Field noteOct 27, 20265 min read1 source

Hallmark is a rebuilt design skill for agents, that refuses to look AI generated.

In one minute

Hallmark is an open-source design skill for Claude Code, Codex, Cursor, and OpenCode that aims to prevent generic-looking agent-generated interfaces. Its v2 combines selectable visual materials with browser-based linting, review gates, and redesign checks intended to preserve an existing page’s behavior.

What changed: Hallmark introduces an anti-AI-slop design skill for Claude Code, Cursor and Codex, distinct from the prior logo-design skill focused on professional logos.

What happened

Nutlope’s public Hallmark repository describes a design skill for coding agents that “refuses to look AI-generated.” The project is made by Together AI and supports Claude Code, Codex, Cursor, and OpenCode. It is distributed as a skill rather than as a standalone design application: a global installation is available through npx skills add nutlope/hallmark -g, while installation can also be scoped to a single harness with an --agent option. Claude Code users can additionally install it as a plugin.

The repository frames Hallmark as a way to impose a design floor while leaving room for a task-specific page. In v2, the project supplies fifty themes, each described as a material with its own type, color, surfaces, components, texture, and a picture of a finished page. A build reads one theme file, which the repository says is capped at 8 KB. The project deliberately distinguishes a material from a page plan: a theme should not prescribe page content or ordering. Hallmark’s script offers three themes that fit a prompt, and another script derives a page anatomy from the job. The supplied examples span uses such as bank reconciliation, a webhook API, an identity studio, a fragrance atelier, farm records, and construction-site scans.

Hallmark also specifies rules for how a build should proceed. Its pre-flight reads a project’s fonts, palette, framework, and a locked design.md, then asks once for inputs such as photos, screenshots, and real numbers. The README says it does not invent facts. It builds the hero separately, renders and reviews that hero before the rest of the page, and stamps each self-contained HTML and CSS page with its design decisions in the first CSS comment. It includes 755 openly licensed photos selected by subject and credited on the page.

The project’s checking system has three tiers. The non-shipping “Floor” covers contrast, focus, reduced motion, invented facts, and a broken phone layout. “Reflex” gates address defaults the repository says models often fall into, though a build may overrule them with a recorded reason. “Finish” checks are craft details that do not block shipping. sloplint --render renders at eight widths; a subagent then reviews the result in a fresh context. Browser checks and screenshots require Node 18 or later, Chrome, and a global puppeteer-core installation, although static checks can run without those dependencies.

Hallmark also has explicit redesign and audit paths. /hallmark redesign <path> is described as rebuilding a page while retaining its content and wiring, whereas /hallmark audit <path> reports without rebuilding. Before redesigning, sloplint --before records links, fields, handlers, and headings; sloplint --kept fails the redesign if one is missing. The handoff is a side-by-side before-and-after comparison.

What the archive adds

This is a new implementation on a known thread. Our earlier note, Teams are narrowing in on skills as the standard for providing domain knowledge to agents, covered skills as selectively loaded packages of domain knowledge for coding agents. Hallmark adds a design-specific package whose claims extend beyond instructions into scripts, rendered-page checks, and redesign preservation tests.

Trend signal

Hallmark supports the pattern Design systems become agent context. Its themes, project inputs, and design.md act as structured context for a build, while the linting tiers turn part of the desired result into executable constraints. It also fits Skills become a portable capability layer: the same package can be installed across several coding-agent harnesses, though its practical value still depends on the checks actually catching bad output.

Why it matters

For teams using agents to create or revise product UI, the notable contribution is not a promise of better taste by itself. It is an attempt to make several review concerns operational: accessibility-related basics, responsive rendering, factual inputs, recognizable visual failure modes, and retention of existing interactions during a redesign. That can give a team a more inspectable handoff than a prompt such as “make it less generic.”

The repository also shows a useful division of labor. A skill can suggest a visual direction and page structure, but browser rendering and a separate review pass are used to test the artifact. Teams can treat that as a design workflow pattern even if they do not adopt Hallmark’s themes: capture source inputs, render at relevant widths, inspect the result independently, and verify that a redesign retained known page behavior.

Limits and uncertainty

The repository describes Hallmark’s intended workflow and bundled assets, not comparative evidence that it produces more effective or more original interfaces than other skills. It does not provide success rates for the linting gates, results from user studies, or evidence that a fresh-context subagent reliably catches subjective design problems. “AI-generated” is also a broad aesthetic judgment rather than a measurable outcome established by the material. Browser checks require extra local dependencies, and the supplied material does not establish how consistently those checks work across project frameworks or deployment environments.

Skill to build now

Pick one existing marketing or product page and create a pre-redesign inventory of every link, form field, handler, heading, and required mobile state. Then run a visual review at several viewport widths before and after an agent-assisted change. Make the preservation list a failing check, not a review comment. Hallmark’s redesign flow makes this concrete: recording behavior before a visual rewrite gives a team a way to distinguish a prettier page from a working replacement.

Reader takeaway

Hallmark packages design direction, UI checks, and redesign safeguards as a portable skill for coding agents. Its useful idea is that an agent’s visual output should be evaluated as a rendered, functioning artifact—not only as code or a persuasive design prompt.

Open question

Will Hallmark’s browser checks and fresh-context review reduce generic UI without turning its own rules and themes into another recognizable template?

Evidence / sources

  1. 1github.com/nutlope/hallmark
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicClaude CodeExplore the claude-code topic hub.
  3. 03topicCodexExplore the codex topic hub.
  4. 04archiveField NotesOpen the latest editorial analysis.
  5. 05source ledgerSource LedgerInspect the public source evidence graph.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract