pbakaus/impeccable skill-v3.5.0

Skill 3.5.0

on GitHub

Dramatically better in GPT and Codex. Impeccable now compiles harness-specific rules into the skill at build time. Codex and GPT-5.5 get rules written for the exact defects they produce, that no other harness ever sees. No other design skill does this; everyone else ships one generic file to every model.

Far less AI-default slop. Bias mining against ~190 samples found the skill's own lines were causing defects, and cut them. The cream / beige background tell (74% on GPT) is banned across the warm-neutral band. Category-to-aesthetic recipes that drove extreme letter-spacing and italic-serif heroes are gone: the shape depends on the brand, not its category. New bans cover all-caps eyebrows, numbered section markers, more than three fonts, and a hard contrast floor.

It treats a new project differently from an existing one. On an existing codebase the skill reads your tokens, theme, and components first and works within them: preserving your identity wins over imposing a fresh look. On a blank slate it draws a single seed color from 129 hand-curated anchors, each carrying a mood and a composition strategy, then builds the full palette around it by hand. One seed becomes a dark jazz club or a light hospitality brand depending on the brief. Either way it ends the cold-start drift toward the same safe palette every time.

Live Mode is Beta. Out of alpha, and it now works at two scales. Type a direction into the new Steer bar, or speak it, and the agent reads the whole page and edits it in place, no element selected and no variant cycling. Or pick a single element, steer it in plain language, and accept the variant straight back to source. You can also edit copy in place: change an element's text right in the browser, and on Apply a subagent rewrites the real source it renders from and repairs anything wired to it, like card labels that double as image keys. Insert mode scaffolds brand-new elements between the ones already there. Recovery survives HMR, hidden heroes, and dev-tool overlays, and an experimental streaming poll mode cuts pickup from seconds to sub-second.

/impeccable teach is now /impeccable init. Renamed to match what it does: one command to set up a project. From a single codebase scan it writes PRODUCT.md, offers a DESIGN.md, configures Live Mode so it just works the first time you run it, then points you at the best command to start with. The old name still works as an alias.

A bare /impeccable recommends your next move. Run it with no command and, instead of a static menu, it reads the project, your dirty git tree, and your latest critique, then leads with the two or three highest-value commands and why (no DESIGN.md yet, run document; unresolved findings in the files you're editing, run polish). It always asks before running anything, and the full menu is still right below.

A faster detector with no jsdom. The HTML/CSS engine was rebuilt from the ground up on htmlparser2 and a real CSS cascade resolver, replacing jsdom. On the same 160-file HTML corpus it runs about 20x faster under Node: 0.34s where the old jsdom engine took 6.8s, roughly 2 ms per file instead of 43 ms. Dependency-free and small enough to bundle straight into the skill and run inline, not just in the CLI and the extension.

Detector: 14 new rules. cream-palette, em-dash-overuse, marketing-buzzword, numbered-section-markers, aphoristic-cadence, theater-slop-phrase, oversized-h1, extreme-negative-tracking, gpt-thin-border-wide-shadow, repeating-stripes-gradient, image-hover-transform, broken-image, text-overflow, and clipped-overflow-container. 41 deterministic rules total, one canonical registry feeding the CLI, the browser extension, critique, and the evals.

The skill keeps itself current. On the first session of the day, Impeccable quietly checks whether a newer version shipped. If one has, it offers to run npx impeccable skills update for you. It always asks first, never nags about a version you declined, and never interrupts the task you're on. Set IMPECCABLE_NO_UPDATE_CHECK=1 to turn it off.

Sharper craft under the hood. Beyond the bans, the craft itself got tighter. Small defaults landed where they pay off, like text-wrap: balance on headings, which cleaned up ragged hero type across the ablation runs. And the instructions themselves got leaner: orphan reference files folded into their commands, context loading simplified, and a new LLM-backed test suite that catches instruction-following regressions across three providers on every change. Plus dedicated /changelog and /faq pages.