Stop Guessing: Add Taste-Driven Review to Your Agent Workflow

Where this fits

You're mid-refactor, polishing a landing page headline, or deciding between two button labels. The code works, but the call feels subjective. Instead of guessing or pinging a designer, you want a fast, reasoned second opinion. That's exactly the gap taste-review fills—it shells out to the Claude Code CLI from your repo root and returns a grounded recommendation you can apply immediately.

Why agents benefit

• Makes judgment calls without human round-trips – Delegate subjective decisions to Claude's taste model and keep momentum.
• Provides structured recommendations with alternatives – The prompt asks for a primary pick and explains trade-offs, so you understand the reasoning.
• Adapts depth to decision complexity – A naming call gets a short answer; a design decision gets several paragraphs. No over- or under-explaining.
• Works reliably in sandboxed environments – Handles macOS OAuth token retrieval from Keychain with fallback to normal authentication.
• Reusable across many task types – UI polish, prose phrasing, naming, formatting—one skill covers a wide range of ambiguous calls.

Practical scenarios

Naming a new state variable

During a refactor, you're torn between isLoading and isBusy. Running taste-review with the file paths and the question yields a recommendation—say, isLoading for consistency with existing patterns—plus alternatives like isFetching when the variable tracks network requests.

Polishing a microcopy button label

Your checkout button currently reads "Submit." You want clarity without sounding robotic. The skill can weigh options like "Place Order," "Complete Purchase," or "Pay Now," and suggest the best fit based on the surrounding UI and tone.

Choosing between two CSS palette values

You're torn between two shades of blue for a primary action. Instead of eyeballing it, you pass the CSS file and the two hex values. taste-review factors contrast, accessibility, and brand feel, then recommends one with alternatives considered.

Add it to your agent workflow

Install the skill with the community CLI:

npx skills add bholmesdev/skills --skill taste-review

Then invoke it from your terminal with a plain prompt:

prompt="We need a new name for a function that validates and sanitizes user input. Files to consider: src/utils/validate.js. Weigh a few options, give your recommendation, and share alternatives."
claude -p "$prompt"

The skill wraps this call, handles authentication, and returns a decision you can act on immediately.

Compare before adopting

Before relying on taste-review, compare it against alternatives:

• Quality signals – Does the output consistently match your team's design instincts? Test it on a few past decisions.
• Maintenance freshness – Check the GitHub repo for recent commits and issue responses. A stale skill may break with new Claude CLI versions.
• Alternatives – You could write your own prompt or use a generic agent loop. taste-review's value is its opinionated prompt structure and OAuth handling.
• Workflow fit – Does it integrate with your existing sandbox and auth setup? The macOS Keychain path is a nice touch for long-lived tokens.

Why it is worth tracking

With 167 stars and a clear, focused purpose, taste-review shows healthy community traction. Its quality signals are strong: a well-tested prompt template, cross-platform auth handling, and a design that encourages reusable approval. For teams building agents that need taste-driven decision-making, this skill is worth trying now and revisiting as the ecosystem evolves. You can also find more details on the skill's directory page to see community feedback and usage patterns.