From our experience: where Hallmark helps, and where it will not save you
At MTD Technologies, we have spent nine years shipping client websites and web applications, and the pattern Hallmark targets is one we recognise from the other side of the desk. The most common failure in agent-built pages is not ugliness — it is undifferentiated competence. A client cannot articulate why their new page feels generic, but their visitors can feel it, and it shows up in bounce rates and enquiry quality long before anyone blames the design system.
Where Hallmark genuinely helps, in our judgement, is as a team standard. If three developers and two agents all build pages, a shared, checkable floor beats a style guide nobody re-reads. The redesign mode’s kept-link guarantee and the hero-first review loop are the two features we would actually mandate in a client workflow, because they catch the two most expensive failure classes: broken functionality disguised as a redesign, and a weak first screen nobody revisits.
Where it will not save you: strategy and copy. Hallmark can stop your page looking machine-made, but it cannot decide what your page should say, who it is for, or what action matters most. A beautifully non-generic page with a muddled offer still loses to a plain page with a sharp one. Treat Hallmark as a quality floor for execution, and keep human judgement on positioning, messaging and the offer itself. Also be realistic about the theme catalogue: fifty themes is a strong starting set, but a brand with a mature identity will still need custom direction — which is why the custom mode, not the theme picker, is the long-term path for serious brands.
Cost and effort are refreshingly honest here. The skill is free and MIT-licensed; the real costs are Chrome and Puppeteer setup, the extra review pass on every page, and the discipline to let a gate block a ship. Teams that skip the rendered check to save ten minutes skip the feature that justifies the tool.
Getting started with Hallmark in five steps
- Install Node 18 or later, then run
npx skills add nutlope/hallmark -g— or drop the-gflag for a single-project install. - For rendered linting and screenshots, install Puppeteer Core and Chrome globally; verify the static checks still run without them so you know what you lose.
- Start with a low-stakes page — a new landing page, not a client redesign — and invoke
/hallmark <prompt>with real material: real copy, real numbers, real screenshots. - Let the hero-first review run, and read the lint output at all eight widths before approving the rest of the page.
- Only then try
/hallmark audit <path>on an existing page, and/hallmark redesign <path>once you trust the kept-link guarantee on a page you can verify by hand.
Frequently asked questions
Is Hallmark free to use?
Yes. Hallmark is published under the MIT licence, so teams can use it, fork it and ship with it at no licence cost. The only costs are your agent’s usual model usage and the optional Chrome and Puppeteer setup for rendered checks.
Which coding agents does Hallmark work with?
Claude Code, Codex CLI, Cursor and OpenCode. Installation places the skill in the harness’s skills directory, and it is invoked as /hallmark <prompt> (or $hallmark in Codex). A Claude Code plugin route is also documented.
Does Hallmark replace a designer?
No. Hallmark sets a quality floor and checks a page against recognisable machine-made tells; it does not make positioning, messaging or brand-strategy decisions. For a mature brand, pair it with human design direction — its custom mode is built for exactly that.
Can Hallmark redesign an existing page safely?
That is one of its strongest features. The redesign mode records every link, field, handler and heading before rebuilding, fails if any go missing, and presents a side-by-side before and after. Audit mode reports problems without changing anything, which is the safe first step on a live page.
Do I need Chrome installed for Hallmark to work?
Only for the rendered checks. The static lint checks run without a browser, but the eight-width rendered review and screenshot capture need Puppeteer Core and Chrome. If you skip them, you lose the review step that makes the skill worth adopting.
Hallmark is worth trying precisely because it is verifiable: install it, build one real page, run the audit on an existing page, and judge the before-and-after yourself. If your team ships agent-built pages for clients, a shared floor that catches generic output before clients see it is not a nice-to-have — it is the difference between looking like an agency and looking like a prompt. If you want help folding tools like this into a client delivery workflow — from AI integration and automation to a full web development engagement — explore our AI Tools Directory for more vetted tools, or talk to us about your next build.
What it can do: fifty themes, built from the job
Version 2 of Hallmark ships fifty themes. Each theme is a material — paper, type, colour, surfaces, its own components and texture — with a picture of a finished page, and the build lays out its own page inside it. The examples in the repository range from Dusk, a dark theme shown with a bank-reconciliation product, to Atelier, a fragrance-atelier look, and Furrow, a farm-records aesthetic. The important detail is what a theme is not: it never dictates what a page contains or in what order. Two prompts in the same theme should produce two different pages, because the page’s structure comes from its job, not from a template.
