Impeccable Guru
impeccable is a design-quality layer for AI coding tools: one skill, 23 commands, a detector that flags anti-patterns as you edit, a documented design system the product itself must pass. this guru distills how the repo builds, enforces, and documents design quality — not how to run the tool.
the distilled areas
- product definition — users, purpose, principles: practice what you preach, show don't tell, editorial over marketing.
- the four modes — Persuade, Operate, Read, Experience: name the visitor's success before designing the surface.
- the craft floor — the mechanical quality checks and the category defaults to refuse.
- direction-finding — how new visual worlds are chosen: derivation from the audience's world, dice against the rut, verdicts and donations.
- refinement vs redesign — the semantic distinction that keeps scope honest.
- the command router — one skill, 23 commands, context-aware routing instead of menu pollution.
- documented design systems — DESIGN.md as a portable token export, written from the build, not before it.
why
every mechanism attacks one failure: AI-generated interfaces converge on category defaults. unnamed refusals mean every surface reinvents the same shapes — icon+heading+text card grids, hero metrics with eyebrow labels, gradient text, glass blur. the surfaces look generic, interchangeable, untrustworthy: "auto-generated, the company didn't decide what it should be." the countermeasures are structural:
- a roll that breaks the ranking rut
- a detector that flags the reflex
- a floor that refuses the lazy container
- a contract that survives the production build
each one makes the default answer wrong.
applied in this corpus
- honest states over polished fakes — truthful product states.
- token systems written from the build — tailwind token system · design systems.
- refinement discipline in product changes — iteration.
good and bad
bad
design review: make it feel more premium
no product mode, visual authority, or testable craft boundary. work is judged by feel: review debates taste while the mechanical defects — missing states, weak contrast, category reflex shapes — survive. iteration loops without a stop condition and converges on whatever the last reviewer preferred.
good
surface mode: operatescope: refinementquality floor: visible focus, semantic tokens, loading and error states
the decision gives the next design change an owner and verification criteria:
- named mode settles the trade-off arbiter
- named scope prevents silent identity replacement
- quality floor converts "feels premium" into checks that pass or fail on the built result
repeated across surfaces, this keeps a product looking decided rather than generated.