What is Hallmark?
Hallmark, by Hassan El Mghari (nutlope) and published under Together AI's name, is an Agent Skill that changes how a coding agent designs web pages. It sets a Floor of 16 numbered rules no build may break, a longer table of Reflex gates for the defaults models fall into, and a catalogue of 50 themes, each one file of at most 8 KB describing paper, type, palette and surfaces but never layout. A build asks once for real material, deals three themes, builds the hero first and runs a bundled linter, sloplint, before a fresh-context subagent reviews the page.
It installs with npx skills add nutlope/hallmark -g into Claude Code and Codex CLI folders, which Cursor and OpenCode also read, and ships as a Claude Code plugin. It needs Node 18 or later, plus puppeteer-core and Chrome for the browser checks. The verbs /hallmark redesign and /hallmark audit work on existing pages, which it otherwise leaves alone.
The repo is MIT-licensed with about 43,800 GitHub stars and 2,300 forks, created in April 2026. There are no GitHub releases; the skill is at version 2.0.0 and the last commit landed on 29 October 2026. Nothing is paid: the site says no API key is needed, and an optional TOGETHER_API_KEY only switches on generated images through Together AI.
Is Hallmark worth it?
Hallmark is the most prescriptive design skill we have read, and that is its strength. Where Anthropic's frontend-design and Impeccable mostly give an agent better taste to draw on, Hallmark writes the tells down as checkable rules and then checks them. SKILL.md bans, without exception, a purple-to-blue gradient, a card inside a card, an emoji as an icon, one italic word inside a roman heading, the eyebrow label above a heading and any label set in tracked capitals. Every colour and font must go through a token at :root, transition: all is forbidden, the headline, one sentence and the main button must fit in a 1280 by 800 window, and no screen may carry more than about 80 words. The fuller table in gates.md goes further: Inter, Roboto or Poppins as a display face, a Tailwind stock swatch such as indigo-500, gaps under 120px between sections and an email sign-up as the closing section all count as Reflex tells the build must fix or justify in writing.
What changes in the output is the shape of the page as much as its paint. The agent asks one question up front for photos, screenshots and real numbers, and will not invent metrics, testimonials or logo walls; without photos it places credited pictures from a list of 755 openly licensed images on Wikimedia Commons. It deals three of 50 themes, avoiding those in a log of earlier builds so consecutive pages do not converge, builds and reviews the hero alone, then runs sloplint at eight widths and hands the page to a subagent with fresh context. The result is self-contained HTML and CSS whose first CSS comment records the decisions taken. The maker's site shows Opus 5.5 alone scoring 26 issues against 0 with Hallmark, but the counter is Hallmark's own linter marking its own rules, which measures conformance more than quality.
The cost is context. For a page build SKILL.md tells the agent to load the layout, anatomy, imagery, craft, genre, theme and patterns files plus four sections of gates.md. In the repo those files come to about 95 KB before the 39 KB gates.md, which by the usual rule of four characters to a token is roughly 25,000 to 35,000 tokens of instructions before any markup is written. The skill itself calls loading more 'the largest avoidable cost', and the roadmap lists measuring a build's time and token cost as unfinished work, so there is no official figure. A fast mode exists for prototypes, and without puppeteer-core and Chrome only the static lint runs.
Two limits matter more than the token bill. A skill is a prompt, not a guarantee: the agent can skip steps, and the maintainers' eval notes that skills did not trigger on their own under claude -p and had to be named. And the taste is a house style. Bans on tracked-caps labels, cream paper with a terracotta accent or a closing newsletter band are defensible but opinionated; a team with its own system should write it into a design.md, which Hallmark reads first and lets override everything else. With no tagged releases and 68 open issues and pull requests, it is also young. For a new landing page from a one-line prompt it is the stronger tool than Impeccable; for a button or a settings screen its machinery is more than the job needs.
Where it earns its place
A solo founder launching a landing page for a webhook API
Get a page that does not look like every other generated SaaS site
Run /hallmark with the product's URL; it captures screenshots of the running app, deals three themes in the modern-minimal genre and lints the page before handing it back. Free; the cost is a large instruction load and two review subagents in agent tokens.
