub weekly chart" roundups, since they need material every week; submit to skills.sh and other skill directories. When installs hit a round number, post a milestone.
Timing: the "teaching AI taste" wave is peaking now, and models' own taste is also improving. The window for this kind of skill is probably a few months (inference), so shipping before year-end is the safer bet.
Signs you've misjudged: if four straight weeks of posts with video each stall at a few thousand views and the stars stay in double digits, the topic isn't painful enough or your reputation isn't big enough yet. Go back and write articles to build an audience first. Conversely, if over the next month a similar project from someone without a big name keeps growing, the category's demand hasn't been fully captured, and a narrow topic can go live right away.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one frontend-design in anthropics/skills Big-company features 2025-09 188k stars (whole repo) Official general-purpose frontend design-taste hint, one skill; it splits into 14 by discipline and ships with review reports nextlevelbuilder/ui-ux-pro-max-skill Open source 2025-11-30 146K stars A searchable design knowledge base that only gives advice and doesn't change code; it edits your existing code directly pbakaus/impeccable Open source 2025-11-16 104k stars, +8.8k this week Generates pages from scratch, with 62 detectors and CI support; it's rules only, no linter emilkowalski/skills Open source 2026-03-16 54K stars The closest rival in the same niche; Emil focuses on motion. It covers more disciplines: typography, color, accessibility, copy, and more Nutlope/hallmark(Together AI) Open source 2026-04-27 49k stars, +6.7k this week Builds marketing landing pages that don't look like AI output, with theme and render checks; it targets apps that already have a design system alchaincyf/huashu-design Open source 2026-04-19 29K stars Turns one sentence into a finished piece (web / slides / animation); it doesn't build from scratch, it polishes ibelick/ui-skills Open source 2026-01-07 10.7K stars Also a directory site for design-engineer skills, about the same scale; its discipline split is more systematic heydamir/unslop Open source 2026-10-29 9.6k stars (this week) A single skill for removing AI taste; it's a whole suite - Where users come from
- Verified The author's own X account (25k followers), publishes new skills + demo videos almost every week: 07-12 launch post 240k views, 07-24 676k, 07-28 588k, 09-03 220k, 10-25 /teardown 245k (TikHub live)
- Verified A big account posting usage on its own: @marcelkargul 09-14: built a CRM in two days with better-ui plus emil-design-eng, 1.35M views, 12k bookmarks (x-radar)
- Verified Platform listings / directory sites: Perplexity Computer listed /teardown in 10-29, reposted by AravSrinivas, free for the first 11-06; skills.sh shows 352k installs; also listed on ui-skills.com, craftwork.design, agenticskills, Dojo, tessl and others
- Verified Chinese / Japanese repost accounts and weekly chart roundups: Chen Facai 10-09, Skills Studio 10-23, Captain Luo 11-02 "Top 10 on GitHub this week, half of them teach AI to have taste", Zhishi Mao 11-01; 11-02 +407 in a single day is the peak of the whole period
- Verified Long-term content assets (personal site + Interfaces magazine): the OKLCH article hit HN with 2025-08 points; the Interfaces magazine hit HN with 2026-05 points 859 100 ; the README carries badges linking the magazine and skills.sh to drive traffic to each other.
- Inferred Company backing Interfere (YC S255, $M seed round): the October funding and design system launch were shared widely by the founders and colleagues; direct contribution to the stars not verified
jakubkrehel/skills: a design engineer turns his interface craft into agent skills
Channel: GitHub. Research date: 2026-11-06. Data sources: GitHub repo database (repo_profile / repo_day), GitHub API, X live (TikHub, via the agent-im machinex user/x live), x-radar, skills.sh, setproduct.com, Jina search.
