A Claude Code skill for engineering taste and quality judgment. Evaluates code, architecture, product, design, and communication through a calibrated 5-dimension lens.
Use this skill when the question is "is this good?" rather than "does this work?"
- Quick Judgment — verdict + highest-leverage fix for single files or snippets
- Deep Review — dimension-by-dimension analysis for large codebases, architecture docs, or PRDs
- Guide — taste-informed code generation and refactoring with explicit design rationale
- Compare — judgment-oriented comparison between approaches, labeling taste vs objective tradeoffs
| Dimension | Core Tension | Ultimate Test |
|---|---|---|
| Code | Consistency > cleverness | Can a tired engineer at 2am understand this? |
| Architecture | Resilience > elegance | Will this survive the next 3 requirement changes? |
| Product | Composability > completeness | Does removing this make the product better? |
| Design | Intentionality > decoration | Does every pixel earn its place? |
| Communication | Teaching > describing | Can a newcomer understand the intent without asking? |
The skill activates on:
- "taste", "/taste", or the Chinese equivalent
- Subjective quality signals: "feels off", "amateur", "overengineered", "is this good"
- Explicit quality goals: "build X with good taste", "review the design quality"
Does not activate for: bug fixes, performance optimization, security audits, or functional requests with no quality signal.
taste/
├── SKILL.md # Main skill instructions
├── references/
│ ├── code-taste.md # Naming, functions, errors, cognitive load, consistency, deps, testing
│ ├── architecture-taste.md # Boundaries, coupling, evolution, state, failure modes, scaling, data
│ ├── product-taste.md # Feature composition, simplicity, mental models, defaults, scope, edge cases
│ ├── design-taste.md # Hierarchy, typography, spacing, color, motion, components, responsiveness
│ ├── communication-taste.md # Naming, API design, documentation, commits, PRs, error messages
│ └── scoring-rubric.md # 1-5 scale descriptors, calibration examples, surface scoring rules
└── agents/
├── reviewer.md # Deep review executor for large targets
└── advisor.md # Pre-generation taste guidance for greenfield code
Copy the taste/ directory to your Claude Code skills path:
cp -r taste/ ~/.claude/skills/taste/The skill will appear in Claude Code's available skills automatically.
Just ask naturally:
How's the taste of this code?
Deep review these 5 files for taste
Refactor this component with good taste
Which of these two implementations has better taste?
This skill was built with its own principles in mind:
- Judgment first, structure second. Quick prose judgment is the default. Scoring, templates, and dimension breakdowns are available but never forced.
- Evidence-driven. Only score dimensions with direct evidence. Skip the rest with a note, not a guess.
- No fake precision. No weighted overall scores. Dimension scores are optional shorthand, not the point.
- Boundary clarity. Each dimension owns specific territory. Naming mechanics go to Code; domain language goes to Communication. The same issue is never penalized twice.
- Respect the codebase. In generation mode, existing project patterns take priority over reference principles unless they're actively harmful.
Unix Philosophy, Dieter Rams, Paul Graham, Cognitive Load Theory, Christopher Alexander, Steve Krug, Rich Hickey, Sandi Metz, Kent Beck, Linus Torvalds.