Hallmark v2: Turning "AI Taste" into an Executable Checklist
The Intuition
You can spot an AI-generated webpage at a glance.
Not because it's ugly—often it's quite "pretty." It's because of an indescribable "flavor": an eyebrow line above every heading, tracked-capital labels everywhere, one colored word in every headline, and every page ending in an email signup.
These aren't mistakes. They are AI's verbal tics.
Hallmark's job is to list those tics as a checklist, then let AI check itself against it.
Not an Aesthetic Problem, an Engineering Problem
The name comes from hallmark—the purity stamp on precious metals. Together AI's message is clear: AI-generated UI needs a purity certification.
But Hallmark doesn't try to teach AI "what good design is"—that's too subjective. Instead, it catalogs AI-specific bad habits and checks them one by one.
This framing matters. Traditional design review is additive: what should good design have? Symmetry, hierarchy, contrast, rhythm. Hallmark is subtractive: what should AI-generated design NOT have?
Subtraction is easier to enforce than addition. It's hard to get AI to understand "elegance," but easy to have it check "is there an eyebrow above this heading"—a binary judgment.
Three Gate Layers
Hallmark's review system has three layers, from hard to soft:
Floor: Things that can never ship. Insufficient contrast, missing focus states, unhandled reduced motion, fabricated facts, broken mobile layouts. These are basic usability issues, not aesthetic ones.
Reflex: Default model behaviors. Eyebrow above every heading, tracked-capital labels, one coloured word in a headline, a page of empty frames, a hero that misses its screen, a form that says thanks without sending anything.
These are AI's "conditioned reflexes"—patterns seen so often in training data they became defaults. Hallmark marks them as "violable with justification." If the agent can justify it ("this eyebrow separates sections"), it stays; otherwise it must change.
Finish: Craft details. They don't block shipping, but they add quality points.
"Fresh-context reviewer gets the last word"
This is the smartest part of Hallmark's design.
Hallmark doesn't let the generating agent review its own work—that's a student grading their own exam. It uses a fresh-context reviewer—a brand-new context that never saw the generation process—to make the final call.
Why does this matter? Because generators and reviewers have different biases. Generators know their own intent and mentally fill in missing context. Reviewers see only the result and notice what the generator overlooked.
It mirrors human design workflows: a designer hands off to another designer for review. Not because the reviewer is better, but because distance produces insight.
Hallmark internalizes this into agent architecture: a generation agent builds the UI, a review agent checks it. Two agents, different prompts, different contexts, same skill.
What's New in v2: From Checklist to Theme System
V1 was a checklist—"don't do this, don't do that." V2 adds a new dimension: themes.
50 themes, each a "material"—paper, typography, color, surfaces, components, textures. Each theme ships with an image of a finished page that the agent references while building.
This isn't just "reskinning." Hallmark themes aren't CSS templates—they're design languages. A "paper" theme changes the agent's default choices around spacing, fonts, and color, not just CSS variables.
This matches how human designers work: you don't design a generic page and then apply a skin. You pick the material first—paper? metal? glass?—then design around it.
"Built from the job"
Another v2 change: building pages from the task.
A booking page should be able to book; a shop page should be able to sell; a newsletter page should be able to capture subscriptions. Hallmark doesn't generate a generic layout and fill in content—it chooses a different build path per page type.
This sounds obvious, but most AI UI generators don't do it. They produce something that "looks like a booking page"—with a heading, a form, a button—but the form may not work and the button may not submit. Hallmark requires pages to be "built from the job," meaning functionality isn't decoration.
What Exactly Is the "AI Look"?
What do Hallmark's "AI tells"—eyebrows, tracked capitals, one colored word—have in common?
They're all low-cost visual differentiation.
Faced with a blank page, someone without design training (human or model) instinctively adds things: an eyebrow to make the heading "stand out," a colored word to give it "hierarchy," capitals to make labels "more formal."
These moves aren't wrong—professional designers use them too. But professionals use them once; AI uses them every time. The essence of the AI look isn't a particular style—it's the tendency to overuse low-cost differentiation.
Hallmark's checklist targets exactly this tendency. It doesn't ban eyebrows; it bans "an eyebrow above every heading." It doesn't ban colored words; it bans "a colored word in every headline."
The Two-Layer Logic of Design Skills
Hallmark represents a new skill-design pattern:
Layer one: what to do. Hallmark generates UI—the functional layer.
Layer two: what not to do. Hallmark checks AI tells—the constraint layer.
Most skills only have layer one. Hallmark's uniqueness is turning layer two into a system—three gate layers, a fresh-context reviewer, 50 themes. The constraint isn't a "make it beautiful" line in a prompt; it's an executable checklist.
The pattern generalizes. Any AI-generated content has its own "AI look"—AI writing has verbal tics, AI code over-abstracts, AI presentations follow formulaic narratives. Every domain can have its own Hallmark: list the domain's AI tells, turn them into a checklist, and have the generator check itself.
The AI look is not an aesthetic problem—it's an engineering problem. Aesthetic problems are solved by taste; engineering problems are solved by checklists.
---
Project: https://github.com/Nutlope/hallmark Stars today: 2,859 Author: Together AI Status: v2, 50 themes, supports Claude Code / Cursor / Codex / OpenCode