# "Codex's design ability isn't bad, you're just using it wrong": a frontend dev locked out of Claude built an open-source Claude Design anyone can use · GitHub open-source project

> Source: https://stringers.io/en/github/gh_Innei__ryugi.html (Stringers / 什么值得做, generated 2026-11-08)
> Stringers answers one question every day: what small product can one or two people build in a few months and actually make money from. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these verdicts and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools today / hot / products / directions / direction / data. Setup: https://github.com/komako-workshop/stringers (the user logs in at stringers.io → "配置 agent" to get a personal URL). Site index: https://stringers.io/llms.txt

> Chinese developers hit by Claude bans and quotas, switching to Codex, couldn't use Claude Design's canvas, and this filled the gap. The author has reach in the Chinese X dev circle, and he posted a rant right when the biggest upstream, Superpowers, got flak for its new version, which drove two waves of stars.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-6

Innei/ryugi · Innei

# "Codex's design ability isn't bad, you're just using it wrong": a frontend dev locked out of Claude built an open-source Claude Design anyone can use

**What it is** A workflow for coding agents plus a local canvas: the agent draws several options side by side, and you pick, comment, and sign off. Works with Claude, Codex, Cursor, and others. Free.

**Why it took off** Chinese developers hit by Claude account bans and rate limits, switching to Codex, lost access to Claude Design's canvas, and this fills that gap. The author has pull in the Chinese X scene, and he posted a rant at the moment the largest upstream, Superpowers, released a new version and got criticized. That brought two waves of stars.

**B** **How you could get in on it** Build the English-language version, or just the "side-by-side options + pixel comments + sign-off" piece, as a standalone tool that plugs into Superpowers and gstack. For the launch post, lead with a counterintuitive takeaway and a good-looking screenshot.

Grade B: the direction of "open-source versions of big-company features that are locked behind subscriptions or regions" holds up, and the English-language space is still empty. But most of ryugi's stars come from the author's own influence, the repo has only 350 stars, the skills-workflow niche is already crowded, and a big company could fill this gap at any time. Chatter in small circles

Data (updated daily) · Repo created 2026-10-19 · 367 stars · 7 days +124

GitHub · stars

Data 2026-10-23 → 2026-11-08 (17 days, 17 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

ryugi (流儀) is an open-source project by Innei, a frontend developer at LobeHub: a workflow for coding agents plus a local canvas where the agent first draws several interface options side by side, you pick and comment, and then it writes the code using TDD. The repo is 18 days old with 353 stars (GitHub, 11-06), and each wave lines up with a post the author made on X. The star count isn't large, but it hits a reusable pattern: take a big-company feature that just launched but is locked behind subscriptions or regions, ship an open version anyone can use, and push it out in the few weeks when people still can't access the original.

### A two-sentence post

On the evening of Oct 2026 21, X user "拾一" (@__oQuery, Innei's Chinese account, with LobeHub's verified badge next to the avatar) posted: "I want to revise my view that Codex is bad at design. He's good, maybe you're just using the wrong method." The only attachment was a link, which led to a canvas screenshot on ryugi.innei.dev.

The post got 95k views and 358 bookmarks (X, 11-06). Nobody in the replies asked "what tool is this?" Instead people discussed how they get Codex to make good-looking interfaces: some said to have Codex list design tokens before starting, others posted sites they'd built with Codex. The post never pitched the tool. It just offered a counterintuitive claim, and curious people clicked through on their own.

The backdrop was the author's own experience that month. In his monthly review on his 10-24 blog, he wrote that after upgrading a backup account to Claude Max, it was banned immediately and the money wasn't refunded. Many Chinese developers were getting banned or rate-limited by Claude at the same time, switching to Codex and DeepSeek. But Anthropic's Claude Design canvas (launched 2026-04, later moved into Claude Code's `/design` in August) was only available to Claude subscribers. Switch models and you lose it.

### What it actually does

Two parts, installed together:

- **A set of skills** : brainstorm → write plan → TDD → debug → verify → worktree, forked from obra/superpowers (313k stars) with many steps removed.