One line
Jakub Krehel (Slovak design engineer, formerly at OpenSea, now founding design engineer at Interfere (YC S25, raised $5.1M seed led by October)) wrote the interface details from his personal site jakub.kr and his self-published magazine Interfaces (radius, optical alignment, shadows, type, OKLCH color, accessibility, copy) into 14 agent skills that install into Claude Code / Codex / Cursor: better-ui, better-typography, better-colors, better-layout, better-accessibility, better-writing, better-interface (the overall review), plus interface-review, explain-interface, break, state-machine, variant, build-design, teardown.
Verdict
It took off because of three things:
- Timing: by 2026 agents were already writing frontend code everywhere, "AI taste" (purple-blue gradients, Inter everywhere, cards inside cards) became a recognized pain point, and design skills that "teach AI taste" grew across the whole category. In the week of 10-25 to 11-01, four of the GitHub weekly top ten were skills of this type (weekly roundups by @slashui and @GeekCatX).
- The author's own X account is the main channel, and it keeps posting: 25k followers, and from July to November he posted almost every week a "new skill + before/after video", several of which passed 200k views. Each star acceleration lines up with one of his posts.
- The author is himself an authority on this: he was already known for writing about interface details (his OKLCH article hit HN with 859 points in 2025-08, and "Details that make interfaces feel better" was widely shared), so people believe the claim that you can install a top design engineer's taste.
Its differentiation: it has the most complete discipline split, and it only polishes existing code and respects the project's existing tokens instead of inventing a theme; add playful skills like break, state-machine and teardown that are easy to turn into videos, which spread naturally. The differentiation is indeed why it was picked up (a comparison review calls its "discipline coverage the broadest"). But its stars are an order of magnitude below the rivals (11k, versus Emil 54k, impeccable 100k), which shows product quality alone can't carry the attention. setproduct's own words: "stars say more about marketing than quality".
Data
- Repo 2026-07-10 (July merged the previous 3 repos into this one; see the setproduct review). Stars: 07-30 1,933 → 10-12 7,832 → 10-28 8,837 → 11-06 10,988 (github repo_day); the GitHub API shows 11-06 live at 11,196 stars and 432 forks.
- Last 7 days +1,848; single-day peak +11-02407 (GitHub repo database). Has never made it onto GitHub trending (trending_days_any=0).
- Installs: the skills.sh page shows 20 skills with 352k total installs (better-ui the highest at 44.5k). On X, the author 11-06 said that his skills together have "over 15k stars and 425k installs" (that count includes the make-interfaces-feel-better repo at 3,925 stars, oklch-skill at 280 stars, and loading at 367 stars). 09-03 he was talking about 160k installs, which more than doubled in two months.
- Author's X account @jakubkrehel has 25,449 followers (TikHub 11-06); GitHub 1,132 followers.
Growth timeline: every acceleration lines up with one of the author's posts
| Date | Event | Exposure |
|---|---|---|
| 2026-07-12 | Launch post "encode your judgment into agents with skills" + npx skills add jakubkrehel/skills | 2,893 likes, 240k views |
| 07-15 | Quote-tweet with his own addition | 130k views |
| 07-24 | "/better series adds 4 new skills" | 2,293 likes, 676k views |
| 07-28 | Before/after video | 4,499 likes, 588k views |
| 08-11 | /interface-review | 80k views |
| 08-20 | /explain-interface | 41k |
| 09-03 | "Skill installed 160k times" + explains why he built it (it was originally for his own use) | 3,914 likes, 220k |
| 09-14 | @marcelkargul (57k followers): built a CRM admin backend in two days with better-ui plus emil-design-eng | 4,988 likes, 12k bookmarks, 1.35M views (x-radar) |
| 09-17 | Author's post | 95k |
| 10-08 / 10-11 | /state-machine and demo | 77k / 47k |
| 10-09 / 10-23 | Chinese @Chenfacai0, Japanese @sumika45379 (Skills Studio) repost introductions | 6k / 20k |
| 10-25 | /teardown: generates an interactive technical exploded view of any object, exportable as SVG / PNG | 6,090 likes, 245k views |
| 10-29/30 | Perplexity Computer listed /teardown in its own skills marketplace, personally reposted by AravSrinivas, free for the first 11-06 | @AskPerplexity 22k, Aravind 19k |
| 11-01/02 | Chinese weekly roundup: @slashui "Top 10 on GitHub this week, half of them teach AI how to have taste", @GeekCatX names "jakubkrehel/skills +1,272 this week" | 11k / 3k; 11-02 +407 in one day, the highest of the whole period |
| 11-03 | setproduct published "13 design skills compared", listing it as the representative of "polished by discipline" | — |
| 11-05 | Chinese accounts @vikingmute, @dngzsn37461 recommend teardown | 6k / 4k |
Sources: the author's posts come from TikHub live pulls of x user jakubkrehel --days 130; other people's posts come from x live jakubkrehel --days 30; the Marcel post comes from x-radar.
