Design rules for AI coding agents: 62 deterministic checks that run without an LLM or API key
Synopsis
Impeccable, a frontend design skill for AI coding agents, ships design constraints as runnable commands and detector rules. Earlier design skills of this kind leaned on the model's own judgment, whereas Impeccable's 62 deterministic detector rules run in a CLI and browser extension with no LLM and no API key. It also provides 24 commands as a shared design vocabulary, and one init step writes durable product truth into PRODUCT.md so later commands keep product facts separate from surface visual direction.
Interpretation
AI coding agents gain a callable frontend design skill that constrains interface output through commands rather than free improvisation. Earlier design skills of this kind are represented by Anthropic's frontend-design, from which Impeccable started; without guidance, models tend to repeat the same recognizable tells, such as Inter for everything, purple-to-blue gradients, cards nested in cards, and gray text on colored backgrounds. The skill carries an explicit avoid list naming overused fonts, gray text on colored backgrounds, pure black or gray, card wrapping and nesting, and bounce or elastic easing.
Design checking splits in two: deterministic rules run in local tooling, and the remaining critique goes to an LLM. The deterministic part does not depend on model inference, so the CLI and browser extension run those rules with no LLM and no API key. 62 deterministic detector rules sit alongside LLM-only critique checks, with the former executed by the CLI and browser extension.
One init step establishes shared context, and 24 commands form the design vocabulary used to talk to the agent. Init inspects the project, asks only for material gaps in durable product truth, and writes the result to PRODUCT.md; visitor mode and visual direction are chosen later per surface, and incumbent or newly built visual systems are recorded separately in DESIGN.md. Commands include polish, audit, critique, distill, animate, bolder and quieter, all reached through /impeccable, and pin creates standalone shortcuts.
Perspective
The skill targets frontend projects using supported AI coding tools, installed with npx impeccable install and configured per tool as hooks or a plugin, with a reload afterward. The material points to the Neo Mirai case study as a before/after reference for a real project but reports no quantitative result.
The material reports no quantitative evaluation, and the case for better interfaces rests on the avoid list and one before/after case. Which problems the deterministic rules cover versus the LLM-only critique checks is not spelled out.