The shortlist
7 projects
The wider reading list: projects that stood out across the week, in their own words. Positions reflect the shortlist, not GitHub rank.
Also worth watching
Fast risers from off the boards — worth a look anyway.
The wider reading list: projects that stood out across the week, in their own words. Positions reflect the shortlist, not GitHub rank.
Fast risers from off the boards — worth a look anyway.
The weekly Star
A project by pbakaus
A design skill for AI coding agents that adds 24 shared design commands and deterministic frontend anti-pattern detection.
The page your agent ships this month is warm. A cream background, an orange accent, a serif heading that feels editorial, three cards below the fold with a drawn icon each. It looks tasteful, and it looks like everyone else's, because it is the average of the same web every model was trained on. Two years ago the average was a purple gradient and Inter. The average moved. It did not get more personal.
Impeccable is a skill that makes your agent argue with that average. Paul Bakaus built it; he created jQuery UI and later worked on Chrome DevTools at Google, and he started this from the design skill Anthropic ships to keep Claude away from generated-looking pages.
Stars at each capture
Rank on each board
It sat on the daily board four days and the weekly board six this week, adding 3,527 stars. When we installed it the repo stood at 78,866 with 63 open issues, which for a project this size is a quiet issue tracker.
Three things. A vocabulary of two dozen commands (polish, audit, critique, distill, bolder, quieter) so you and your agent have words for design work instead of "make it pop". A written constitution, the craft floor, that names the habits models fall into and bans them: emoji standing in for icons, gradient text, the little label above every heading, same-size cards as a page structure. And a detector, a CLI that scans your HTML and CSS for 59 deterministic anti-patterns. The detector runs without an API key and without a model. It's a spellchecker where design is the spelling.
We built a page with the old sins on purpose. The detector came back in under a second, contrast ratios computed to the decimal, the purple gradient named, "Supercharge Your Workflow" flagged as a buzzword.
Then we rewrote the page by the craft floor. Real copy, drawn icons, a serif chosen for personality, a warm background chosen to feel human. Two findings came back, both about us. The serif was on the tool's overused-font list, because every other site fleeing the old look picked it too. The background got flagged in the tool's words as "the default tasteful AI surface". We had escaped the average straight into the next average.
Bakaus wrote the sentence for this before we tripped over it: "Today's antidote to slop becomes tomorrow's slop the moment everyone reaches for it." The detector files 27 findings on impeccable.style itself, four against the dark, cyan-accented scheme its own homepage is built on.
Keith redesigned a blog that had drifted through five migrations and got "one site with different rooms, not seven sites stitched together". His caveat is fair: "It can't look at a yellow background and feel that it's too aggressive." On Hacker News, people who tried this whole category report the same ceiling from the other side: consistency and accessibility improve, while the design itself still converges "unless you are extremely explicit". Bakaus agrees with them. You can't bottle taste. What the tool can do is catch the converging, in your agent and in you, and show its work.
From a project directory:
npx impeccable detect .No account, no key, one download. Read what it says about a page you
already own; the findings are specific enough to argue with, which is
the point. If the argument persuades you, npx impeccable install adds
the skill to your agent. Two notes from our install. "Project" means the
nearest directory it treats as a project root, and without a
package.json it climbs upward, so anchor the folder first. And the hooks
that watch your agent edit run chatty until they're calibrated; the
author spent weeks on his own, and an off switch ships in the box.
A fair first test: run the detector on the last page your agent built you. Count the findings you agree with. If you agree with more than a couple, keep the skill installed before the next one gets built.
A CLI harness that runs AI coding agents like Claude Code and Codex as managed teams booted from YAML topologies.
Days on board: 5 daily · 1 weekly · 0 monthly
Desktop app for local voice AI workflows: voice cloning and design, video dubbing, dictation, and long-form audio production.
Days on board: 3 daily · 7 weekly · 1 monthly
A runtime that sandboxes autonomous AI agents, enforcing kernel-level policies on file, syscall, and network access.
Days on board: 4 daily · 1 weekly · 2 monthly
An agent memory system that lets LLM agents learn over time by retaining, recalling, and reflecting on experiences, not just chat history.
Days on board: 2 daily · 7 weekly · 6 monthly
A plugin for AI coding agents that steers them to write minimal code via a ladder favoring reuse, stdlib, and native platform features.
Days on board: 5 daily · 0 weekly · 7 monthly
Self-hosted control plane that orchestrates teams of AI agents with org charts, goals, budgets, and governance.
Days on board: 2 daily · 7 weekly · 6 monthly
A design skill for AI coding agents that adds 24 shared design commands and deterministic frontend anti-pattern detection.
Days on board: 4 daily · 6 weekly · 0 monthly
Photoshop, re-implemented clean-room in pure Rust — 31k stars in five days without one day on our boards — the flagship of a seven-repo suite quietly rebuilding the whole Adobe line.
Point Claude at your game library: skills and tools that let Claude Code mod almost any PC game you own — recon, reverse engineering, generated art and audio, in-game testing, and a showcase video at the end.
Turn off Apple Intelligence, the analytics, the ads, and the pop-ups — a native macOS app and CLI where every change can be undone — the switch Apple didn't ship.