Note: before 10-28 we only had two points, 07-30 and 10-12; there is no daily curve for the July wave. GitHub's stargazers endpoints (REST and GraphQL) return empty for this repo, so per-day star timing can't be obtained, and matching "post → stars" can only be checked day by day after 10-28.
Where users come from (acquisition)
- The author's own X (confirmed): since July, 40 posts, each carrying
npx skills add jakubkrehel/skills, in a fixed format of "new skill + short video demo / before-and-after". This is the main channel. - Big account posting usage on its own (confirmed): Marcel Kargul's post got 1.35M views and 12k bookmarks, and the same post also promoted Emil's skill. The two sides send traffic to each other.
- Platform listings (confirmed): Perplexity Computer's skills marketplace listed /teardown and its CEO reposted it; skills.sh has rankings and install counts; in addition, more than a dozen skill directories including ui-skills.com, craftwork.design, agenticskills.io, skillvault.md, Dojo, tessl, lobehub and mcpmarket have it, so the SEO long tail is covered.
- Chinese and Japanese repost accounts (confirmed): Chen Facai, Skills Studio (Japanese), Captain Luo, Zhishi Mao, Viking, ThirdBro and others. The Chinese-language "GitHub weekly roundup" pieces line up exactly with the 11-01 and 11-02 peaks.
- Long-term content the author has built up (confirmed): articles on jakub.kr (the OKLCH article hit HN with 859 points, 2025-08-25), the Interfaces magazine started in 2026-05 (hit HN with 100 points; by October he said subscriptions had passed his goal of 500 people); the README's first line links the magazine and skills.sh badges, so they send readers to each other.
- Backing (indirect): Interfere (YC S25) announced a $5.1M seed round in October and published its design system. The founders and colleagues shared Jakub's posts. This amplifies the "top design engineer" persona, but its direct contribution to the stars is unverified.
- Not found: paid ads, Product Hunt, big Reddit threads, TikTok. Also no HN thread about this repo itself.
Competitors
| Name | Where | Launched | Scale (early November) | How it differs |
|---|---|---|---|---|
| frontend-design in anthropics/skills | Big-company features / open source | 2025-09 | 188k stars (whole repo) | An official general-purpose frontend aesthetics prompt, one skill; it splits the work into 14 disciplines |
| nextlevelbuilder/ui-ux-pro-max-skill | Open source | 2025-11-30 | 146K stars | A searchable design knowledge base (colors, fonts, styles) that only gives suggestions and doesn't change code; it edits existing code directly |
| pbakaus/impeccable | Open source | 2025-11-16 | 104k stars, +8.8k this week | Generates pages from scratch, with 62 deterministic detectors that can run in CI; it has rules only, no linter |
| emilkowalski/skills | Open source | 2026-03-16 | 54K stars | The closest rival in the same track; Emil (author of Sonner / Vaul) focuses on motion; it covers more disciplines like typography, color, accessibility and copy |
| Nutlope/hallmark(Together AI) | Open source | 2026-04-27 | 49k stars, +6.7k this week | Focuses on marketing landing pages that "don't look AI-generated", with 50 built-in themes and render checks; it targets React apps with an existing design system |
| huashu-design (花叔) | Open source | 2026-04-19 | 29K stars | Turns one sentence into a finished piece (web / PPT / animation); it doesn't build from scratch |
| ibelick/ui-skills | Open source | 2026-01-07 | 10.7K stars | Also a "skills for design engineers" pack, with the directory site ui-skills.com of similar scale; its discipline split is more granular |
| MengTo/Skills(Meng To) | Open source | 2026-02-03 | 7.3k stars | A skill set from designer celebrities |
| heydamir/unslop | Open source | 2026-10-29 | 9.6k stars (this week) | Turns AI-looking UI into designed UI; a single skill |
| Jakubantalik/transitions.dev | Open source | — | 6.1k stars | A motion code library plus skills, partly paid; it's pure rules |
| google/elephant-goldfish-model | Open source | 2026-09-13 | 12k stars, +6.6k this week | A software design review skill from a big company, more engineering-oriented |
Source: GitHub repo_profile (11-06); setproduct.com 2026-11-03 comparison review.
