Taste Stack

Keep Claude Code UI consistent across pages and projects

By Taste Stack · Updated · First-party guide

Design consistency needs a shared source of truth. Give Claude Code the same tokens, components and active visual rules for every page, then check that it actually used them. A skill can carry those rules; it cannot enforce them without review.

Preview EditorialInstallation guide

Find which source is drifting

Before rewriting prompts, compare two screens and name the disagreement: font, type scale, spacing, button radius, shadow, chart treatment or navigation. Check whether both pages import the same stylesheet and components. Two pages can read the same prose instruction while using different CSS values.

If you already have a component library, make it the first reference. Add a visual skill only where it complements that system. Installing several conflicting design skills is not a reliable way to combine their strengths.

Keep rules and implementation together

Use one token file for color, type, spacing and motion, and reuse components for recurring UI. Point the agent at the real files instead of describing a sidebar from memory. Taste Stack web packages include CSS variables, Tailwind v3 and v4 formats, and neutral JSON; select the format your project actually consumes.

  1. Record the active skill and version with the project, along with the model and brief used for a reference build.
  2. Choose one shared token source. Avoid copying a different set of literal color and spacing values into each screen.
  3. Create shared navigation, buttons, fields and table components before asking for several similar pages.
  4. Make existing project rules and the chosen skill agree. When they conflict, decide which source wins and remove the conflicting instruction.

Ask for a small, inspectable change

Build the first page, review it, then ask for the next page to reuse named files. A request to “make it consistent” leaves the agent to guess which decisions matter. Name the artifacts you want reused and ask it to report new design values.

Suggested prompt
Before changing the interface, read the project’s shared tokens, navigation and form components, and the installed taste skill. Build the pricing page by reusing those files. Do not introduce a second font scale, button component or shadow system. If a required component is missing, add it to the shared layer. Report any new design token and explain why an existing one was not enough.

Review each feature against your Tailwind and shadcn system

Before a feature starts, name the token files and customized components it must reuse. Review the actual imports, new values and small-screen output before accepting the work.

Use the prompt below as a review checklist. Enforcement depends on your own project checks; the published web examples do not validate enforcement in a customized shadcn or Expo app.

  1. Find actual token CSS and custom component paths. If components.json exists, inspect its aliases and CSS path; the file is optional.
  2. Record new literal values and variants, with exception reasons.
  3. Capture long-content and empty states at 375/768/1440px. Run existing checks and identify anything not run.
Suggested prompt
Implement this feature using the project’s existing tokens and customized components. First list their actual paths; inspect components.json only if present. Keep the current theme; do not reinstall or reinitialize shadcn. Name any missing capability before adding a component or variant. List new literal style values, variants and exception reasons. Capture actual screenshots at 375, 768 and 1440px with long content and empty states. Inspect wrapping and overflow. Run existing project checks. Report commands, results and screenshot paths; mark checks or captures not run and why.

Across repositories, pin the same version

A skill installed for one repository is not automatically present in another. Install the same approved version in each project, or use the agent’s machine-wide scope when that is appropriate. Keep the shared component or token package version aligned as well.

For a paid Taste Stack package, follow its license when sharing it with a team. Do not publish the package into a public repository. A different machine needs its own authorized install and the same project references; the agent does not inherit another project’s visual memory.

Use a repeatable visual check

Compare the same viewport sizes and the same kind of content on every screen. Include navigation, a form, a table and an error or empty state. Inspect the page with the longest heading and the largest figure, not just the one that makes the nicest screenshot.

  1. Capture both a desktop and phone view of the screens you are comparing.
  2. Check type families, column starts, spacing steps, corners and interactive states against the shared source.
  3. Fix an issue in the shared component when it affects multiple pages, then recheck those pages.
  4. Keep a small reference set and check it after a model, dependency or skill update. A successful download or a skill appearing in a list does not establish visual consistency.

What the published example establishes

Editorial’s published example includes a landing page and dashboard from one standard brief, at desktop and phone widths. It is a useful reference for seeing one identity across different screens, not proof of perfect consistency across every framework or repository.

If you mainly need layout and component enforcement, start with your actual design system. If your missing piece is a chosen visual direction, preview a taste skill and evaluate it on one of your own tasks before expanding its use.

Continue with a real project

Open Editorial to inspect the current package, price, license and supported agents. For another direction, compare Swiss and Editorial.

Follow the quickstart or browse landing-page tastes.

Related field guides

Sources and scope

This is Taste Stack’s own product guidance. The package files and published product examples are the source for its visual rules. The suggested prompts are exercises, not additional completed benchmarks.