KEEP IT HANDS-ON functional ~ tested 2026-09-24
// sandboxed in host macOS Β· aarch64 Β· Node 24.13.1 Β·install log Β· why not fully functional: Verified install fidelity (13 of 13 skills byte-identical across two agent targets), structural integrity (strict YAML frontmatter, link resolution), and 23 of 24 documented API claims against the real npm packages (sonner 2.0.8, motion 13.4.2, Base UI 1.0.0-rc.0). The skill bodies are prompt contracts executed inside a coding agent session, which was not exercised end to end, and one documented claim (toast.getActiveToasts) fails against current sonner. Β·functional log

emilkowalski/skills

by Emil Kowalski · https://github.com/emilkowalski/skills · MIT · vunversioned (no tags or releases; repo state of 2026-09-23) · updated 2026-09-23
Don't install your next skill blind. One email a week: what's worth installing.

A working designer's taste, compressed into agent skills that cite exact curves, durations, and APIs instead of adjectives.

⚙ ⚙ ⚙ ⚙ ⚙ 4 / 5
quality 5/5
documentation 4/5
setup 5/5
value 5/5
ecosystem fit 4/5
// bottom line

This is the best-written doc-only design pack GearScope has tested. Every rule comes with a number or a named failure mode, registry installs land byte-identical, and 23 of 24 documented API claims verified against the real packages. Deductions are small but real: the only documented install command needs a TTY, ask-sonner still documents one API call that sonner 2.0 removed, and the repo ships no versions, tags, or changelog.

Get the skill digest The best and the broken from 100+ hands-on tests. One email a week.
install via skills CLI
$npx skills@latest add emilkowalski/skills
non-interactive (CI, cron, scripts)
$npx skills@latest add emilkowalski/skills --agent  -y

the README command opens an interactive picker that aborts without a TTY

install if

  • Frontend developers whose agents animate badly. The animate and review-animations pair replaces "make it feel polished" with a decision framework an agent can follow and a review it can fail; the escalation trigger list (transition: all, scale(0), ease-in on UI) names the exact patterns agents emit by default.
  • React Native and Expo teams. animate-expo (263 lines plus 385 lines of recipes) covers the RN-specific layer most web animation advice omits: which thread motion runs on, gesture handoff, haptics, and screen transitions with Reanimated and Expo Router.
  • Anyone building Sonner toasts or picking frontend libraries. ask-sonner is a full prop and troubleshooting reference (one stale entry notwithstanding), and pick-ui-library replaces "let the agent hand-roll a toast" with a curated list that names when to reach for motion versus plain CSS.
  • Swift developers. write-swift (396 lines) is a bonus outside the design lane: value types, Swift 6 concurrency, actors, and Swift Testing guidance for agents that default to dated Swift idioms.

What It Does

Thirteen agent skills that encode Emil Kowalski's design engineering practice (Vercel and Linear alumnus, author of Sonner and the Animations course) as prompt contracts for coding agents. The flagship emil-design-eng is a 674-line philosophy and ruleset for UI polish and animation decisions; around it sit build, review, audit, and lookup skills (animate, animate-expo, review-animations, improve-animations, find-animation-opportunities, animation-vocabulary, apple-design, mobile-native, pick-ui-library, prototype, ask-sonner, write-swift). The audience is frontend and React Native developers whose agents keep picking the wrong easing, the wrong duration, or the wrong library. Everything is markdown: 26 files, 7,284 lines, zero scripts, zero dependencies.

The Good

Every rule carries a number or a named failure mode. The skills never say "make it smooth". The animate skill's easing decision tree ends in concrete cubic-beziers (a strong ease-out of cubic-bezier(0.23, 1, 0.32, 1), a drawer curve of cubic-bezier(0.32, 0.72, 0, 1) credited to Ionic). The duration table assigns budgets per element class: button press feedback 100-160ms, tooltips 125-200ms, modals 200-500ms, with a hard "UI stays under 300ms" rule. The frequency table is a full decision path: actions seen 100+ times a day get no animation ever, with Raycast named as the reference implementation of that discipline.

The routing is the best in the design tier. Each frontmatter description states when to fire, when not to, and which sibling handles the adjacent job: animate points critiques to review-animations and whole-codebase audits to improve-animations; mobile-native points motion work to animate and React Native work to animate-expo. Descriptions run 155 to 738 chars, all under the 1024 ceiling. Three lookup skills (pick-ui-library, prototype, review-animations) set disable-model-invocation: true so they only run when explicitly summoned, which is the correct discipline for opinionated overrides.