A frontend developer at a ten-person agency handed a client's dated marketing site
Restyle it without losing a form, a link or a tracking hook
/hallmark redesign ./site records every link, field, handler and heading with sloplint --before and fails the run with --kept if one goes missing; say keep the layout for a restyle or polish for only the tells. The handoff is a side-by-side before and after.
A design lead whose engineers generate pages with coding agents
Catch the generated look before a page reaches review
Run node skills/hallmark/scripts/sloplint.mjs on the output with --render, or /hallmark audit for a report that changes nothing; the linter reads HTML, CSS, framework components and Tailwind classes. Some rules are house opinions, so expect to waive a few on the record.
Before you commit
- Setup
- npx skills add nutlope/hallmark -g (drop -g for one project, add --agent to target one harness), or /plugin marketplace add nutlope/hallmark then /plugin install hallmark@hallmark in Claude Code. Needs Node 18+; npm i -g puppeteer-core and Chrome for the render checks. Invoke with /hallmark <prompt> ($hallmark in Codex).
- Integrations
- Claude Code, Codex CLI, Cursor and OpenCode, with GitHub Copilot as an install target; reads a project's design.md, Tailwind config and tokens; optional Inspo MCP server; an opt-in edit-time Floor hook for Claude Code; optional Together AI image generation (FLUX.1.1-pro by default).
- Data and privacy
- Runs locally and writes into your project, including a .hallmark/ folder with a build log and pre-flight cache. Library photos are hosted on Wikimedia Commons and keep their own CC0, public-domain, CC BY or CC BY-SA licences; prompts go to Together AI only if TOGETHER_API_KEY is set. The README and SKILL.md describe no telemetry. MIT licence.
- Cost at scale
- No paid tier or hosted edition. The real cost is agent tokens: a page build loads about 95 KB of instruction files plus parts of the 39 KB gates file and runs two fresh-context reviews. Generated images bill on your own Together AI account.
- Limits
- Built for websites and new pages, with reduced scopes for components and app screens; it ignores existing designs unless named. No tagged releases and 68 open issues and pull requests; skills may not trigger on their own; the maintainers have not published a measured token cost.
What’s good
- Concrete, checkable bans on the common tells of generated pages
- Bundled sloplint checks the page in a real browser at eight widths
- Refuses to invent metrics, quotes or logos; credits every photo
- Redesign mode fails if a link, field or handler goes missing
What’s not
- Heavy instruction load per build; no published token cost
- Opinionated house style that some teams will fight
- Browser checks need puppeteer-core and Chrome
- No tagged releases; may not fire unless named
Best for
Developers who build marketing pages and landing pages with Claude Code, Codex, Cursor or OpenCode and want them to stop looking like every other generated site.
Hallmark alternatives
Hallmark: common questions
- Is Hallmark worth it?
- Hallmark turns design taste into rules an agent can fail, which makes its pages look made rather than generated, at a context bill nobody has yet measured. We rate it 4.1 out of 5.
- How much does Hallmark cost?
- Free and open source (MIT). An optional TOGETHER_API_KEY enables generated images, billed by Together AI.
- What are the downsides of Hallmark?
- Heavy instruction load per build; no published token cost; opinionated house style that some teams will fight; browser checks need puppeteer-core and Chrome; no tagged releases; may not fire unless named.
- Who is Hallmark best for?
- Developers who build marketing pages and landing pages with Claude Code, Codex, Cursor or OpenCode and want them to stop looking like every other generated site.
- How much does Hallmark cost for a team or in production?
- No paid tier or hosted edition. The real cost is agent tokens: a page build loads about 95 KB of instruction files plus parts of the 39 KB gates file and runs two fresh-context reviews. Generated images bill on your own Together AI account.
- Is Hallmark safe for company data?
- Runs locally and writes into your project, including a .hallmark/ folder with a build log and pre-flight cache. Library photos are hosted on Wikimedia Commons and keep their own CC0, public-domain, CC BY or CC BY-SA licences; prompts go to Together AI only if TOGETHER_API_KEY is set. The README and SKILL.md describe no telemetry. MIT licence.
- What are the best Hallmark alternatives?
- The closest alternatives we have reviewed are Impeccable, OpenDesign and Anthropic Skills.
Answers checked Oct 31, 2026.
Make Hallmark? Show this rating on your site
