ISSUE 086  ·  17 NOVEMBER 2026 ELEN RSS
CodeCity
DESIGN № 0002

Impeccable — the design judgment for your AI coding agent

Your agent knows how to write code, but not how to judge design. Impeccable gives it that judgment. It is a skill, not a plugin and not a framework. It sits beside your tool and gives it vocabulary and boundaries.

Impeccable — σχεδιαστική κρίση για τον AI coding agent σου

What it does

Your agent knows how to write code. It does not know how to judge design. Impeccable gives it that judgment.

It is a skill. It is not a plugin, and it is not a framework. It sits beside your tool and gives it vocabulary and boundaries.

Why is it needed? Models grew up on the same SaaS templates. You can see the result everywhere. Inter everywhere. Floating blurs. Cards inside cards. Gray text on colored backgrounds.

Impeccable does not kill creativity. It puts a brake on the clichés.

It ships with 24 commands and 60 detector rules. The rules do not need an LLM or an API key. So they also run in CI.

What it does not do

It will not create design or a new visual identity for you. Do not ask it to generate mockups, redesign components, or replace a designer’s judgment. It does not automatically fix code. It gives findings and reasons. You decide what changes to make.

We found technical limits. The detectors work only on direct changes to UI files and lose much of their value when the UI comes from component libraries, templating, or styles generated at runtime (CSS-in-JS, theme tokens). In Codex, it needs a manual approval hook. In Claude and Codex, notifications come after the change; only in Cursor does it block first and write later. The rules are opinionated and change with each release.

Who it is for

For anyone shipping UI with an AI agent. For anyone who has been hoping the model will remember to look at the contrast of the text.

Installation

Run this from the root of the project:

npx impeccable install

The installer looks for the harness folders that exist (~/.claude, ~/.codex, .cursor). It asks whether you want project-local or global. It writes the right skill files for each tool.

Then open your tool and run:

/impeccable init

This creates PRODUCT.md. There you write who the users are, what the product does, and what it is not. If you skip it, the commands will give generic results.

Use in each LLM

Claude Code

This is the simplest path. Two options.

With the installer:

npx impeccable install --providers=claude

It installs only the Claude Code part.

Or through the plugin marketplace:

/plugin marketplace add pbakaus/impeccable

That downloads the plugin. Then you install it from /plugin.

The installer also writes a hook manifest in .claude/settings.local.json. The hook runs the detector on every file edit in the UI and sends back the findings. One warning matters: command hooks do not go through model approval. If you run unattended, read the hook before you leave it free.

Codex CLI

Codex has no /prompts: commands. Impeccable comes in as a skill. After installation, open /skills or write $impeccable.

There is one step that does not happen by itself. Go to /hooks and approve the project hook. Without approval, it does not run.

Cursor

The installer creates .cursor/ and .cursor/hooks.json. In Cursor, the hook works more directly. It blocks the bad change before it lands in the file. In Claude and Codex, the notice comes later.

Do not want hooks at all? Pass –no-hooks and skip them.

Where it wins, where it loses

Its strength is that the checks are deterministic. Same input, same finding. Every time.

Its weakness is coverage. The hook catches only direct changes to UI files. If the UI comes from component libraries or templating, it loses a lot. In Codex it needs manual approval. And the number of detector rules changes from release to release.

Test

I set up an HTML file full of clichés. Inter, mud gradient, border-left accent, bounce easing, gray text with low contrast. Then I ran:

npx impeccable detect slop.html

It returned ten findings, each with its reason: gray text on a colored background, contrast 1.3:1 instead of 4.5:1, Inter used as an overused typeface, cards inside cards, and a mud palette. The count depends on the file — the same classes on another HTML file give another total. Fast, no API key, on Node 26.

What I did not run: the /impeccable commands inside a harness. They need the tool installed. I also did not run npx impeccable install, because it is invasive and writes into harness folders. Everything else I say here comes from the documentation.

Verdict

Worth it. Do not expect it to design for you. It will give you the vocabulary and a safety net. The net matters. A few seconds, no cost, and it catches the same pattern that makes every AI frontend look like the one before it.

Score: 8.5/10.

Also see

Related, from the same category:

  • Lighthouse: the checker that opens your page and tells you what is failing
  • @shadcn/lint: the linter that corrects the practice, not the polish
  • shadcn-ui-mcp-server: the helper that reads component code, not guesses it

Sources:

https://github.com/pbakaus/impeccable

https://raw.githubusercontent.com/pbakaus/impeccable/main/README.md

https://api.github.com/repos/pbakaus/impeccable

https://github.com/pbakaus/impeccable/releases/tag/skill-v4.5.0

Εγκατάσταση

3 ΒΗΜΑΤΑ
  1. npx impeccable install
  2. npx impeccable install --providers=claude
  3. npx impeccable detect slop.html
Impeccable — the design judgment for your AI coding agent · ΑΠΟΤΕΛΕΣΜΑ 8.5/10
ΠΗΓΗ
github.com/pbakaus/impeccable