musebot
New to Muse? Join at muse.ai/join and use invite code Referral: using this code gives the site owner credits.

muse skills

Muse skills for artifact crafting and building

How to add it to Muse

Copy the skill's SKILL.md (from the copy box or the source) and paste it into a Muse chat, asking Muse to follow it. This lasts for that conversation; saving skills isn't documented by Meta. In Muse Code, install it with `muse skills install` per the README.

New to Muse? Join at muse.ai/join and use invite code Referral: using this code gives the site owner credits.

There's no official Muse install link, so musebot never makes one up. Skill For the Muse app, web or WhatsApp.

Skill text to copy

---
name: "artifact_craft"
description: "Practitioner quality bar for interactive web artifacts and dashboards: information hierarchy, chart selection, low-friction interactivity, mobile/touch behavior, accessibility, performance, and a pre-ship checklist. Use when reviewing, hardening, or refining an artifact that is already planned or built."
---

# Artifact Craft

## Purpose
Turn a working artifact into a good one. This is the review and hardening pass, drawn from current practitioner guidance (2026 dashboard audits, Core Web Vitals practice, WCAG): what separates artifacts people keep using from ones they abandon. Complements `artifacts_builder`, which owns the build routes, design tokens, and build workflow.

## Workflow
1. **Audit first.** Run `references/pre_ship_checklist.md` against the artifact. Fix fails before adding polish.
2. **One screen, one decision.** Cut to 5-9 key metrics on the main view. Every remaining number must inform a decision; kill vanity metrics and the metric graveyard.
3. **Hierarchy.** Primary KPI top-left and large, secondary supporting, tertiary in tables. Sections get headings and breathing room. No same-weight-everywhere layouts, no important content in the bottom corner.
4. **Right chart for the data.** Use `references/chart_selection.md`. Let the data pick the chart, not preference.
5. **Context for every number.** Compare against a previous period or a target. Annotate spikes and dips. Show last-updated or data freshness. A lone number tells nothing.
6. **Frictionless interactivity.** Sensible filter defaults so the page answers the core question immediately on load, no five-click setup. Every control must do something; a decorative control is a bug.
7. **Mobile pass.** Three breakpoints: ≥1024 desktop layout unchanged, 768-1023 two-column cards, <768 single column. Filters collapse behind a toggle with an active-count badge. Tables become cards. 44px minimum touch targets. Modals go full-screen. KPI strip in a 2-column grid. 16px body text.
8. **Accessibility pass.** WCAG AA text contrast (4.5:1 minimum). Never encode meaning by color alone: add labels or patterns, then grayscale-test. All controls keyboard-operable. Honor `prefers-reduced-motion`.
9. **Performance pass.** Good band at p75: LCP ≤ 2.5s, INP ≤ 200ms, CLS ≤ 0.1. Keep initial JS under ~200KB gz. Lazy-load below the fold. Explicit width/height on all media. Defer or facade third-party scripts. `font-display: swap` on custom fonts.
10. **Re-run the checklist, then ship.**

## Output Contract
- The hardened artifact, or a precise list of checklist fails with fixes.
- All copy real, all numbers sourced or labeled as the user's own data.
- Stale data flagged with its age; never silent.

## Operating Rules
1. Audit before adding: a failing bar is never fixed by more widgets.
2. Cut before adding: a metric that cannot change a decision leaves.
3. Mobile is not a shrunken desktop; reflow content, don't scale it.
4. Performance is a design constraint, not a final step; re-audit after every major change.
5. Never present generated or sample data as fact.
6. One artifact, one slug, one job; changes go through the artifact's edit route, never a second create.

skills/artifact-craft/SKILL.md from the repo, MIT licence. Read it before pasting. Muse asks before it acts.

What the source says

Jev decision layer plus a 23-skill library for Muse agents: typed evaluations, shadow-mode routing, and a working terminal demo.

Quoted from the source, by theglitcharchitect. Licence: MIT.

Where we found it