AGENT SKILL

impeccable

Use when the user asks for UI/UX work, wants to fix or polish a page/component, or when an interface looks bland, inconsistent, or like a default template.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
1

01 前言 Front Matter

name
impeccable
triggers
[ "user" ]
description
Use when the user asks for UI/UX work, wants to fix or polish a page/component, or when an interface looks bland, inconsistent, or like a default template.

03 SKILL.md 正文

Impeccable

Design vocabulary for production-grade frontend interfaces. Adapted from github.com/pbakaus/impeccable @ c3bb418 (Apache 2.0) + direction protocol from github.com/alchaincyf/huashu-design (MIT). Muscle: extensions/design-intelligence/ (design_lint.py, search.py).

When to use

  • User asks for UI/UX work: landing page, dashboard, component, app shell, form, settings, onboarding, empty state.
  • Existing UI looks bland, loud, inconsistent, or like a default template.
  • Need to audit, polish, or iterate a frontend surface.
  • Not for backend-only or non-UI tasks.

Context files

Two files hold durable truth. Everything else is transient.

PRODUCT.md - durable product truth (schema 1)

Fixed sections, in order: Platform - Users - Product Purpose - Positioning - Operating Context - Capabilities and Constraints - Brand Commitments - Evidence on Hand - Product Principles - Accessibility & Inclusion. Emit a <!-- impeccable:product-schema 1 --> marker near the top. Never invent claims, testimonials, benchmarks, pricing, or deployment facts - Evidence on Hand records what exists; absence is recorded, not fabricated.

DESIGN.md - visual system

YAML frontmatter (normative, the portable token export): name, description, colors: map in oklch/hex with per-token comments, plus typography/space/radius/motion maps as needed. After the frontmatter, fixed sections in order: 1. Overview - 2. The Kit - 3. Colors - 4. Typography - 5. Elevation and Material - 6. Components - 7. Do and Do Not. When a stylesheet/tokens file is the source of truth, say so in a comment inside the frontmatter and mirror values verbatim.

Command vocabulary (24, synced to upstream @ c3bb418)

Command Use when
impeccable General design request; pick next step.
shape Plan UX/UI before writing code (see Direction protocol).
init One-time capture of PRODUCT.md per the schema above.
document Generate DESIGN.md from existing project code (frontmatter + fixed sections).
extract Pull reusable components/tokens into a design system.
critique Design review with scoring and detection.
audit Technical quality check (layout, performance, accessibility).
polish Meticulous final pass between good and great.
bolder Push safe designs toward impact.
quieter Tone down designs that are shouting.
distill Ruthless subtraction; strip to essence.
harden Production-ready edge cases, i18n, errors, overflow.
onboard First-run experiences, empty states, paths to value.
animate Purposeful motion that conveys state.
colorize Add strategic color without going garish.
typeset Fix typography that feels generic or accidental.
layout Fix spacing, rhythm, alignment.
delight Small moments of personality.
clarify Rewrite confusing UX copy.
adapt Responsive or cross-context adaptation.
optimize UI performance (LCP, bundle size).
overdrive Push past conventional limits.
craft Deprecated alias for an ordinary new-work request.
live Iterate UI in the browser with variants.
generate Produce variants/alternatives of a named element to choose from.

Direction protocol (huashu port, used by shape and new surfaces)

  1. sec0 fact-verify - before any direction work, verify product facts (name, purpose, platform, versions, audience) with web_search/repo evidence. Write verified facts to product-facts.md; if a material fact cannot be verified, ask the user - never invent.
  2. Five Form Questions - answer all five against the facts: papel narrativo, distancia de audiencia, temperatura visual, capacidade/densidade, motivo visual. Map each to extensions/design-intelligence/search.py --domain (style, product, colors, typography) - the corpus dial, not taste.
  3. Direction gate - generate 6 direction candidates; select 3 real directions + 3 documented alternatives with rationale. Write direction-approved.md recording the choice, the alternatives rejected, and every exemption verbatim (exact user words granting each exception). Exemptions are never paraphrased.
  4. Only after the gate, produce the brief and (if asked) implementation.

Muscle wiring (deterministic first pass)

  • audit / critique: run python extensions/design-intelligence/design_lint.py <target> --json first; report its findings verbatim, then apply LLM judgment only to what the linter cannot see (narrative, hierarchy intent, brief fit). Exit 2 findings are deterministic facts, not opinions.
  • colorize / typeset / layout: query search.py "<brief>" --domain <colors|typography|style> before proposing; cite the retrieved rows.
  • React/Next/React Native/Expo targets: also consult extensions/design-intelligence/data/frontend-rules/<stack>/ before implementing.

Push vs Pull for Design Patterns

  • Push (to critique/audit): the reviewer or audit pass receives the design context - PRODUCT.md, DESIGN.md, direction-approved.md, anti-references, brand tone, surface type, and lint findings. Push these constraints before the critique so the reviewer verifies against them.
  • Pull (by implementer): the implementer reaches for the right command (shape, layout, harden, etc.), the lint output, and the direction artifacts before writing UI, instead of waiting for critique to catch generic design.