- **A local workbench** (`npx ryugi-artifacts`): the canvas draws several options side by side, lets you leave a comment on any pixel and submit your selections; design docs can be commented sentence by sentence; an HTML test bench with sliders (tweak animation easing, font size, and similar parameters); and Verify puts the built page next to the design you originally chose, so you click Pass or Send Back.

The README's first line tells you to send "Let's make a react todo list," and it will brainstorm first, "draw the options instead of describing them in text." It supports Claude Code, Codex, Cursor, Gemini CLI, Kimi Code, OpenCode, Qwen Code, and a dozen-plus other tools. Everything is free, with nothing paid.

### How the stars came in waves

| Date | What happened | star |
|---|---|---|
| 10-19 | Repo created, npm package published the same day | — |
| 10-21 | The "Codex design ability" post, 95k views | Cumulative 10-23 reached 52 |
| 10-27 | Second post: "The new Superpowers is dumb as hell... that's why ryugi exists," 25.7k views | 10-28 single day +104, the peak of the whole run |
| 11-01 | Acquaintance wibus_wee (mx-space collaborator) posted "installed across the board, feels amazing" | 11-02 +34,11-03 +42 |
| 11-03 | License changed from MIT to a custom license | About +10 per day afterward |

(Star data from GitHub daily stats, 10-23 to 11-06; view counts from the X post pages, 11-06)

The second post was the key to this growth. Superpowers is the largest package in the skills scene. Its v7.0.0 changed the brainstorming flow, and users complained the questions were too formulaic and wordy; the v7.0.1 release notes admitted as much. Innei chose this moment to go off on it directly, and replies quickly had people agreeing: "It's not just the new version, the old one was dumb too, it easily triggers the whole pipeline, so I kept only the brainstorm skill and deleted everything else." Baoyu (@dotey) replied, "Maybe you don't need Superpowers at all..." When a big account weighs in on the replies, the post brought in +104 the next day.

Beyond those two posts, the channel was the author himself: his blog innei.in has versions in Chinese, English, Japanese, Korean, and Traditional Chinese, and his monthly review had a section titled "Open-source Claude Artifacts." A few acquaintances reposted on X, and some people opened PRs. It didn't hit GitHub Trending, and I found no coverage on HN, Reddit, Product Hunt, or in the media.

The npm package shipped about 130 versions in 18 days, with 10-22 to 11-05 downloads of 21k (npm). But since it releases almost daily and every npx run re-fetches the package, that number shouldn't be read as a user count. The direct use of frequent releases is that the author has a new, screenshot-worthy feature almost every day to post about, like "UI verification with comments on the canvas, which Claude Artifacts doesn't have."

### Who's nearby

| Project | Scale | Compared with ryugi |
|---|---|---|
| obra/superpowers | 313k stars | Upstream, text-only workflow, no canvas |
| garrytan/gstack | 140k stars | Also a skills workflow package, backed by a famous name, no canvas |
| EveryInc/compound-engineering-plugin | 26k stars | Skills workflow, no canvas |
| Claude Design / Claude Code `/design` | Anthropic (official) | The canvas prototype is right here, but it requires a Claude subscription |
| devdotfast/whiteboard | 3,451 star | A whiteboard shared by humans and agents, mostly for reviewing code and drawing architecture |
| plannotator/artifact-server, plivo-labs/glance | 262 / 53 star | Only "publish outputs as links" |

Both the "skills workflow package" and "open-source Artifacts" lines have plenty of players. Combining the two, working with any agent, and centered on "draw it out and pick an option" is, as far as I can see, only ryugi. Copycats haven't reached scale either: of the 15 forks, all are contribution forks with 0 stars, and one port to ZCode no longer opens.

One thing to watch: 11-03 the author switched the license to a custom "Ryugi License 1.0," which states that major Chinese companies such as Tencent, Alibaba, ByteDance, Baidu, and Huawei, along with products like Qwen, Trae, and Qoder, are "not granted any rights," and it bans putting the skills into a marketplace. The workbench npm package also went closed-source from 1.51.2. In other words, you can't take its code and fork a version out of it; you'd have to write your own from scratch.