“Built from the job” is Hallmark’s core design idea, and it is the part most agent workflows skip. A booking page books, a shop sells, a newsletter subscribes: the action sits where that category of page puts it, and the page ends on it. That sounds obvious until you watch an agent build a booking page that buries the booking form beneath three screens of marketing copy. Hallmark’s build process deals the page’s anatomy from its job first, then dresses it in a theme — structure before style, in that order.
- Fifty themes — each a single file of at most 8 KB with a reference picture; a build reads exactly one, so themes stay cheap in context.
- Custom mode — say “custom” and Hallmark drops the themes entirely, designing from your prompt alone under the same floor and gates.
- Real photographs — 755 openly licensed photos, picked per subject and credited on the page, rather than stock imagery or invented placeholders.
- Your real product — point it at a URL, a dev server or an app and it captures the screenshots itself instead of designing around lorem ipsum.
- Redesign and audit —
/hallmark redesign <path>rebuilds an existing page while keeping its content and wiring;/hallmark audit <path>only reports what is wrong.
The redesign mode deserves emphasis for agency work. Most “redesign” tools quietly drop links, form handlers or headings. Hallmark records every link, field, handler and heading before the redesign, then fails the redesign if one went missing, and hands over a side-by-side before and after. That is the kind of boring, checkable guarantee that makes a tool safe to run on a client’s live pages rather than only on fresh demos.
How the anti-AI-slop design skill Claude Code workflow runs
Installation is deliberately low-friction. The documented route is a single command, requiring Node 18 or later:
npx skills add nutlope/hallmark -g
The -g flag installs it for every project; leave it off to install into the current project only. It lands in ~/.claude/skills/hallmark for Claude Code and ~/.agents/skills/hallmark for Codex CLI, and Cursor and OpenCode pick up those copies. There is also a Claude Code plugin route and a from-clone install script that validates every theme and reference before copying anything. Invoke it with /hallmark <prompt>, or simply describe a new build — and if it does not fire on its own, name it explicitly: “use hallmark to build…”.
A build then runs in four stages. First, pre-flight reads the project — fonts, palette, framework, any locked design document — and asks once for the real material: photos, screenshots, real numbers. Nothing is invented, which is itself a quiet rebuke to the usual agent habit of filling gaps with plausible fiction. Second, a theme script deals three themes that fit the prompt. Third, the hero section is built alone, rendered and reviewed cold before the rest of the page exists — the highest-leverage review point, because a bad hero poisons everything below it. Fourth, a lint script checks the page at eight viewport widths, and then a fresh-context reviewer agent gets the last word.
The gates come in tiers, and the tiering is what makes the system usable rather than preachy. The Floor never ships: contrast failures, broken focus states, missing reduced-motion support, invented facts, a broken phone layout. Reflex gates are the defaults a model falls into; a build with a reason may overrule one, on the record. Finish checks are craft detail that never blocks a ship. That distinction matters — a linter that blocks everything gets disabled; a linter that distinguishes “never” from “prefer not to” survives contact with real deadlines.
One practical note: the browser checks (rendered linting and screenshots) also need Puppeteer Core and Chrome installed globally. Without them, the static checks still run, but you lose the rendered review — which is arguably the whole point. Budget the setup step rather than discovering it mid-build.
Why it is trending now
Hallmark arrived at the exact moment its problem became unavoidable. Agent coding tools went mainstream in 2025 and 2026, and suddenly every team could generate a competent-looking page in minutes. Competent-looking was the trap: when everyone generates from the same models with similar prompts, output converges. The web filled up with pages that are individually fine and collectively identical. A tool that names the convergence and checks against it was always going to find an audience — and 39,000 stars in a young repository suggests it found one fast.
From our experience: where Hallmark helps, and where it will not save you
At MTD Technologies, we have spent nine years shipping client websites and web applications, and the pattern Hallmark targets is one we recognise from the other side of the desk. The most common failure in agent-built pages is not ugliness — it is undifferentiated competence. A client cannot articulate why their new page feels generic, but their visitors can feel it, and it shows up in bounce rates and enquiry quality long before anyone blames the design system.
Where Hallmark genuinely helps, in our judgement, is as a team standard. If three developers and two agents all build pages, a shared, checkable floor beats a style guide nobody re-reads. The redesign mode’s kept-link guarantee and the hero-first review loop are the two features we would actually mandate in a client workflow, because they catch the two most expensive failure classes: broken functionality disguised as a redesign, and a weak first screen nobody revisits.