Takeaway: there are many competitors, and most have more stars. It wasn't the first (ibelick in January, Emil in March, and the author's own make-interfaces-feel-better in Mar 13), nor the one with the most stars. But in the niche of "polishing existing code by discipline", it and Emil are the two most often mentioned together.
Differentiation
- Split by discipline, composable: seven better-* skills each cover one discipline, and better-interface orchestrates them into one sorted report. setproduct rated it "the broadest discipline coverage".
- Respects existing design systems: it invents no themes and uses the project's existing tokens and components. Suited to mature React apps that already have a design system, not just landing pages.
- Backed by personal taste: the rules come from the author's public writing, so it's "one person's judgment", not a list aggregated from elsewhere. The author himself says "many skills are slop; mine was originally for my own use".
- Demo-friendly "toy skills": /break (stress-tests a component against every edge case), /state-machine (renders all states of a component on one page), /variant, /teardown (exploded view). These are ideal for short videos, and teardown was even adopted by Perplexity in its product. Differentiation in distribution may matter more than in content.
- Weakness: no linter or render checks; it relies on the agent voluntarily reading the rules (setproduct points this out). This is one reason impeccable / hallmark pulled ahead in stars.
Visibility
It's spreading in small circles: well known among design engineers and the vibe coding crowd (Marcel's post with 1.35M views, the Perplexity CEO's reshare, Chinese and Japanese repost accounts, the setproduct review), but no mainstream tech media has covered it. It doesn't make money (MIT, free), and the author publicly shares stars and install counts. The category itself is already crowded.
Copycats
The category is heavily copycat (heydamir/unslop, a wave of taste-skill and anti-slop skills that reached tens of thousands of stars within a week). Those directly reusing its content include affaan-m/everything-claude-code, which copied make-interfaces-feel-better verbatim into its own repo; skillsy and lobehub host several forks of make-interfaces-feel-better. Dozens of skill directory sites repost its skill page.
What others can borrow
- Build the audience and reputation first, then turn the content into skills: the articles are the source of the skills and the source of trust.
- One repo, keep shipping: add a new skill every 1–2 weeks, and each release gets a post with a demo video and a one-line install command, using the same repo to accumulate stars and installs.
- Build one or two visually impressive "toy skills" (teardown, break) to break out, and aim for platform adoption (Perplexity, skills.sh).
- Regularly share milestones (160k → 425k installs) to create social proof.
Unknowns
- The day-by-day star curve from July to mid-October is missing (our database starts at 10-28, and the stargazers endpoint can't retrieve it), so we can't separate how many stars 07-24, 07-28, and 09-14 viral posts each brought.
- Install counts come from skills.sh statistics and may include CI or repeat installs, so they don't equal the number of people actually using it.
- Whether Perplexity's inclusion involved any business cooperation or payment is unverified.
- The peak around 11-02 is the combined result of a Chinese weekly-list roundup and Perplexity's free period; we can't tell how much each contributed.