### Why this one, and why now

- **Timing** : Chinese developers were being banned and rate-limited by Claude, and after switching to Codex and DeepSeek they found the Claude Design-style canvas gone. ryugi effectively moved "Claude Design" onto any agent.

- **The author is the channel** : a design engineer at LobeHub who posted 11k posts on X, and two other posts that same month each got 7–90k views. His work looks good, and one screenshot can spread.

- **Riding upstream momentum** : forked the largest skills package, and launched "better brainstorming" in the same week its new version was getting criticized.

- **Something new every day** : canvas, Verify, Cloud share links, multi-harness support, released one after another, which kept the stars coming in several waves.

The differentiation is real, but it only accounts for half. The other half is timing and the author's pull in the Chinese X scene. If someone without X followers posted the same thing, I estimate they'd not reach 350 stars within 18 days (inference).

### If it were you

**Verdict** : B. The direction of "open versions of locked big-company features that anyone can use" holds, and ryugi shows that people want it. But ryugi itself has only 350 stars, so its ceiling isn't clear yet; the skills-workflow niche is already crowded; and Anthropic or OpenAI could build the canvas into their own tools or lift the restrictions at any time.

**Who it suits** : frontend or design engineers with several thousand followers on X, Jike, or English developer circles who can make interfaces look good and have screenshots worth sharing.

**Who it doesn't suit** : anyone without a posting channel who expects GitHub Trending to bring the stars. ryugi never reached Trending, and every wave of stars came from the author's own posts.

**How to do it** (pick one):

1. **Build the English-language version** . ryugi has only circulated in the Chinese X scene, and in the English world a "Claude Design canvas that any agent can use" is still an empty slot. Its license forbids modifying the code, so you'd have to write your own.

2. **Build only the canvas, not the workflow** . Package "side-by-side options + pixel comments + Pass / Send Back" as a standalone MCP or CLI that plugs into large workflow packages like Superpowers and gstack, without competing with them on process. The reply "kept only the brainstorm skill, deleted everything else" suggests many people only want this piece.

3. **Swap in a different locked feature** . Apply the same playbook to the next new feature that's limited to subscribers or only available in certain regions.

**How to get users** : Innei's playbook. The launch post doesn't pitch a tool. It makes a counterintuitive claim ("X model is actually good at design, you're just using it wrong") with a nice canvas screenshot. The second wave rides the upstream version that just got roasted. After that, ship one screenshot-worthy feature a day.

**When to move** : as soon as possible, ideally before Anthropic or OpenAI's next big release. For example, if Codex ships a built-in canvas, or Claude Design drops its subscription requirement, this gap closes.

**What would prove this wrong** : (1) ryugi is still under 500 stars a month later, meaning "pick from rendered mockups" doesn't resonate beyond the author's circle; (2) OpenAI adds an official design canvas to Codex, or Anthropic opens `/design` to non-subscribers; (3) an English-language clone appears and passes 1,000 stars within two weeks, meaning the gap is already taken.

[GitHub repo ↗](https://github.com/Innei/ryugi)

**Research notes: competitor table, acquisition evidence, full research report**

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| obra/superpowers | Open source | 2025-10 | 313k stars | ryugi's upstream. Text-only workflow. Users found v7's brainstorming redundant. ryugi trims the flow and adds a visual canvas. |
| Claude Design / Claude Code /design canvas | Big-company features | 2026-04 / 2026-08-17 | Anthropic (official) | Official version requires a Claude subscription. ryugi runs locally and works with Codex, DeepSeek, and any other harness. |
| Claude Code Artifacts | Big-company features | 2026-06-18 | Anthropic (official) | Tied to claude.ai. ryugi adds multi-option canvases, commentable docs, and Verify. |
| garrytan/gstack | Open source | 2026-03 | 140k stars | Skills workflow pack. No visual canvas for picking between options. |
| EveryInc/compound-engineering-plugin | Open source | 2025-10 | 26k stars | Skills workflow. No canvas. |
| devdotfast/whiteboard | Open source | 2026-08 | 3,451 star | Open-source desktop whiteboard shared by humans and agents. Focused on code and architecture review, not UI option comparison or TDD flow. |
| plannotator/artifact-server / plivo-labs/glance | Open source | 2026-08 / 2026-06 | 262 / 53 star | Self-hosted publishing and commenting only. Covers the same ground as ryugi's Cloud piece. |