Where it will not save you: strategy and copy. Hallmark can stop your page looking machine-made, but it cannot decide what your page should say, who it is for, or what action matters most. A beautifully non-generic page with a muddled offer still loses to a plain page with a sharp one. Treat Hallmark as a quality floor for execution, and keep human judgement on positioning, messaging and the offer itself. Also be realistic about the theme catalogue: fifty themes is a strong starting set, but a brand with a mature identity will still need custom direction — which is why the custom mode, not the theme picker, is the long-term path for serious brands.
Cost and effort are refreshingly honest here. The skill is free and MIT-licensed; the real costs are Chrome and Puppeteer setup, the extra review pass on every page, and the discipline to let a gate block a ship. Teams that skip the rendered check to save ten minutes skip the feature that justifies the tool.
Getting started with Hallmark in five steps
- Install Node 18 or later, then run
npx skills add nutlope/hallmark -g— or drop the-gflag for a single-project install. - For rendered linting and screenshots, install Puppeteer Core and Chrome globally; verify the static checks still run without them so you know what you lose.
- Start with a low-stakes page — a new landing page, not a client redesign — and invoke
/hallmark <prompt>with real material: real copy, real numbers, real screenshots. - Let the hero-first review run, and read the lint output at all eight widths before approving the rest of the page.
- Only then try
/hallmark audit <path>on an existing page, and/hallmark redesign <path>once you trust the kept-link guarantee on a page you can verify by hand.
Frequently asked questions
Is Hallmark free to use?
Yes. Hallmark is published under the MIT licence, so teams can use it, fork it and ship with it at no licence cost. The only costs are your agent’s usual model usage and the optional Chrome and Puppeteer setup for rendered checks.
Which coding agents does Hallmark work with?
Claude Code, Codex CLI, Cursor and OpenCode. Installation places the skill in the harness’s skills directory, and it is invoked as /hallmark <prompt> (or $hallmark in Codex). A Claude Code plugin route is also documented.
Does Hallmark replace a designer?
No. Hallmark sets a quality floor and checks a page against recognisable machine-made tells; it does not make positioning, messaging or brand-strategy decisions. For a mature brand, pair it with human design direction — its custom mode is built for exactly that.
Can Hallmark redesign an existing page safely?
That is one of its strongest features. The redesign mode records every link, field, handler and heading before rebuilding, fails if any go missing, and presents a side-by-side before and after. Audit mode reports problems without changing anything, which is the safe first step on a live page.
Do I need Chrome installed for Hallmark to work?
Only for the rendered checks. The static lint checks run without a browser, but the eight-width rendered review and screenshot capture need Puppeteer Core and Chrome. If you skip them, you lose the review step that makes the skill worth adopting.
Hallmark is worth trying precisely because it is verifiable: install it, build one real page, run the audit on an existing page, and judge the before-and-after yourself. If your team ships agent-built pages for clients, a shared floor that catches generic output before clients see it is not a nice-to-have — it is the difference between looking like an agency and looking like a prompt. If you want help folding tools like this into a client delivery workflow — from AI integration and automation to a full web development engagement — explore our AI Tools Directory for more vetted tools, or talk to us about your next build.
If you have shipped a page with a coding agent lately, you probably know the feeling: the layout works, the copy reads fine, and yet something about it feels instantly familiar. The same hero, the same eyebrow above the heading, the same three feature cards in a row. That is exactly the problem the anti-AI-slop design skill Claude Code developers are reaching for Hallmark to solve. Hallmark is an open-source design skill for Claude Code, Codex, Cursor and OpenCode that refuses to let an agent-built page look AI-generated — and with more than 39,000 GitHub stars, it is currently one of the most closely watched answers to a very real quality problem.
This is not another component library or prompt pack. It is a working method: fifty design themes, a build process that starts from what the visitor came to do, and a lint step that checks the finished page in a real browser. Here we explain what Hallmark is, how it works, what it costs (nothing — it is MIT-licensed), and where its limits sit.
What Hallmark is — and what AI slop actually looks like
Hallmark describes itself in one line on GitHub: “A design skill for Claude Code, Codex, Cursor and OpenCode that refuses to look AI-generated.” It is made by Together AI and published under the MIT licence, which means any team can use it, fork it and ship with it. At the time of writing it has around 39,582 stars and 2,070 forks on GitHub, and it was updated as recently as 28 October 2026 — a useful signal, because agent skills go stale fast as the underlying models change.
