# Craft Guard

> UI craft and feedback quality. Use when polishing product UI (motion, loading/error feedback) or when a page reads as AI slop.

- Skill: `bandersnatch0x/craft-guard` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add bandersnatch0x/craft-guard`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bandersnatch0x/craft-guard/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Bandersnatch0x (https://skillmd.com/u/bandersnatch0x)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bandersnatch0x/craft-guard

---


# craft-guard

**Craft** declaration for cross-product quality: hierarchy, wait feedback, purposeful motion, CJK-safe type. Does not own business risk (`domain`) or token inventory (`design`).

When a verified design-baseline binding exists (`status: ready`), use that path as the project-specific visual baseline. Report clear divergence (density, type, spacing, shape, motion, or primitive treatment) to `ui-evaluator` with source set to the bound path; do not replace valid project choices with this skill's generic taste defaults.

When a persistent contract v1 is bound for the run (`contract-bind.json` / `scripts/contract_v1.py`), honor decided visual/craft constraints from that contract and reject unknown contract schema versions. Do not invent page/component inheritance layers.

## Apply (checklist)

Treat as exhaustive for the surface under edit.

### Hierarchy and type

- One clear primary; secondary recedes via density/space, not rainbow chrome.  
- CJK UI uses the project Chinese stack (`var(--font-cn)` or equivalent).  
- Brand solid + brand-gradient **≤ 3** emphasis hits per viewport; neutrals carry the rest.

**Done when:** a cold reader can name the primary action/region in one glance.

### Loading tiers

```text
<300ms      no indicator
300ms–2s    skeleton (layout stable)
>2s         loader + what is happening
>10s        timeout + retry (or cancel)
```

**Done when:** every async path in scope maps to a tier with a next action on failure/timeout.

### Motion

- UI motion ≤ 300ms; animate `transform` / `opacity` only.  
- Each animation names the state change it explains.  
- Honor `prefers-reduced-motion`. Keyboard-triggered actions stay free of decorative motion.

**Done when:** every motion in scope states the state change it explains.

### AI slop → target look

For implemented UI, evaluate the applicability predicates of the registry entries in [`../design-playbook/references/rules.md`](../design-playbook/references/rules.md) — the subset declared by the active spec/contract, or the full catalog when none is declared. Run every entry whose predicate evaluates to `applicable` against rendered UI plus relevant source. Execution protocol and the row format live in [`references/detectors.md`](references/detectors.md). Write exactly one seven-column audit row per **applicable** entry to `.scratch/<run>/craft-guard.md`:

| ID@ver | Applicability | Predicate reason / missing proof | Result | Rendered evidence | Source evidence | Exception check | Positive fix |
| --- | --- | --- | --- | --- | --- | --- | --- |

`Applicability` is the entry's three-state predicate outcome: `applicable`, `not-applicable`, or `blocked`. **not-applicable and blocked both require an observable reason** (blank is invalid — never a silent skip). `Result` is `clear|hit` only for applicable rows, `-` otherwise; `Positive fix` is required on hit rows. Missing rendered or source proof is `blocked`, not a silent clear. Unknown registry IDs fail at this stage. Audit rows are advisory: record evidence, exception check, and positive fix; leave declaration source, severity, and verdict to `ui-evaluator`.

The host model may have no vision (text-only input): renders stay **path-bound** (the `Rendered evidence` column cites the artifact path; never read the image), and assertions run on the **text face** — HTML/CSS source, `a11y tree` text, interaction-trace JSON. A no-vision run evaluates the same registry protocol this way without degrading it; genuinely missing rendered proof still records `blocked` honestly.

| Push toward | Instead of default sludge |
| --- | --- |
| Accent on key noun + primary CTA | Purple–blue gradient wallpaper |
| Glass only on rare floating layers | Blur on every card |
| Project CJK stack | Inter/Roboto as Chinese body |
| Weighted modules | Equal white card grid |
| State-explaining motion | Bounce/elastic decoration |

[`references/craft.md`](references/craft.md) — **required when the surface has L4 interactive zones** (per-zone hover/motion affordance, with its own Done when); also holds edge craft (failure/permission feedback, rounded-corners/shadows, charts).

## Completion

**Done when:** hierarchy, every in-scope wait/fail path, every animation, and — when the surface has L4 interactive zones — every zone's affordance (per `references/craft.md`) pass their checks; implemented UI also records exactly one seven-column audit row for every registry craft entry whose applicability predicate evaluates to `applicable`, with no unexplained missing proof. Residual hits and blocked proof are handed to `ui-evaluator`; audit rows do not assign source, severity, or verdict.