Where users come from

- **Verified** **Author's own X posts (Chinese dev circles)** : post 10-21 got 95k views and 358 bookmarks. Post 10-27 got 25.7k views, and stars rose +10-28 … +104.

- **Verified** **Roasting the upstream Superpowers v7 to drive traffic** : "The new Superpowers is so dumb… that's why ryugi exists." Replies from 宝玉 and yetone.

- **Verified** **Insider KOL reshare** : wibus_wee 11-01 "installed across the board, feels amazing," followed by +11-02/03 / +34 / +42.

- **Verified** **Personal blog (5 language versions)** : innei.in's 2026 monthly recap has a section on "an open-source Claude Artifacts."

- **Verified** **Listed on skills aggregators** : skillfoxx.ru, scriptbyai, and similar lists.

- Inferred **HN / Reddit / PH / press / ads** : none found.

## Why Innei/ryugi (流儀) took off

> Channel: GitHub. Research day 2026-11-06. Data card: repo 2026-10-19,18 days old, 353 stars (GitHub API read 362 that day), 15 forks. Biggest single-day gain +104 (10-28). Last 7 days +139, first 7 days +162.

### Bottom line

A small project that took off in Chinese X circles. The author, Innei (拾一, @__oQuery, LobeHub frontend, 3.5k GitHub followers), got his Claude account banned and then hit Kimi's limits. So he took the Claude Artifacts / Claude Design canvas and built an **open-source version that works with any agent** , on top of a Superpowers fork. He posted two items on X (95k and 25.7k views), and they drove two waves of stars. The selling point is the canvas: Superpowers' text-based option description is too wordy, so ryugi draws the options and lays them side by side for you to pick. The timing fits: Chinese developers were switching from unstable Claude to Codex and DeepSeek, and ryugi fills the gap they were missing.

### What it is

- Two parts: (1) a set of skills (brainstorm → spec → TDD → debug → verify → worktree), forked from obra/superpowers v7.0.0; (2) an "Artifacts workbench" `npx ryugi-artifacts`, including a canvas (multiple options side by side, comments on any pixel, submit a choice), commentable design docs, an HTML test bench with sliders, and Verify (compare the implementation against the design, pass or send back), plus a cross-project local library. There's also Ryugi Cloud (ryugi.innei.dev) for publishing artifacts as share links. (README, [github.com/Innei/ryugi](https://github.com/Innei/ryugi))

- Supports over a dozen harnesses, including Claude Code, Codex, Cursor, Gemini CLI, Kimi Code, OpenCode, Pi, Qwen Code, Hermes, Devin, Droid, and Antigravity.

- Free, no paid tier. After 1.51.2, the npm package `ryugi-artifacts` went closed-source (UNLICENSED). It shipped 130 versions in 18 days (npm registry).

- License: commit "replace MIT with Ryugi License 1.0" (2026-11-03). It states that major Chinese tech companies (Tencent, Alibaba, ByteDance, Baidu, Huawei) and products like Qwen, Trae, Qoder, and WorkBuddy get "no rights granted," and it bans scraping the skills into marketplaces (LICENSE commit history, README).

