Agent Skills

UI audits, typography, docs, PR review, and releases, loaded on demand by your coding agent

26 skills for the parts of shipping that code review never covers: the loading states, the type scale, and whether half the diff is AI slop.

Install

npx skills add mblode/agent-skills -g --agent claude-code codex cursor -y

--agent takes a space-separated list. A comma-separated one is rejected whole as a single invalid name.

The bundle is listed at skills.sh.

<<<<<<< Updated upstream

Agents: these skills. Humans: Taste Training, a course on spotting and fixing AI slop. First unit free.

Cloud agents

A cloud agent runs on a fresh VM that never sees your machine's ~/.claude/skills or ~/.agents/skills, and Cursor's Sync Skills toggle covers only ~/.cursor/skills. Nothing you install locally reaches them.

Claude Code cloud, Cowork, and routines. This repo is a plugin marketplace. On claude.ai, open Customize > Plugins > Add > Add marketplace > Add from a repository, enter mblode/agent-skills, and install the Agent skills plugin. Every cloud session gets all the skills with no per-repo config, and Sync automatically keeps them current whenever the repo changes on GitHub.

To scope it to one repository instead, declare the marketplace there and Claude installs it at session start:

// .claude/settings.json
{
  "extraKnownMarketplaces": {
    "mblode-agent-skills": {
      "source": { "source": "github", "repo": "mblode/agent-skills" }
    }
  },
  "enabledPlugins": { "agent-skills@mblode-agent-skills": true }
}

Cursor cloud agents. Cursor has no account-level equivalent, so install on the VM at build time:

// .cursor/environment.json
{ "install": "npx skills add mblode/agent-skills -g --agent cursor -y" }

That lands in the VM's ~/.agents/skills, which Cursor loads, and leaves nothing untracked in the agent's diff.

Skills

Architecture

• codebase-architecture: Design a structure, deepen an existing one, or harden it with guardrails.
• scaffold-nextjs: Next.js turborepo with Blode UI, Ultracite, GitHub, and Vercel.
• scaffold-cli: TypeScript CLI package: ESM, tsdown, vitest, changesets, CI.
• multi-tenant-architecture: Tenant isolation, routing, and custom domains on Cloudflare or Vercel.

Design

• product-design: What the interface should do, before anyone builds it.
• ui-design: Visual direction, Tailwind builds, screenshot to markup, dark mode, responsive, and a React and Next.js UX audit with a ship verdict.
• ui-animation: Springs, gestures, easing, and curves pulled from a screen recording.
• presentation-creator: Decks with a story spine, speaker notes, and a contrast-checked QA pass, as Marp markdown, a web app, or a handoff to pptx.

Writing

• eli5: Session talk in plain language: one gist, one analogy, next action first, no AI prose.
• copywriting: Product and marketing copy, persuasion frameworks, brand voice charts, AI-ism removal.
• docs-writing: Diataxis doc types, 51 rules, audit and writing modes, agent-readable docs.
• readme-creator: Type-aware README from the manifests: one install, a runnable quickstart, badges only where published.
• seo-program: Keyword and prompt-volume research, writer briefs, AEO question maps, Search Console monitoring.

Personal voice, blog posts, and voice evaluation live in ghostwriter.

Quality

• ax-audit: Agentic experience audit: 27 rules for tool parity, approval gates, and escape hatches, ship verdict.
• dx-audit: Libraries, CLIs, SDKs, npm packages: 38 rules, agent-friendly checks, root-cause findings.
• typography-audit: 78 rules: punctuation, fonts, sizing, spacing, hierarchy, pairing.
• optimise-seo: Next.js App Router SEO: sitemaps, metadata, structured data, redirects, hreflang, AI-crawler policy, Core Web Vitals.

Shipping

• planning: Interrogates intent, writes the plan file plan mode presents, then scores it to 5/5 before any code.
• pr-reviewer: Read-only diff review: bugs, structure, AI slop, security.
• tidy: Hunts complexity across five angles and applies the simplifications.
• pr-creator: PRs with short human descriptions, Linear IDs, templates, drafts, tidied commits.
• pr-babysitter: Watches an open PR: conflicts, CI, comments. Fixes what it can.
• autoship: npm releases with changesets: fix loop, CI watch, Version Packages merge, OIDC publish verify.

Authoring

• agents-md: Wires a repo so Claude Code, Codex and Cursor read the same instructions, then audits and refactors AGENTS.md and CLAUDE.md, with grades.
• agent-skills-creator: Creates, audits, and rightsizes skills to the open format spec, with a validator and eval workflow.
• save-md: Writes a named source as a markdown file the next turn can reread.

License

MIT

---

Crafted by Matthew Blode