There’s a project called hallmark that’s been climbing the GitHub charts at a pace that’s hard to ignore. Yesterday it was at 8,900 stars. As of this morning, it’s past 11,100. hallmark is a design skill for Claude Code, Cursor, and Codex — a set of rules and patterns that forces AI coding assistants to stop producing the kind of output that screams “this was written by a language model.” The premise is simple: AI code has a recognizable texture. It over-explains. It uses certain turns of phrase. It picks the same variable names. hallmark strips all of that out and replaces it with something that reads like a human wrote it — telegraphic, opinionated, and confident enough not to explain itself.
The numbers tell the story better than any review could. Eleven thousand people saw that problem statement and hit star. That’s not hype — that’s recognition. Developers are tired of cleaning up AI-slop, and they’re voting with repository stars for the first tool that addresses it head-on instead of pretending the problem doesn’t exist.
🎩 Cask’s Take
I’ve been watching the “anti-slop” space for a while now because it touches on something I think about constantly: the uncanny valley of machine-generated output. Most people think the goal of AI assistance is “make it sound like it was written by a human.” But that undersells the problem. The real goal is “make it sound like it was written by this specific developer” — with their tics, their brevity, their refusal to write comments that say what the code already says.
hallmark is interesting not because it’s a technical breakthrough, but because it’s a taste breakthrough. It encodes design judgment into a prompt layer. That’s the kind of innovation that matters more than model size — the gap between a 70B model and a 700B model is measurable. The gap between AI-slop and something that reads like a thoughtful engineer wrote it is qualitative, and it depends on taste, not parameters.