### Growth data and timeline

| Date | Event | Star changes (github repo_day / data card) |
|---|---|---|
| 10-19 | Repo created, first commit. npm package went live the same day. | — |
| 10-21 | X launch post "I want to change how you see Codex's weak design ability," with ryugi.innei.dev attached. **95k views, 358 bookmarks** (https://x.com/__oQuery/status/2112990461950660897; the author's blog says 93K views and 700+ bookmarks). | Cumulative 10-23 reached 52 |
| 10-22 | Japanese blog lab.main.jp wrote up the canvas option-picking flow (indexed by Exa). | — |
| 10-24 | Author's monthly blog recap, with a section on "an open-source Claude Artifacts" (https://innei.in/en/notes/2026/10/24/2026-10-recap). | 10-24:+20 |
| 10-27 08:03 UTC | X post "The new Superpowers is so dumb… that's why ryugi exists," **25.7k views** . 宝玉 (@dotey) and yetone replied below (https://x.com/__oQuery/status/2114991486685384730). | **10-28: +104 (highest of the whole run)** |
| 10-30/11-01 | The author posted a string of feature threads ("UI review with comments on the canvas, something Claude artifacts doesn't have"). 11-01 wibus_wee (a mx-space collaborator) posted "installed across the board, feels amazing" (https://x.com/wibus_wee/status/2116822229896073265). | 10-31 +23,11-02 +34,11-03 +42 |
| 11-03 | Switched to a custom license that restricts major Chinese tech companies. | 11-04 +11 |
| 11-06 | New post "You and the agent, on the same page," with video. 2,961 views at time of capture. | 11-05 +13,11-06 +10 |

- npm `ryugi-artifacts` had 21,314 downloads from 10-22 to 11-05. Peak was 10-23/10-24 (2.6k–2.8k per day), then 600–1,800 per day afterward (api.npmjs.org). The count is inflated by 130 versions' worth of releases, mirrors, and every npx install, so don't read it as user count.

- Star growth lines up with the author's X posts one for one. It never made GitHub trending (trending_days_any=0; repo was under 60 days old).

- Third-party chart: top10earth "new Shell repos, last 30 days," ranked #8.

### Competitors

| Name | Where | Launched | Scale | How it differs from ryugi |
|---|---|---|---|---|
| obra/superpowers | Open source | 2025-10 | 313k stars | ryugi's upstream. Text-only workflow, no visual workbench. Users complained that v7.0.0's reworked brainstorming was "formulaic and wordy," and v7.0.1 admitted as much (Superpowers releases). ryugi cuts the redundant steps and draws the options so you can pick. |
| Claude Design / Claude Code `/design` canvas | Big-company features | Claude Design 2026-04; `/design` in Claude Code 2026-08-17 | Anthropic (official) | The canvas-for-picking-options prototype is here. But it's only for Claude subscribers and requires logging into claude.ai. ryugi runs locally and isn't limited to one harness, so Codex and DeepSeek work too. |
| Claude Code Artifacts | Big-company features | 2026-06-18 | Anthropic (official) | Publishes session output as pages on claude.ai, also tied to a Claude subscription. ryugi adds multi-option canvas picking, commentable docs, and Verify. |
| garrytan/gstack | Open source | 2026-03 | 140k stars | A skills workflow pack for coding agents with a famous backer, but no visual canvas for picking options. |
| EveryInc/compound-engineering-plugin | Open source | 2025-10 | 26k stars | Also a skills workflow, no canvas. |
| devdotfast/whiteboard | Open-source desktop app | 2026-08 | 3,451 star | A whiteboard shared by humans and agents, for code review and drawing architecture. The focus isn't UI option comparison or the TDD flow. |
| plannotator/artifact-server | Open source | 2026-08 | 262 star | Self-hosted publishing, commenting, and versioning service. Only covers the Cloud piece of ryugi. |
| plivo-labs/glance, momja/Exhibit | Open source | 2026-06 | 53 / 20 star | Self-hosted Artifacts alternative, only handles "publish as a link." |

Takeaway: both the "skills workflow pack" lane and the "open-source Artifacts" lane are crowded. But **combining the two, working with any harness, and leading with canvas-based UI option picking** is something only ryugi has shown so far.

### Growth tactics (evidence)

1. **Author's own X account (Chinese dev circles)** : confirmed. Posts 10-21 and 10-27 drew 95k and 25.7k views. The 10-27 post brought the highest single-day gain of +104 the next day. The author has built up a following over time (11k posts, LobeHub verified badge). The blog also mentions two other posts from the same month getting 7–90k views.

