Impeccable review: design skills, rule checks and a reproducible CLI test

A controlled test of Impeccable 4.8.0, with a practical distinction between design commands, deterministic detection, project context and Live Mode.

AgentIndex · Published 2026-11-04 · Updated 2026-11-04

Impeccable is worth evaluating when you already have a working interface and want a coding agent to improve specific problems: unclear pricing hierarchy, confusing form recovery, or crowded mobile content. It organizes design work into named commands and supplies a detector for repeatable checks of some interface patterns.

Our recommendation is to start with one existing component. Decide which checks come from fixed rules, which changes depend on the model, and how your team will preserve intentional brand choices. Those questions are more useful for adoption than a repository’s star count.

What we verified

The review date is November 4, 2026. We read the project README and the official detector, context, hooks and Live Mode documentation. We also ran the help command and two local HTML fixtures using the published npm package [email protected]. The detector test did not call a generative model or install skills or hooks into this project.

The hands-on result covers one controlled detector fixture only. We did not test model-generated design quality, user task completion, the browser extension, Live Mode, or differences between models. The adoption advice below is our interpretation of the documented mechanisms and this limited check, rather than a performance ranking.

Evaluate three capabilities separately

CapabilityWhat it suppliesWhat you must evaluate
Design skillsNamed tasks such as critique, typeset, layout and polishWhether model-generated changes serve the product’s users
DetectorReports on identifiable source or rendered-page patternsRule fit, false positives and missing coverage
Live ModeLocal page selection, variant comparison and acceptance into sourceFramework fit, source mapping and a successful build after acceptance

Named commands make a vague request easier to scope. Asking for a review of checkout error recovery gives you something more concrete to assess than asking for a more sophisticated site. The project recommends narrow targets too. The skills need a coding agent to perform the work; running a detector does not complete the design process. Project and command overview

Our detector check: four findings to none

The initial HTML fixture used Inter, a purple-blue gradient and an h1 followed by an h3, with no explicit heading sizes. We pinned the package so future releases would not silently change the reported result:

npm exec --yes --ignore-scripts [email protected] -- \
  impeccable detect --json before.html
FixtureObserved exit codeObserved rule identifiers
Before2overused-font, flat-type-hierarchy, skipped-heading, ai-color-palette
After0An empty array, []

The revised fixture used a system font, a white background and dark text, changed the h3 to an h2, and explicitly set heading sizes to 32px and 24px. There were other manual edits too, so this was not an experiment isolating each variable. The full inputs, captured results and reproduction steps are in our repository’s review fixtures.

The type-hierarchy finding deserves attention: the initial file had no explicit heading sizes, and the result reported headings and body text as 16px. Inspect the style context the detector sees before treating this as a statement about an actual browser screenshot. Applications using external CSS, runtime styling or component libraries should also be checked as running pages.

According to the detector documentation, exit code 2 means findings and code 1 means a target could not be scanned. A CI job should distinguish a reviewable finding from a failed scan. Source and URL modes also have different coverage; checks requiring rendered information, such as contrast, need the page mode. Detector reference

Put product constraints before commands

An incident console may need to communicate outage status quickly. An editor may need to remain comfortable during hours of reading and input. The same request for stronger emphasis should lead to different decisions in those products.

Impeccable uses PRODUCT.md for product context and DESIGN.md for an existing visual system. Review those files for user tasks, constraints and decisions to preserve before asking an agent to change components. Brand typography, spacing conventions and status colors should have reasons that survive the next session. Context-file reference

You do not need to replace a brand font just to clear a warning. Record intentional choices and evaluate readability and usability separately. An empty findings array means no executed rule matched this scan; it does not establish that the interface is well designed.

A practical adoption sequence

  1. Choose one component on a reversible development branch and record its current usability problem.
  2. Run the detector independently to understand its output and your intentional exceptions.
  3. Follow the official skill-installation flow and inspect the saved project context.
  4. Ask the agent to change one region, review the diff, and run the existing build and interaction checks.
  5. Consider automatic hooks or Live Mode after the manual workflow proves useful.

Hooks have a separate setup boundary. Installing a skill does not mean the host has allowed its automatic checks. The documentation lists host-specific trust and configuration steps; verify that a check actually runs rather than merely confirming a configuration file exists. Automatic checks

Accepting a Live Mode variant changes source code, and the documentation currently describes the mode as beta. Keep source review and build verification in the loop when comparing local alternatives. We did not exercise Live Mode in this review. Live Mode reference

Who should adopt it

Teams with an existing interface, clear user tasks and a habit of reviewing changes are good candidates for a small pilot. A detector can also give a team a consistent basis for discussing patterns repeatedly introduced by coding agents.

If your need is code formatting, start with your formatter. For accessibility acceptance, retain keyboard, screen-reader and rendered-page checks. If the product’s users are still undefined, resolve that first. Use a real component improvement to decide whether to expand adoption; the four-to-zero fixture result is insufficient evidence that the tool improves every interface.

Related reading: Cognee memory selection and Parlant customer-agent selection.