The term “AI slop” is doing a lot of work here, so it is worth pinning down. Hallmark’s own documentation names the tells it hunts: the eyebrow label above every heading, tracked-capital labels everywhere, one coloured word picked out in a headline, a page of empty frames or nothing but photographs, a hero section that misses its screen, an email sign-up tacked onto the end of every page, and a form that says “thanks” without actually sending anything. Any one of these is harmless. Together, they are a fingerprint. Visitors may not name them, but they feel the sameness — and for a business, sameness reads as cheap.
What it can do: fifty themes, built from the job
Version 2 of Hallmark ships fifty themes. Each theme is a material — paper, type, colour, surfaces, its own components and texture — with a picture of a finished page, and the build lays out its own page inside it. The examples in the repository range from Dusk, a dark theme shown with a bank-reconciliation product, to Atelier, a fragrance-atelier look, and Furrow, a farm-records aesthetic. The important detail is what a theme is not: it never dictates what a page contains or in what order. Two prompts in the same theme should produce two different pages, because the page’s structure comes from its job, not from a template.
“Built from the job” is Hallmark’s core design idea, and it is the part most agent workflows skip. A booking page books, a shop sells, a newsletter subscribes: the action sits where that category of page puts it, and the page ends on it. That sounds obvious until you watch an agent build a booking page that buries the booking form beneath three screens of marketing copy. Hallmark’s build process deals the page’s anatomy from its job first, then dresses it in a theme — structure before style, in that order.
- Fifty themes — each a single file of at most 8 KB with a reference picture; a build reads exactly one, so themes stay cheap in context.
- Custom mode — say “custom” and Hallmark drops the themes entirely, designing from your prompt alone under the same floor and gates.
- Real photographs — 755 openly licensed photos, picked per subject and credited on the page, rather than stock imagery or invented placeholders.
- Your real product — point it at a URL, a dev server or an app and it captures the screenshots itself instead of designing around lorem ipsum.
- Redesign and audit —
/hallmark redesign <path>rebuilds an existing page while keeping its content and wiring;/hallmark audit <path>only reports what is wrong.
The redesign mode deserves emphasis for agency work. Most “redesign” tools quietly drop links, form handlers or headings. Hallmark records every link, field, handler and heading before the redesign, then fails the redesign if one went missing, and hands over a side-by-side before and after. That is the kind of boring, checkable guarantee that makes a tool safe to run on a client’s live pages rather than only on fresh demos.
How the anti-AI-slop design skill Claude Code workflow runs
Installation is deliberately low-friction. The documented route is a single command, requiring Node 18 or later:
npx skills add nutlope/hallmark -g
The -g flag installs it for every project; leave it off to install into the current project only. It lands in ~/.claude/skills/hallmark for Claude Code and ~/.agents/skills/hallmark for Codex CLI, and Cursor and OpenCode pick up those copies. There is also a Claude Code plugin route and a from-clone install script that validates every theme and reference before copying anything. Invoke it with /hallmark <prompt>, or simply describe a new build — and if it does not fire on its own, name it explicitly: “use hallmark to build…”.
A build then runs in four stages. First, pre-flight reads the project — fonts, palette, framework, any locked design document — and asks once for the real material: photos, screenshots, real numbers. Nothing is invented, which is itself a quiet rebuke to the usual agent habit of filling gaps with plausible fiction. Second, a theme script deals three themes that fit the prompt. Third, the hero section is built alone, rendered and reviewed cold before the rest of the page exists — the highest-leverage review point, because a bad hero poisons everything below it. Fourth, a lint script checks the page at eight viewport widths, and then a fresh-context reviewer agent gets the last word.
The gates come in tiers, and the tiering is what makes the system usable rather than preachy. The Floor never ships: contrast failures, broken focus states, missing reduced-motion support, invented facts, a broken phone layout. Reflex gates are the defaults a model falls into; a build with a reason may overrule one, on the record. Finish checks are craft detail that never blocks a ship. That distinction matters — a linter that blocks everything gets disabled; a linter that distinguishes “never” from “prefer not to” survives contact with real deadlines.
One practical note: the browser checks (rendered linting and screenshots) also need Puppeteer Core and Chrome installed globally. Without them, the static checks still run, but you lose the rendered review — which is arguably the whole point. Budget the setup step rather than discovering it mid-build.
