Capture a design language as a portable design.md. The skill distinguishes public websites from working product interfaces and keeps evidence separate from proposed defaults.
The download includes the supporting references. Unzip capture-design into your agent’s skills directory.
Produce a portable design.md that lets an agent apply a captured aesthetic without seeing the source. Preserve the decisions that govern hierarchy, composition, and components, with enough detail to implement them.
Write lowercase design.md in the user’s target project, or the current project when unspecified. Save evidence under references/capture-design/. Preserve existing user decisions; keep distinct captures in separate user-selected folders.
The guide stands alone. Describe the design through rules and values; keep source images, proprietary components, and external source stylesheets as evidence rather than runtime dependencies. Aim for 600–1,000 words, scaled to the evidence.
Identify the intended use from the request: a public website, working product interface, editorial page, or another medium. Classify each reference by what it actually shows. A product screenshot supplies product evidence even when it appears inside a marketing page.
Use references that match the intended task. When surfaces differ, capture shared foundations and separate modes, with explicit conditions for choosing each. A public website alone establishes website rules; mark product coverage as missing if the intended application is inaccessible. Continue the supported capture and ask only when the missing scope materially changes the work.
For revisions, compare new evidence with existing guidance. Correct rules that the new evidence disproves and retain rules whose scope still holds.
Done when: the target use, reference roles, and coverage gaps are explicit.
Copy available inputs into the evidence folder using neutral filenames. Record source locations, capture date, surfaces, states, viewports, and missing assets in sources.md. Verify saved files exist and open correctly. Record any save failure and continue with inspectable inputs; request input if none is available.
For live URLs, read web-inspection.md before browsing. For static media, distinguish visible traits from unknown interactions and dimensions. Record exact fonts, libraries, and icon sets in the manifest only when verified; visual resemblance supports a description, not package attribution.
Done when: each inspected reference has a manifest entry, and saved evidence can be opened or its absence is recorded.
Identify three or four decisions that explain the design. Connect each to visible evidence and its applicable surface. Inspect relationships alongside primitives: reading order, type hierarchy, content density, alignment, grouping, pane boundaries, scrolling, and action priority.
For distinctive components, capture anatomy, dimensions or proportions, spacing, state, and icon treatment. Distinguish the interface from presentation around it: a backdrop framing a screenshot may belong outside the product window.
Separate measured values, visual estimates, and proposed defaults. Use qualitative descriptions when precision is unsupported. Describe a substitute font or component treatment as a proposal, keeping it distinct from source observations.
Done when: each defining decision has evidence, scope, and usable rules; material unknowns remain explicit.
Read output-template.md before drafting. State rules as role, value, and condition. Define the small set of type, spacing, color, shape, and component choices that carries the aesthetic. Explain which rules are shared and which belong to a specific mode.
Preserve task relationships so a new product can choose its own layout and content. Translate physical references into interface defaults only when requested, labeling that translation as proposed.
Done when: the guide contains implementation decisions for the captured scope and meets the neutrality contract below.
Read design.md as the only document reaching a new project. Check that it supports two different tasks or page types within the captured scope. Resolve contradictory rules, undefined values, unsupported precision, and rules applied beyond their evidence. Check neutrality across prose, examples, links, and code.
If the user also requests a sample, implementation, or correction, read application-review.md before applying the guide. A routine capture ends with document review.
Done when: the guide and manifest exist, evidence gaps are recorded, and each reported validation was actually performed.
Use an evocative, unbranded title, generic component roles, and semantic token names such as --color-surface. Describe typefaces by visual traits and generic family categories. Keep source product/company names, trademarks, logos, slogans, copied product text, source domains, branded font names, and branded identifiers out of design.md, authored captions, and filenames.
Raw evidence may retain source identity. Keep source URLs and implementation provenance in sources.md; keep reference images separate from reusable assets.
Report the guide’s path, defining decisions, captured surfaces, and evidence gaps. Building or publishing remains separate work unless requested.
When revising this skill or resolving its scope, read design-md-guidance.md. It records the authoring sources and their adaptation for portable guides.