2. **Driving traffic by roasting upstream** : confirmed. "The new Superpowers is so dumb" rides on the frustration of Superpowers v7 users (there are also Reddit posts complaining about "obra/superpowers overkill"), and big accounts like 宝玉 and yetone replied below.

3. **Insider circle reshare** : confirmed. wibus_wee (a collaborator on mx-space) recommended it in 11-01. Forks include familiar names like wibus-wee and equt, who opened PRs #6 and #7.

4. **Personal blog** : confirmed. innei.in's monthly recap comes in five language versions: Chinese, English, Japanese, Korean, and Traditional Chinese.

5. **Picked up by skills aggregators** : confirmed but low traffic (skillfoxx.ru, scriptbyai "480+ skills," and similar).

6. HN, Reddit, Product Hunt, press, ads: none found.

### Why it took off

- **Timing** : 2026 October. The author and Chinese developers around him were stuck with Claude bans and limits (the author's blog describes an alt account getting banned right after upgrading to Max, with no refund). Codex and DeepSeek became the replacements, but neither has Claude's Artifacts / Design canvas. ryugi fills the gap of "Claude Design in a non-Claude environment."

- **The proof point works** : the launch post doesn't say "I built a tool." It says "Codex's design ability isn't bad, I was using it wrong," and the before-and-after hooks curiosity. The author is a design engineer at LobeHub, and the output looks good on its own, so screenshots spread by themselves.

- **The upstream just blew up** : after Superpowers v7.0.0 changed its brainstorming, users complained loudly. ryugi jumped on it with "a better Brainstorming skill" and pulled some of that unhappy crowd away from the biggest skills pack.

- **Fast shipping** : 18 days, about 130 npm versions and v1.54+ releases. Almost every day brought a new feature worth posting about (whiteboard, Verify, Cloud, multi-harness support).

### Differentiation

"Sketch a few options, compare them side by side, comment on the pixels, then go into TDD." It's harness-agnostic, runs locally, and is free. Next to Superpowers it has a visual workspace and less redundant process; next to Anthropic's official tool it doesn't require a Claude subscription. The difference is real and explains some of the traction, but half of it goes to timing (Claude is unstable in China) and the author's own reach. Without his pull in the Chinese X dev circle, a differentiator alone would hardly get 350 stars in 18 days.

### Visibility and copycats

- Visibility: **it's spreading in small circles** . It only circulated in the Chinese developer corner of X (the author, wibus_wee, 宝玉 and yetone's replies). No English media, HN or Reddit coverage, and nothing in our x-radar corpus. It's free, so there's no revenue to disclose.

- Copycats: none at scale. The 15 forks are all contribution forks with 0 stars; LIPiston/ryugi-zcode ports it to ZCode (search engines indexed it, now 404); skillfoxx.ru lists the plugin. The author 11-03 changed the license to bar big Chinese companies and marketplaces from using it, which suggests he worries about being copied.

### Takeaways

- Take a feature a big company just launched that is "locked by subscription or region" (Claude Design's canvas), build an open-source version that works with any agent, and ship it at the moment accounts get banned or quotas bite.

- Attach to a huge upstream (fork Superpowers, 310k stars), and when the upstream's new version takes flak, pitch "a leaner alternative."

- Open with a counterintuitive claim like "model X is actually strong, you're just using it wrong." It spreads better than a straight tool pitch.

- Release every day, and every day ship a new feature you can screenshot, so stars come in waves.

### Uncertain

- Daily star counts from 10-19 to 10-22 aren't on record (our database only starts on 10-23); the GitHub stargazers API returned 404 this time, so I couldn't verify timestamps.

- How many real users: npm downloads are inflated by frequent releases, so no reliable conversion.

- Why LIPiston/ryugi-zcode is now 404; whether the license change (excluding Chinese big tech) caused controversy in the community. I found no discussion on X.

- Whether Cloud will be paid later: right now it's free with signed access.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.