Why it is trending now
Hallmark arrived at the exact moment its problem became unavoidable. Agent coding tools went mainstream in 2025 and 2026, and suddenly every team could generate a competent-looking page in minutes. Competent-looking was the trap: when everyone generates from the same models with similar prompts, output converges. The web filled up with pages that are individually fine and collectively identical. A tool that names the convergence and checks against it was always going to find an audience — and 39,000 stars in a young repository suggests it found one fast.
From our experience: where Hallmark helps, and where it will not save you
At MTD Technologies, we have spent nine years shipping client websites and web applications, and the pattern Hallmark targets is one we recognise from the other side of the desk. The most common failure in agent-built pages is not ugliness — it is undifferentiated competence. A client cannot articulate why their new page feels generic, but their visitors can feel it, and it shows up in bounce rates and enquiry quality long before anyone blames the design system.
Where Hallmark genuinely helps, in our judgement, is as a team standard. If three developers and two agents all build pages, a shared, checkable floor beats a style guide nobody re-reads. The redesign mode’s kept-link guarantee and the hero-first review loop are the two features we would actually mandate in a client workflow, because they catch the two most expensive failure classes: broken functionality disguised as a redesign, and a weak first screen nobody revisits.
Where it will not save you: strategy and copy. Hallmark can stop your page looking machine-made, but it cannot decide what your page should say, who it is for, or what action matters most. A beautifully non-generic page with a muddled offer still loses to a plain page with a sharp one. Treat Hallmark as a quality floor for execution, and keep human judgement on positioning, messaging and the offer itself. Also be realistic about the theme catalogue: fifty themes is a strong starting set, but a brand with a mature identity will still need custom direction — which is why the custom mode, not the theme picker, is the long-term path for serious brands.
Cost and effort are refreshingly honest here. The skill is free and MIT-licensed; the real costs are Chrome and Puppeteer setup, the extra review pass on every page, and the discipline to let a gate block a ship. Teams that skip the rendered check to save ten minutes skip the feature that justifies the tool.
Getting started with Hallmark in five steps
- Install Node 18 or later, then run
npx skills add nutlope/hallmark -g— or drop the-gflag for a single-project install. - For rendered linting and screenshots, install Puppeteer Core and Chrome globally; verify the static checks still run without them so you know what you lose.
- Start with a low-stakes page — a new landing page, not a client redesign — and invoke
/hallmark <prompt>with real material: real copy, real numbers, real screenshots. - Let the hero-first review run, and read the lint output at all eight widths before approving the rest of the page.
- Only then try
/hallmark audit <path>on an existing page, and/hallmark redesign <path>once you trust the kept-link guarantee on a page you can verify by hand.
Frequently asked questions
Is Hallmark free to use?
Yes. Hallmark is published under the MIT licence, so teams can use it, fork it and ship with it at no licence cost. The only costs are your agent’s usual model usage and the optional Chrome and Puppeteer setup for rendered checks.
Which coding agents does Hallmark work with?
Claude Code, Codex CLI, Cursor and OpenCode. Installation places the skill in the harness’s skills directory, and it is invoked as /hallmark <prompt> (or $hallmark in Codex). A Claude Code plugin route is also documented.
Does Hallmark replace a designer?
No. Hallmark sets a quality floor and checks a page against recognisable machine-made tells; it does not make positioning, messaging or brand-strategy decisions. For a mature brand, pair it with human design direction — its custom mode is built for exactly that.
Can Hallmark redesign an existing page safely?
That is one of its strongest features. The redesign mode records every link, field, handler and heading before rebuilding, fails if any go missing, and presents a side-by-side before and after. Audit mode reports problems without changing anything, which is the safe first step on a live page.
Do I need Chrome installed for Hallmark to work?
Only for the rendered checks. The static lint checks run without a browser, but the eight-width rendered review and screenshot capture need Puppeteer Core and Chrome. If you skip them, you lose the review step that makes the skill worth adopting.
Hallmark is worth trying precisely because it is verifiable: install it, build one real page, run the audit on an existing page, and judge the before-and-after yourself. If your team ships agent-built pages for clients, a shared floor that catches generic output before clients see it is not a nice-to-have — it is the difference between looking like an agency and looking like a prompt. If you want help folding tools like this into a client delivery workflow — from AI integration and automation to a full web development engagement — explore our AI Tools Directory for more vetted tools, or talk to us about your next build.