Hallmark, created by Together AI, establishes the baseline rules that every output must follow: colors and fonts must use :root variables, width and height must not be animated, keyboard focus styles and reduced-motion alternatives must be present, the title and primary button must be visible within 1280×800 in the first viewport, and data or customer testimonials must not be fabricated. The workflow begins by reading the project's fonts, color scheme, and framework, asking once about assets and audience, then selecting one of fifty themes and laying out the page according to its purpose. It starts by building the first viewport in isolation and checking the render, then uses the sloplint script to re-check across eight widths, and finally has a sub-agent with a fresh context review it. There are also two commands, /hallmark redesign and /hallmark audit, for modifying or merely auditing existing pages.
Why it matters It prohibits common model habits like gradient purples, eyebrow labels, and cards nested within cards, and then uses scripts to re-check in a real browser across multiple widths after completion.
skills.sh installs 118228 (+1569 / 24h)
Security Audit: Risk warnings present
- References unscanned scripts: scripts/sloplint.mjs, scripts/capture.mjs, scripts/theme.mjs
- Audit scope: GitHub nutlope/hallmark/skills/hallmark/SKILL.md (only the SKILL.md text was scanned)
Reasoning
- skills.sh installs 1569 in 24 hours, with a cumulative total of 118228 (indicating only attention level)
- SKILL.md explicitly requires not fabricating metrics, reviews, or customer logos; if assets are insufficient, it asks the user first; it only modifies existing designs when Hallmark is explicitly named
Usage
- Tell the agent, "Use Hallmark to create a landing page for my coffee subscription service," or run /hallmark audit ./index.html on an existing page to generate only an audit report
Risks
- The audit warning references unscanned scripts (sloplint, capture, theme), which aligns with their purpose: they are used for render checks, taking screenshots of your product, and extracting themes; the editor only read SKILL.md and README, not the scripts
- capture.mjs will open the URL, dev server, or local app you provide to take screenshots; if TOGETHER_API_KEY is set, it adds a step to call Together AI to generate reference images