Output contracts are enforced, not suggested. emil-design-eng and review-animations both require findings as a markdown table with Before, After, and Why columns, show the wrong format as an anti-example, and review-animations closes with an explicit Block or Approve decision plus a nine-step remedial hierarchy that starts at "delete the animation" and ends at polish. That is a review skill an agent can actually lose, which is rare.

Registry fidelity is perfect. The skills CLI landed all 13 skills for both the claude-code and hermes-agent targets in isolated environments, byte-identical across all 20 markdown files (SKILL.md plus the 7 progressive-disclosure references: two RECIPES, STANDARDS, AUDIT, PLAN-TEMPLATE, PICKER, API), and wrote a 3,193-byte skills-lock.json with per-skill SHA-256 hashes. Demand is real and current: 1,592,415 all-time installs across the 13 canonical skills.sh entries (emil-design-eng alone at 293,137, up from the scanner's 286K a day earlier), with 84 to 99 derivative hits per skill as other packs vendor the content.

Documented API claims survive verification. I asserted the ask-sonner API tables against sonner 2.0.8 installed from npm: the documented defaults all match the source constants (visibleToasts AMOUNT = 3, gap GAP = 14, duration LIFETIME = 4000, position bottom-right, hotkey alt+T), Infinity persistence, toasterId multi-toaster support, mobileOffset, containerAriaLabel, onDismiss and onAutoClose all present. The animate skill's useSpring example with stiffness and damping matches motion 13.4.2's motion/react export, and Base UI 1.0.0-rc.0 really does drive popover origins from the --transform-origin CSS variable the skills tell agents to use. All 27 external links cited across the pack returned HTTP 200.

The Bad

The only documented install command needs a TTY. README says npx skills@latest add emilkowalski/skills, which opens an interactive skill picker. With stdin closed it aborts: "Interactive prompt required but stdin is not a TTY. Nothing was installed. Use --agent (or --agent '*') and -y to run non-interactively." A human at a terminal is fine; a CI job, a cron, or any scripted bootstrap hits a dead end, and the recovery flags appear only in that error message, not in the README.

ask-sonner documents an API call that sonner 2.0 removed. Both SKILL.md and API.md tell agents to read active toasts with toast.getActiveToasts(). Sonner 2.0.0 (released 2025-02-18, current 2.0.8) ships getToasts() and getHistory() instead; the function the skill names does not exist at runtime. It is one entry in an otherwise accurate 64-line API table, but a guide skill whose selling point is exact props and defaults should not carry a 19-month-old rename of its own author's library.

No versions anywhere. Zero git tags, zero releases, no changelog, no version field in any of the 13 frontmatters (0 of 13 declare one). At 40,844 stars and 1.59M installs there is no way to pin, diff, or talk about a version of this pack, and the lock file hashes are the only change-detection mechanism consumers have.

Small hygiene residue. A 0-byte file named .pl sits at the repo root (stray artifact, no references), performance-cheatsheet.md (15 lines) lives outside any skill directory and is linked nowhere, and none of the 13 skills declare a license field in frontmatter (the root MIT LICENSE covers the repo, but per-skill declaration is the convention skills.sh consumers rely on). There is also no CI at all: no frontmatter lint, no link check, which is how the sonner drift and the .pl file survive unnoticed.

Smoke Test Results

Host-based run (macOS aarch64, Node 24.13.1, python3 with a PyYAML venv). The sbx microVM daemon is down pending a Docker update, so the canonical structural run executed directly on host with full logs captured. The repo under test: clone at commit d16ebe6 (2026-09-23).

Structural validation

$ git clone --depth 1 https://github.com/emilkowalski/skills.git

βœ… 548KB clone, 26 markdown files, 13 skill directories

$ python3 structure.sh (strict PyYAML frontmatter parse, 13 SKILL.md)

βœ… {"PASS": 50, "NOTE": 5, "FAIL": 0}: all 13 parse, names match directories, descriptions 155-738 chars

$ python3 structure.sh (relative link resolution, all .md files)

βœ… 26 of 26 relative markdown links resolve on disk

$ python3 structure.sh (progressive disclosure reachability)

βœ… 7 of 7 reference files (RECIPES x2, STANDARDS, AUDIT, PLAN-TEMPLATE, PICKER, API) linked from their SKILL.md

$ python3 structure.sh (stub detection)

βœ… 0 stubs: median SKILL.md 181 lines, min 85 (pick-ui-library), max 674 (emil-design-eng)

$ python3 structure.sh (opt-in gating)

βœ… 3 of 13 skills set disable-model-invocation: true (pick-ui-library, prototype, review-animations)

$ python3 structure.sh (license and hygiene)

βœ… root LICENSE present (MIT); NOTE: 0-byte .pl artifact at root, 0 of 13 skills declare a license field

$ curl -s skills.sh/api/search?q=emil-design-eng (registry probe)

βœ… canonical entry emilkowalski/skills/emil-design-eng at 293,137 installs; 13 of 13 repo skills indexed, 1,592,415 total

Pass rate: 8 of 8. Zero structural failures. The notes (.pl artifact, missing per-skill license fields) are recorded above in The Bad.