Frontend loop

When the user says "fix the design" without naming a command, run in order and stop when the result is good enough:

  1. audit - design_lint.py first; then a11y labels, focus rings, contrast, motion preference, no-JS fallbacks, overflow. Findings to .devin/plans/impeccable-<surface>-audit.md.
  2. critique - lint findings + top-3 judgment issues to .devin/plans/impeccable-<surface>-critique.md.
  3. distill (optional) - strip non-essential elements.
  4. layout - spacing, rhythm, alignment, breakpoints.
  5. typeset - hierarchy, line-height, weights, casing.
  6. colorize - all color uses map to theme tokens (primary, on-primary, surface); fix mismatched tokens.
  7. harden - reduced motion, keyboard nav, screen-reader labels, error states, long text, small viewports, dark mode.
  8. polish - shadows, transitions, radius consistency, icon sizing.

Bounded iteration (eval R7-13)

Loop critique -> fix -> re-lint until the gate passes or the budget is spent, whichever first. Gate: design_lint.py exits 0, or every finding is waived verbatim in direction-approved.md, plus the full verification checklist below. Max 3 critique-fix cycles per surface. Each cycle must close at least one finding or checklist item; a no-progress cycle ends the loop. If a cycle grows the file >20% without closing findings, stop and distill. Gate unmet at 3: write remaining findings to the audit plan file and stop.

Reference bar (optional, for high-stakes surfaces)

When a stronger gate than lint-clean is wanted, anchor scoring to references, not self-assessment (pattern from Leon Lin's cook skill, evidence .devin/research/2026-11-14-yt-batch3-reels10-survey.md T1):

  1. Attach 6-12 reference images before prompting; they are the minimum bar, not inspiration.
  2. Declare a numeric target per surface (for example 9/10 vs the chosen reference).
  3. Score against the reference with a separate judge or subagent - the builder never self-scores; fix the weakest dimension each round.
  4. Finalize the static design before the motion pass; motion references are screen recordings, not screenshots.
  5. Human approval is the ship gate, not the score.

State and motion coverage (eval R7-12)

During audit/harden, verify interactive components cover seven states: default, hover (pointer only, never on touch), focus (visible ring, never suppress outline), pressed (feedback <= 100ms), loading (in-control indicator, label preserved), disabled (visibly inert), and checked/toggled where applicable. Motion checks: transform over layout properties, translate over scaleX-stretch, ease-out (exponential preferred, no bounce/elastic per lint), <= 150ms hover, and a prefers-reduced-motion fallback for every animation.

Output rule

After each pass that produces code changes: run the project's tests/build; commit with a focused message.

Verification checklist

  • [ ] design_lint.py clean or every finding waived verbatim in direction-approved.md.
  • [ ] Color tokens are correct for each surface.
  • [ ] Focus states are visible.
  • [ ] Motion respects prefers-reduced-motion.
  • [ ] No-JS fallback works where applicable.
  • [ ] Tests and build pass.

Boundaries

  • Missing PRODUCT.md alone does not make a project greenfield; infer the incumbent world from code first, offer init after.
  • Refinement preserves the incumbent identity; redesign replaces the look but keeps product truth - never split the difference into polish on a discarded look.
  • The brief wins: honor pinned aesthetics, eras, fonts, palettes even when they collide with lint warnings - record the collision as a verbatim exemption.
  • Emblems tied to militarism, supremacy, or hate movements are never motifs or ornament; content documenting them as fact stays.

Cross-skills

  • a11y-audit - WCAG verification (contrast, focus, labels) before shipping.
  • creative-engineering - named visual styles/archetypes (CRT, dither, brutalist): load its retrieval step-0 and technical anatomy.
  • context7 - named UI framework (React, Vue, Svelte, Tailwind, shadcn): fetch docs before proposing implementation.

References

  • extensions/design-intelligence/ - local muscle (lint + BM25 corpus).
  • https://impeccable.style/slop/ - anti-pattern gallery behind the 62-rule registry.

04 同仓库更多 Skills

Leostruka
Leostruka/devin-bundle

montador-de-ideias

Use when assembling, expanding, or elevating a raw project idea or initial prompt into an ultra-detailed, first-principles dossier and compiled master prompt with journalistic OSINT market research, Tech Lead and Staff Architect justifications, atomic and modular hardening/optimization, and high-impact contact surfaces.

★ 0
Leostruka
Leostruka/devin-bundle

git-github-adaptive-engineering

Use when any task touches git history, GitHub issues/PRs/releases/CI, repo governance, or branch and merge decisions. Runs the Fase Zero situational-awareness battery, picks one of five delivery flows via the adaptive decision engine, and enforces commit/PR information-density standards. Orchestrates the gh, git-workflows, github-actions, pr, and debug/finish skills.

★ 0
Leostruka
Leostruka/devin-bundle

adaptive-agenda-calendar-routine-engine

Use when planning or executing work across hours, days, weeks, or sprints. Turns the calendar into a strict time-blocked agenda (HH:MM-HH:MM blocks carrying a routine profile, expected inputs, and a handoff output artifact), runs the 5-stage cognitive flow (Radar and Hunt -> Triage plus Steve Jobs conception -> workbench manifest -> >=32-cycle loop -> Close plus Open-Minded Scrum), and unfolds macro calendar milestones into micro agenda blocks through extensions/gcal/wrapper.py.

★ 0
Leostruka
Leostruka/devin-bundle

demo

Use when demonstrating a skill in a fixture project.

★ 0
Leostruka
Leostruka/devin-bundle

duplicate-skill

Use when testing duplicate skill name detection.

★ 0
Leostruka
Leostruka/devin-bundle

duplicate-skill

Use when testing duplicate skill name detection.

★ 0