md-to-pdf

Render HTML/CSS in anthropic.com visual aesthetic — warm cream background (#faf9f5), Poppins headings + Lora serif body, orange accent (#d97757) for CTAs and emphasis, rounded filled pill buttons, editorial card grids, abstract SVG illustrations, low-saturation data visualizations (soft blue/gray/teal), hand-drawn SVG architecture diagrams with orange/blue/green node categorization and diamond decision gates. TRIGGER when the user says 'anthropic 风格' / 'anthropic style' / 'claude 官网风格' / 'Anthropic 品牌', or asks for an editorial/long-form page, research article layout, pricing card grid, or a filled-button-with-warmth feel. DO NOT TRIGGER for generic 'beautiful web page' requests (use frontend-design) or Apple aesthetic (use apple-design).

Render HTML/CSS in apple.com visual aesthetic — white/pale-gray alternating sections, SF Pro typography, minimal text links (no filled buttons except Buy CTA), large statistic callouts, product-photography-driven layout, isometric infographics, hand-drawn SVG flow/architecture diagrams. TRIGGER when the user says 'apple 风格' / 'apple style' / '苹果官网风格' / 'like apple.com', or asks to design a landing page / slide / doc / diagram / spec page / configurator matching Apple's web look. DO NOT TRIGGER for native iOS/macOS UI (use an Apple HIG skill instead).

Use when the design harness should improve ITSELF — proposing better generator rules, templates, or techniques and keeping only the ones that measurably score higher. The evolutionary/self-improvement loop (harness component 09). TRIGGER on requests like "make the design skills better over time", "let the system evolve its own rules", "auto-tune the generators", or after a batch of pages reveals a recurring weakness in one skill. Runs a Darwin-style ratchet: diagnose the weakest critic axis → propose one change → regenerate → score with the FROZEN evaluator → keep iff it strictly beats the locked baseline and no held-out canonical regressed, else git-revert. Inspired by Karpathy's autoresearch (hill-climb on one editable asset against a frozen metric) and the Darwin / EvolveR self-improving-agent line. Pairs with design-review (the frozen evaluator) and learning-loop (which codifies one-off catches; design-evolve optimizes the rules themselves).

Use when the user gives a vague one-line brief for a web page ('做个 X 的页面' / 'we need a dashboard for Y') and a design skill (anthropic/apple/ember/sage/glass) will generate it. Expands the brief into a concrete plan: infer page-type + audience, pull the sprint contract via bin/design-review --plan (unknown page-types fall back to the nearest canonical, stamped LOW-CONFIDENCE), then write a section list with required content per section plus the hard numbers (diagram density, bilingual, brand presence). DO NOT TRIGGER when the user already provides a section-level spec, or for non-page work (components, refactors).

Independent evaluator for any design-skill output (anthropic-design / apple-design / ember-design / sage-design / glass-design). TRIGGER when a demo / template / landing page has just been written with one of the 5 design skills and is about to be shipped. Runs four gates — structural verify (placeholders, BEM, undefined classes, bilingual toggles), rendered visual-audit (contrast, hero diagram sizing, orphan cards, SVG text, 26 known-bugs), full-page screenshot, and LLM taste judgment (solo design-critic or 4 parallel specialists — composition / copy / illustration / brand). Pairs with design-learner to codify every critic miss so the same bug is never caught twice. Inspired by GAN's discriminator: this skill deliberately lives outside the generator skills so the reviewer does not inherit the generator's assumptions.