impeccable: a design language for AI harnesses

TL;DR: impeccable describes itself as “The design language that makes your AI harness better at design.” At 73,704 stars it is the largest of this morning’s trending repos, and the interesting part is that it targets the harness rather than the browser.

Reading the description carefully

“The design language that makes your AI harness better at design.”

Parse that and the target becomes clear. This is not a design system you consume at runtime in the browser. It is something an AI coding agent consumes while writing code. The audience is the model, not the end user.

That reframes what a design system has to contain. A conventional system ships components, tokens, and documented variants, all of which resolve at build time. A design language for a harness has to ship something a language model can follow: rules, constraints, and patterns stated in a form that survives being read in a prompt.

It is an unusual artifact. Design guidance normally assumes a human at a design system browser. This assumes a machine at a code editor, which is a different reading behaviour with different failure modes.

The numbers

git clone https://github.com/pbakaus/impeccable

Field	Value
Stars	73,704
Forks	4,439
Language	JavaScript
License	Apache-2.0
Created	2025-11-16
Last push	2026-10-02
Size on disk	380 MB
Home page	impeccable.style

The growth rate is the headline. The repository was created in November 2025 and crossed 73,000 stars inside eleven months, with commits landing on the day this was written. That is sustained interest rather than a launch spike.

The fork ratio sits at about seventeen to one, which is unusually high for a guidance project. People are not just reading it. A meaningful share have taken it into their own work, which is the behaviour you expect when the content is meant to be adapted to your own house style.

Two figures deserve a flag before you get excited. The primary language is JavaScript rather than TypeScript, which is unusual for a design tool shipping in 2025 and suggests the core is plain configuration and rules rather than a typed SDK. And at roughly 380 MB the repository is large for something that sounds like documentation. Expect assets, examples, or bundled tooling.

Apache-2.0 means you can fork, adapt, and ship it without asking.

Why this exists

AI coding agents produce competent code that looks like every other competent code sample. Spacing is correct, the layout is valid, and nothing is wrong in a way a linter can catch. The failure is that the result is generic.

A design language aimed at a harness tries to fix that by giving the model something better than generic to aim at. Whether it works depends on something no repository can measure for you: how much design specificity survives the round trip through a prompt.

What to check

• Read the license and any usage terms. Apache-2.0 covers code, and guidance content can carry separate terms worth a look.
• Find out what is inside the 380 MB before cloning it into a constrained environment.
• Test it on your own project, not a demo. Generic output is the problem, and a demo will not reproduce it.

Judge it on whether your agent’s output improves, not on the star count. Though at 73,704 stars, plenty of people already have.

Source and Accuracy Notes

• Project page: https://github.com/pbakaus/impeccable
• Source repository: pbakaus/impeccable (verified via GitHub API html_url)
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• Stars at time of writing: 73,704
• Counts, language, license, and dates were read from the GitHub API. Installation steps, rule content, and harness integration methods were not independently verified against the README.