Human Taste: Code
Evaluate code and software design through human taste -- the trained judgment that detects whether abstractions are right-sized, complexity is managed, and the system will be cheap to change.
This complements the UX-focused human-taste skill. For full research citations see references/research-sources.md.
Why This Matters
LLM-generated code is functional but measurably lower in design quality. Studies show 42-85% more code smells in AI-generated code compared to human-written code. Human taste for maintainability, abstraction quality, and structural elegance is what separates code that works from code that lasts.
Key insight: taste in code is not aesthetic preference -- it is the ability to anticipate future change cost and act on that foresight now.
Quick Start
When asked to evaluate code:
- Identify scope -- single function, module, class hierarchy, or system architecture
- Run the rubric below across all six dimensions
- Produce a Human Taste: Code Report using the output template
- Cite specific code -- reference actual lines, names, and patterns
Evaluation Rubric
Score each dimension 1-5. Anchor every score with concrete evidence from the code.
1. Abstraction Depth (weight: high)
Are modules deep -- simple interface, rich functionality hidden behind it?
| Score | Meaning |
|---|---|
| 1 | Shallow -- classes/functions expose implementation details; interface is as complex as internals |
| 2 | Leaky -- abstractions exist but callers need to know how things work inside |
| 3 | Adequate -- most modules hide internals, some leak |
| 4 | Deep -- simple interfaces hide substantial complexity (Unix file I/O pattern) |
| 5 | Elegant -- abstractions feel inevitable; you cannot imagine a simpler interface |
Look for: interface-to-implementation ratio, information hiding, whether callers need internal knowledge, general-purpose vs over-specialized APIs.
2. Conceptual Integrity (weight: high)
Does the codebase feel like one mind designed it?
| Score | Meaning |
|---|---|
| 1 | Fragmented -- multiple conflicting patterns, naming conventions, and styles |
| 2 | Inconsistent -- some unified areas but noticeable clashes |
| 3 | Mostly consistent -- follows conventions with occasional drift |
| 4 | Cohesive -- one clear style, one approach to common problems |
| 5 | Unified -- every part reinforces the same design philosophy |
Look for: naming consistency, error handling patterns, data flow conventions, one way to do common things vs many.
3. Change Cost (weight: high)
How expensive will it be to modify this code in six months?
| Score | Meaning |
|---|---|
| 1 | Brittle -- any change risks cascading failures; high coupling everywhere |
| 2 | Rigid -- changes require touching many files; dependencies are tangled |
| 3 | Manageable -- most changes are localized but some require careful coordination |
| 4 | Flexible -- clear boundaries; changes stay contained in their module |
| 5 | Supple -- designed for change; new requirements slot in naturally |
Look for: coupling between modules, dependency direction, use of interfaces/protocols, feature toggles, test coverage of boundaries.
4. Simplicity (weight: medium)
Is the code as simple as the problem allows -- and no simpler?
| Score | Meaning |
|---|---|
| 1 | Over-engineered -- abstractions for hypothetical futures; patterns for pattern's sake |
| 2 | Complex -- more indirection than the problem demands |
| 3 | Balanced -- complexity matches problem complexity |
| 4 | Clean -- direct solutions; easy to trace logic flow |
| 5 | Minimal -- nothing to remove; every line earns its place |
Look for: premature abstraction, unused generality, configuration surface area, inheritance depth vs composition, "astronaut architecture."
5. Readability (weight: medium)
Can a new team member understand this code without an oral tradition?
| Score | Meaning |
|---|---|
| 1 | Opaque -- requires significant effort to understand basic flow |
| 2 | Dense -- understandable with effort but easy to misread |
| 3 | Clear -- straightforward logic, reasonable naming |
| 4 | Transparent -- intent is obvious; naming tells the story |
| 5 | Self-documenting -- reads like well-written prose; no surprises |
Look for: naming precision, function length, nesting depth, comment quality (explains why, not what), consistent formatting.
6. Robustness (weight: low)
Does the code handle the real world -- not just the happy path?
| Score | Meaning |
|---|---|
| 1 | Fragile -- crashes on unexpected input; no error handling |
| 2 | Weak -- some error handling but inconsistent; edge cases ignored |
| 3 | Adequate -- common errors handled; some gaps |
| 4 | Solid -- errors handled consistently; graceful degradation |
| 5 | Resilient -- anticipates failure; recovers cleanly; observable |
Look for: input validation, error propagation strategy, timeout handling, null/undefined safety, logging, retry logic.
Output Template
Produce your evaluation in this format:
# Human Taste: Code Report
**Subject:** [what was evaluated -- file, module, system]
**Language:** [primary language]
**Date:** [date]
**Overall Score:** [weighted average, 1-5, one decimal] / 5
## Scores
| Dimension | Score | Key Evidence |
|-----------|-------|-------------|
| Abstraction Depth | X/5 | [specific observation with code reference] |
| Conceptual Integrity | X/5 | [specific observation] |
| Change Cost | X/5 | [specific observation] |
| Simplicity | X/5 | [specific observation] |
| Readability | X/5 | [specific observation] |
| Robustness | X/5 | [specific observation] |
## Strengths
- [concrete strength citing specific code]
- [concrete strength citing specific code]
## Issues
- **[severity: Critical/Major/Minor]**: [specific issue] -- [why it harms long-term quality] -- [suggested refactor]
## Verdict
[2-3 sentences: what works, what does not, and the single highest-impact refactor]
Weighted average formula: (AbstractionDepth*3 + ConceptualIntegrity*3 + ChangeCost*3 + Simplicity*2 + Readability*2 + Robustness*1) / 14
Comparing Implementations
When comparing two approaches:
- Run the rubric on each independently
- Add a Comparison Table with side-by-side scores
- Identify which approach wins on change cost specifically -- this is usually the deciding factor
- Note tradeoffs honestly -- sometimes the "uglier" code is the right choice for the constraint
Reviewing AI-Generated Code
AI-generated code has specific taste failure modes:
- Shallow modules -- many small functions/classes that just pass data through without hiding complexity
- Over-abstraction -- interface + abstract class + factory + builder for a problem that needs one function
- Inconsistent error handling -- some functions throw, some return nulls, some use result types in the same codebase
- Copy-paste variation -- similar but slightly different implementations of the same pattern
- Missing edge cases -- happy path works perfectly; error paths are afterthoughts
- Naming theater -- verbose names that sound precise but don't help (
AbstractSingletonProxyFactoryBean)
Flag these explicitly when you detect them.
When Not to Use This Skill
- UX/visual design evaluation (use the
human-tasteskill instead) - Writing/content quality (use the
human-taste-contentskill) - Pure performance optimization (taste is about design, not benchmarks)
- Style-only reviews (formatting, linting -- those are automated)
Additional Resources
- For full research citations and sources, see references/research-sources.md
- For worked examples of the rubric in action, see examples.md