uncodixify

name	uncodixify
description	Use when building or refactoring React and Tailwind UI that keeps drifting into generic AI-dashboard styling and needs a repeatable visual audit loop.

Uncodixify

Overview

Use this skill when a frontend task needs strict anti-template UI constraints and an explicit audit loop. The goal is not "fancier" UI. The goal is restrained, product-shaped UI that avoids generic AI-dashboard tropes.

This package is intentionally isolated for eval work. Treat it as an eval target, not as part of the shared .agents/skills catalog.

When to Use

• React or Tailwind work is trending toward gradients, glassmorphism, oversized radii, or decorative dashboard filler.
• A task needs a repeatable check against banned UI patterns before completion.
• You want a with-skill vs baseline eval loop for frontend styling behavior.

Do not use this skill as a general design system replacement. It is a narrow constraint package.

Core Constraints

Before running the audit loop, keep these rules intact:

1. Use Tailwind v4 patterns.
2. Prefer dark, restrained base surfaces such as bg-zinc-950, bg-black, or bg-neutral-900.
3. Do not use purple accent classes.
4. Keep corner radii small: rounded-sm, rounded-md, or none.
5. Use subtle separators such as border-white/10 or divide-white/10.
6. Ban glassmorphism, gradients, dramatic shadows, floating panels, and eyebrow-label copy.

Audit Loop

1. Generate or edit the component.
2. Run ./bin/run-uncodixify.sh <path-to-file>.
3. If the validator fails, fix the specific issues it reports.
4. Repeat until the command exits with code 0.

The wrapper script resolves the package paths for you. Do not assume the current working directory is the package directory.

Eval Iterations

Use python3 ./bin/run-evals.py to scaffold an eval iteration under workspace/iteration-N/.

Each eval case gets:

• with_skill/ for the package-guided run
• without_skill/ for the baseline run
• prompt.txt, inputs/, outputs/, timing.json, and grading.json

After runs are complete, aggregate the benchmark with:

python3 ./bin/run-evals.py --aggregate-only --iteration-dir <absolute-iteration-path>

To grade mechanical assertions in the generated outputs/ directories and refresh benchmark.json in one step, run:

python3 ./bin/grade-evals.py --iteration-dir <absolute-iteration-path>

To run the package end to end against its bundled fixtures, use:

python3 ./bin/run-package-evals.py

Package Files

• manifesto.md: expanded design rules and banned patterns.
• evals/evals.json: starter eval cases for iteration work.
• toolchain/: package-local evaluator implementation.
• workspace/: ignored eval outputs, iteration folders, and review artifacts.