codeswithroh/tastemaker

Selection

Select when

• You want Claude Code to build or improve a UI from a product brief.
• You have screenshots or images whose actual colors should ground the design.
• You want fresh project palettes with machine-checked contrast pairings.
• You need project styling to remain consistent across later sessions and screens.
• You want preferences learned from explicit keep and reject decisions.
• You need landing-page structure to vary rather than repeat a standard AI template.
• You want photos, icons, illustrations, favicons, and motion handled in the same workflow.
• You need a ranked read-only audit of an existing interface.
• You want mechanical scans for generic UI patterns and motion mistakes.
• Your project can run Python 3 scripts locally.
• You use Claude Code's plugin or Agent Skills system.
• You are willing to review aesthetic judgments and third-party assets before shipping.

Boundaries

Avoid when

• You need a standalone visual editor or GUI.
• You require mathematically verified aesthetic quality rather than contrast verification.
• You cannot allow an agent or local scripts to read and write project design files.
• You prohibit network access to all external asset services.
• You need full Cursor support through SKILL.md conversion today.
• You expect bundled unDraw artwork without maintaining a local illustration library.
• You cannot add Pillow but require deterministic palette extraction from images.
• You cannot add Cairo dependencies but require PNG or favicon export paths that depend on them.
• You need an audit mode that automatically edits the interface.
• You must use a fixed corporate design system without any generated visual direction.
• You cannot review licensing, quality, and suitability of fetched third-party assets.
• You need independently audited security assurances.

Strengths

Capabilities

Risk profile

Risks and limitations

• A contrast pass verifies legibility thresholds, not whether a palette or completed design is aesthetically good.
• Taste classification, reference interpretation, mood matching, anti-slop judgments, and preference promotion remain heuristic.
• Font selection is limited to five curated mood pairings, and similar projects can receive the same pairing.
• Illustration matching depends on a finite user-maintained unDraw library and may repeat source art for similar concepts.
• High-quality illustration setup is not fully automatic because unDraw files are deliberately not bundled.
• Deterministic image palette extraction requires Python 3 and Pillow; the documented fallback is a less deterministic vision read.
• PNG and favicon export paths require cairosvg and a system Cairo library.
• Cursor conversion is documented as degraded because supporting references and scripts are stripped.
• Networked asset sourcing exposes asset search requests to allowlisted third-party services and depends on their availability.
• GSAP is the default motion path, which may conflict with projects that avoid that dependency or extensive animation.
• Uncertain: repository-authored security controls and claims were not independently audited in the supplied evidence.
• Uncertain: the root and skill security files conflict on whether vulnerabilities should be reported privately or through a public issue.

First-party files state that bundled Python helpers avoid subprocess, shell execution, eval, exec, and pickle; HTTPS and redirect-aware host allowlisting are enforced by a network guard. CI runs domain checks, security regression tests, and SVG validation on pushes, pull requests, and weekly. Allowed egress is documented for Iconify, Openverse, optional Pixabay, jsDelivr, and cdnjs. These are repository claims, not independently verified here. Reporting guidance conflicts: the root policy requests private advisories or email, while the nested skill policy says to open a security issue.