← Back to blog
AiAbout 7 min read

Two 100,000-Star Repos Are Trying to Teach Agents Taste

Published Nov 7, 2026
Two 100,000-Star Repos Are Trying to Teach Agents Taste

The same week that a design skill from Together AI and a rules engine from a solo developer both climbed past 100,000 GitHub stars, two more repositories showed up with the same goal from different directions. One is a collaborative workspace where a team and its agents build images, video and products together. The other is a set of instructions that make a coding agent build a real company website instead of a generic landing page.

Neither name is familiar yet, but both are worth an afternoon.

open-design: a workspace where the agent sits in the room

open-design describes itself as the first collaborative AI design agent workspace. It is local-first and Apache 2.0 licensed, and it puts motion, prototypes, websites, apps, dashboards, presentations, images and video inside one project that a team shares with its agents.

The installation model is the interesting part, because it targets the mess of agent clients rather than trying to replace them.

```shell od mcp install claude od mcp install codex ```

The same command covers Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code with Copilot, GitHub Copilot CLI, Gemini CLI, OpenCode, OpenClaw, Cline, Trae, Kimi CLI, Mistral Vibe CLI, DeepSeek Reasonix, Raven, Pi, Antigravity and others. The dry-run flag, --print, shows what would be installed before anything happens, and --uninstall removes it cleanly. DeepSeek Harness needs its own dsh CLI in place first.

If you have no CLI installed at all, a bring-your-own-key proxy handles it. The daemon exposes POST /api/proxy/{anthropic,openai,azure,google,ollama,senseaudio}/stream, which runs the same loop without spawning a process. You paste a base URL, an API key and a model name, and there are presets for OpenAI, Anthropic, Azure OpenAI, Google Gemini, Ollama, LM Studio, vLLM or any OpenAI-compatible endpoint. Internal IPs, link-local addresses and CGNAT ranges are blocked at the daemon edge, which is the sort of detail you want to see in a tool that proxies other people's requests.

For a team producing images and video, the practical appeal is straightforward: the artifacts, the prompts, the reference material and the feedback live in one place that a human and several agents can edit.

taste-skill: rules a browser can check

taste-skill is MIT licensed with about 111,000 stars, and it is a single agent skill with a narrow goal. Its README states it plainly: the skill exists so that coding agents build professional company websites at the level of figma.com or x.ai rather than assembling a generic landing page.

```shell npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend" ```

Claude Code users can go through the plugin marketplace instead, and Codex has its own plugin command. The repository also ships install.sh, which symlinks the skill into your agent so a git pull updates it in place. On Windows under Git Bash it creates a directory junction when symlinks are unavailable, and it tells you when an older copied folder is still lying around.

The requirements are modest and specific: Node 22 or newer, and a Chrome, Edge or Chromium binary, with CHROME_PATH set if it lives somewhere unusual. ffmpeg on the PATH is optional, used for WebP output and the phone crop of a hero photograph. Nothing is installed, no API key is needed, and the bundled Wikimedia Commons photographs work offline.

The rules are specific in the same way. Plain copy, one design system, one section per screen, equal left and right edges, real photographs, real brand icons, real product UI, and product renders when a product has no photographs. Micro-interactions stay quiet. Shader effects, creative moments and scroll scenes are reserved for tech, gaming and software pages. Every build starts by classifying the business as software, gaming, tech, product, service or place, and planning screen by screen before any markup gets written.

What makes it unusual is how the output gets verified. A checker runs in a real browser and enforces those rules. The agent cannot grade its own work by describing how good it looks; a browser can check edge alignment, section counts and image sources.

Why this wave of tooling exists

The problem these repos attack has a name now: AI slop. It is the term for the specific look of a page built by a model asked for something and given no further instruction, and the list of tells is remarkably consistent. Purple-to-blue gradients. Cards nested inside cards. Gray text on colored backgrounds. A rounded-square icon tile above every heading. An eyebrow label above every headline. A hero image that misses the screen it was placed on. A signup form at the bottom of every page, including the pages where nobody is signing up.

Earlier answers came in two shapes. Impeccable shipped 62 deterministic rules that flag those tells, ranked findings from P0 to P3, and needed no model or API key to run. Hallmark, from Together AI, took a different approach: 50 themes, each one a material with its own paper, typography, color and components, plus 755 openly licensed photographs chosen per subject and credited on the page, and a browser-based lint that runs before delivery.

open-design and taste-skill belong to the same wave but pull in different directions. One is infrastructure for a team, the other is a single opinionated skill you install in an afternoon. Both assume the fix is procedural.

What this does not solve

Checkers verify what they can measure. A browser can tell you the left and right margins match, that each screen carries one section, that the hero image is a real photograph rather than a gradient, that the icon set came from somewhere. It cannot tell you whether the page says anything worth reading, whether the headline is the right headline, or whether the product in the screenshot is the product being sold.

The gap is permanent. Passing a rule set gets you a page that is structurally sound and visually coherent. It does not get you a page anyone wants to look at, and the repos that claim otherwise are selling a workflow, not a guarantee.

There is also a second-order effect worth naming. Once enough pages are built to one reference implementation, the checker becomes the house style. The pages stop looking identical to each other and start looking identical to the repository. That is a smaller problem than AI slop, but it is the same shape of problem, and the fix is the same one that works everywhere in this industry: use the tooling to produce work that has a point of view, and not as a filter for work that does not.

Two different bets on the same problem

Both projects assume that the missing piece in agent-built design is judgment, and both replace judgment with mechanism. open-design puts the judgment in the workflow, where a person and several agents iterate on the same artifact. taste-skill puts it in a checker, where rules decide whether the page passes.

The two projects split along a real divide in how teams think about AI output. One approach assumes you need more people in the loop. The other assumes you need a stricter definition of done. The expensive part of both is not the code, it is the material, and taste-skill's answer to that is a curated set of openly licensed photographs that ship with the skill and get credited.

The open question is whether either changes what a user actually sees, or whether taste simply moves from the model into the repository. Every AI slop detector so far has produced the same result: the pages stop looking like each other, and they start looking like the reference implementation.

Either way it is an improvement: a page built to a known standard, with real photographs and a checker that verifies it, beats a page built to an average of whatever the model felt like.

Related articles