Skip to content
Impeccable: 62 Deterministic Rules to Eliminate Generic AI Look from Generated Frontend Pages

Impeccable: 62 Deterministic Rules to Eliminate Generic AI Look from Generated Frontend Pages

Impeccable provides coding agents with 24 design commands and a non-LLM checker containing 62 deterministic rules (31 for AI-style patterns, 31 for quality issues) to detect and remove repetitive AI-generated UI traits, plus a live in-browser editing mode that writes accepted changes directly back to source code across major frameworks and AI tools.

Architecture Digest
Architecture Digest
Architecture Digest

Problem: AI-Generated Pages Look the Same

When asking coding agents to build landing pages, the output consistently follows a predictable template: beige background, large serif headings, italic emphasis, pill-shaped buttons, rounded cards with flat vector illustrations. Changing models, projects, or prompts does not alter this uniform aesthetic. Impeccable addresses not whether the result looks good, but why it always looks the same.

What Impeccable Is

Impeccable is a design vocabulary for coding agents plus a deterministic checker that does not rely on an LLM. It consists of 24 commands grouped into five categories:

Direction (9): bolder, quieter, distill, colorize, typeset, layout, animate, delight, overdrive — each maps to a concrete design action, not vague adjectives.

Review (2): critique scores usability heuristics and orders fixes; audit checks accessibility, performance, theming, responsiveness, grading findings P0–P3.

Polish (6): polish, harden, clarify, adapt, optimize, onboard.

Browser Iteration (2): live and generate drive in-browser iteration.

Planning & Knowledge (4): shape, init, document, extract. The 24th command, craft, is deprecated (alias for a generic build request).

Author Paul Bakaus created jQuery UI (2007) and now leads Renaissance Geek (a16z seed); Impeccable is their first product.

Highlight 1: 62 Deterministic Rules Turn "AI Feel" into Checkable Conditions

Rules are defined in crates/foundation/src/registry.rs with explicit IDs, names, and descriptions. There are 31 slop rules (detecting AI-like patterns) and 31 quality rules (detecting genuine problems).

Slop Rules — Visual Patterns

side-tab

: thick colored bar on card side, described as "instantly recognizable trait of AI-generated interfaces." icon-tile-stack: rounded square icon above heading. hero-eyebrow-chip: small letter-spaced tag above main heading. cream-palette: beige background. italic-serif-display: italic serif display type. ai-color-palette: purple gradients and cyan on dark backgrounds.

Slop Rules — Copy Patterns

em-dash-overuse

: triggers only when ≥8 em dashes and density approaches 1 per 500 characters; comment notes humans use em dashes legitimately, so rule fires only at saturation. marketing-buzzword: catches words like "streamline", "empower", "supercharge". aphoristic-cadence: detects three consecutive sections ending with a short rebuttal sentence.

Thresholds Are Evidence-Based

In crates/foundation/src/rules/text.rs, constant LEADING_DISPLAY_TYPE_PX = 24.0 is derived from corpus review: every reading block judged "cramped" used ≤22px; false positives for display type started at 24px and continued to 51px, so the boundary was set at 24px. Validation sites listed in comments: culturedcode.com, mercury.com, stripe.com, apple.com.

Limitation: the 62 rules only enforce a floor. They remove predictable AI styles (colored bars, gradient text, nested cards) but cannot guarantee a distinctive, thoughtful design — that remains a human task.

Highlight 2: Live In-Browser Editing Writes Back to Source

Running /impeccable live overlays the running dev server. Select an element, give it a direction, it swaps variants in place; clicking Accept writes the change to the source file. Official tutorial screenshot shows selecting section#book-club, a yellow outline, and a bottom bar showing 2/3, Tune 2, ✓ Accept. This solves ambiguous feedback like "this feels cramped" — instead of guessing, the developer cycles variants and the chosen one commits directly.

Supported: Vite, Next.js, SvelteKit, Astro, Nuxt, plain HTML. Constraint: only modifies local checkout; do not inject into deployed sites or relax production CSP.

Highlight 3: One Rule Set Distributed to 24 Tool Formats

Repository root contains 24 distribution directories: .claude, .cursor, .agents, .gemini, .grok, .hermes, .kiro, .omp, .codex, .vibe, .veto, .dsh, plus cursor-plugin, plugin, skill. Each holds a full copy (70–78 files), including the 1 MB font-index.json duplicated 23 times (24.1 MB total, ~28% of the 86 MB repo). Manual copying loses hooks and subagents; updates require re-running npx impeccable update. Worth it if you work across two or three agents daily — avoids re-explaining design preferences in each tool.

Potential Issues

Web-only: Detector, live mode, and edit hooks target browser HTML/CSS. Native iOS/Android only have audit and adapt via direct source reading (SwiftUI, Compose, React Native, Flutter), not rendered output.

Hooks run outside model approval: In Claude Code, installed command hooks execute even if you reject the model's own command, so the first edit may download the engine. Review installed hooks or disable with --settings JSON containing "disableAllHooks": true.

Live Apply bypasses approvals: If no agent has polled in the last minute, Apply spawns codex (fallback claude) with --dangerously-bypass-approvals-and-sandbox or --permission-mode bypassPermissions, inheriting your environment variables. Disable by setting IMPECCABLE_LIVE_COPY_AGENT=chat or off. Not recommended on machines with payment/banking data.

Installation & Configuration

Prerequisite: Node ≥18.19 (only for npx; skills run on bundled engine).

Four Steps: npx impeccable install Installer detects your tools, asks project vs global scope and whether to install design hooks (accept recommended).

Reload your tool.

Run /impeccable init (Codex uses $impeccable init).

First-Time Setup: init reads code, asks only what it cannot infer: target users, product purpose, constraints. Writes PRODUCT.md; every subsequent command reads this file. Existing design systems: run /impeccable document to produce DESIGN.md so changes stay on-brand.

Verification: /impeccable doctor finds stale context files, unused settings, broken hooks. Standalone detector: npx impeccable detect src/ — exit code 2 = findings, 1 = unscannable target.

Common Pitfall: Hooks installed ≠ hooks active. Codex requires one-time approval in /hooks; Grok Build needs /hooks-trust or --trust flag; Hermes Agent must run hermes plugins enable impeccable-design-hook. If all three are installed but unapproved, the detector stays silent.

Time to running: ~10 minutes; hook approvals add ~2 minutes.

Author's Perspective

Evaluation criterion: does the tool turn what you want into something it can execute? Previously, telling an agent "make it premium" added gradients, glassmorphism, neon glows — because those correlate with "premium" in training data. bolder differs by having a fixed meaning: strengthen hierarchy, increase scale, swap to decisive type; gradients and glows are forbidden. An adjective becomes an action checklist.

Bakaus's background matters: jQuery UI taught how to make an interaction library usable by many; Chrome DevTools taught that developer tools avoided if slightly awkward. Both share: the good parts go unpraised, the bad parts get complained about daily.

Notable rejection: automatic mode requests are frequent, but PRs for it are closed because the focus is steering, not letting go.

Shortcoming reiterated: 62 rules protect the floor, the ceiling is human. Rules can say "no left colored bar", but cannot decide that this seaside inn should be seen at night.

Two audiences: frontend/design engineers who daily tweak UI in agents — install and use; backend developers who occasionally assemble a page — skip, the 62 rules in context become overhead.

GitHub repository: https://github.com/pbakaus/impeccable

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

frontend developmentopen sourceAI-assisted codingdesign systemscoding agentsImpeccabledeterministic ruleslive editing
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.