pbakaus/impeccable

The problem

AI coding agents can ship a frontend end-to-end, but they were trained on the same handful of SaaS landing pages. Without extra guidance the output converges: Inter everywhere, purple-to-blue gradients, cards inside cards, gray text on colored fills. The existing fix is a design skill or a long prompt, which relies entirely on the model remembering and obeying it, and gives you nothing to check afterwards. Impeccable packages that guidance as a skill plus a command set and a rules engine the agent and you can both run.

What's different

It starts from Anthropic's widely copied frontend-design skill, then adds two things that skill does not have. First, a persistence model: /impeccable init writes durable product facts (audience, constraints, voice) to PRODUCT.md and visual direction to DESIGN.md, so each later command reads context instead of re-deriving it from chat history. Second, a split between taste and checks: 24 commands (polish, critique, distill, harden, and so on) handle judgement via the LLM, while 62 deterministic detector rules run in the CLI and a browser extension with no model and no API key. That last part is the real differentiator — most design-prompt projects offer opinions only, and a rules engine makes anti-patterns auditable in CI rather than a matter of trust. Compared with a generic agent framework or a docs-only style guide, the trade-off is scope: it is opinionated about what good frontend design looks like, so you inherit its taste.

Who it's for

Frontend and full-stack developers already using Cursor, Claude Code, Codex, Copilot, Gemini CLI, or similar harnesses, who want agent-generated UI to clear a bar before review. Small teams that ship marketing sites or product surfaces fast will get the most from the init/polish/audit loop. It is a poor fit if you have a mature design system with its own tokens and reviewers — the bundled anti-pattern list will fight your choices — or if you work outside the supported harness/provider list and cannot run project hooks.

Cost of adoption

Node is needed only for the npx installer; the engine is a self-contained Rust binary downloaded on first run into ~/.impeccable/bin/, or vendored via Git submodule and linked into harness folders. That means new binaries fetched at install time and hook manifests written into provider configs, which some harnesses (Codex, Grok Build) require you to explicitly trust. Onboarding is a real step: reload the harness, run /impeccable init, and accept that PRODUCT.md/DESIGN.md become project files your team must maintain. Apache-2.0 licensing is permissive. The README is detailed on install paths but thin on how the detector rules are versioned or extended.

Verdict

This is a young project — created in November 2025 — but its traction is not a one-day fluke: roughly 86k stars and about 15.6k gained in the last month, ranking #14 on the monthly all-language board. Momentum is clearly still rising, and the deterministic rule engine gives it a reason to exist beyond prompt fashion. The risk is maturity: a 342-day-old codebase with fast-moving install and hook behavior, one board appearance, and no Hacker News discussion to sample outside opinion from. Worth adopting for a side project or a greenfield UI where you can tolerate churn; evaluate carefully before wiring it into a production team's workflow.