Full structure log β†’

Run A. Install probe, fresh workdir

$ npx skills@latest add emilkowalski/skills (README-verbatim, stdin closed)

❌ "Interactive prompt required but stdin is not a TTY. Nothing was installed." (exit 1, 13 skills found)

$ npx -y skills@latest add emilkowalski/skills --agent claude-code -y

βœ… 13 of 13 skills land under ./.claude/skills (exit 0)

$ npx -y skills@latest add emilkowalski/skills --agent hermes-agent -y

βœ… 13 of 13 skills land under ./.hermes/skills (exit 0)

$ cmp -s repo/skills/.md installed/.md (all 20 markdown files, both targets)

βœ… byte-identical 20 of 20; skills-lock.json written (3,193 bytes, per-skill SHA-256)

Pass rate: 3 of 4. The single failure is the README's only install command in any non-TTY context; the recovery flags exist but are documented only in the error text.

Full fresh log β†’

Run C. Functional verification (documented claims vs real packages)

$ npm install sonner motion @base-ui-components/react@1.0.0-rc.0

βœ… sonner 2.0.8, motion 13.4.2, Base UI 1.0.0-rc.0

$ node functional.sh (ask-sonner API surface, 15 checks)

❌ toast.getActiveToasts is not a function (documented in SKILL.md line 51 and API.md line 63); the runtime exports getToasts and getHistory instead

$ node functional.sh (documented defaults vs sonner source constants)

βœ… duration 4000, visibleToasts 3, gap 14, position bottom-right, hotkey alt+T, Infinity persistence, toasterId, mobileOffset, containerAriaLabel, onDismiss/onAutoClose, swipeDirections

$ node functional.sh (animate skill claims vs motion 13.4.2)

βœ… useSpring, animate, AnimatePresence all exported from motion/react

$ node functional.sh (transform-origin claim vs Base UI 1.0.0-rc.0)

βœ… --transform-origin referenced in 18 shipped files

$ python3 linkcheck.py (27 external URLs cited by the pack)

βœ… 27 of 27 return HTTP 200

Functional pass rate: 5 of 6. The one failure is a genuine doc defect, not a test artifact: I confirmed the rename by reading sonner's dist/index.js exports directly.

Full functional log β†’

What the runs tell you

Structure and install are spotless: strict-parseable frontmatter, zero broken links, byte-identical installs on both agent targets, and a hash-pinned lock file. The functional layer is where the single defect lives, and it is small: one stale function name in the Sonner guide against an otherwise fully verified API table. What could not be verified is the skills' effect on agent output quality, since that requires live coding-agent sessions; the prompt-contract content itself is the strongest GearScope has reviewed in the design lane.

Setup Walkthrough

  1. Run npx skills@latest add emilkowalski/skills in a real terminal. The picker lists all 13 skills; space-select or take all.
  2. In scripts, add --agent -y (for example --agent claude-code or --agent hermes-agent); without a TTY the bare command installs nothing.
  3. Note the landing directory is the current directory's agent config (./.claude/skills or ./.hermes/skills), not your global home config, unless you run the command from home.
  4. There is nothing else to configure: no dependencies, no environment variables, no API keys. Verify a skill landed with head ./.claude/skills/emil-design-eng/SKILL.md.

Alternatives

  1. pbakaus/impeccable - 59 deterministic anti-pattern detector rules that run with no LLM and no API key, plus browser iteration. Complementary rather than competing: impeccable catches violations in code, emilkowalski/skills teaches the agent to not produce them.
  2. Leonxlnx/taste-skill - the CN design-taste tier's demand giant (3.46M installs across 13 skills), broader frontend taste coverage including image generation prompts, less depth per animation decision than emil-design-eng.
  3. anthropics frontend-design skill (via claude-plugins-official) - the Anthropic seed skill both impeccable and much of this lane derive from; thinner, but first-party and bundled with Claude Code.
  4. nextlevelbuilder/ui-ux-pro-max-skill - a searchable catalog (88 styles, 192 palettes, 74 font pairings) for style-level decisions, where emilkowalski/skills is motion-and-craft-level; different layers of the same problem.
// review provenance
reviewed by
GearScope
tested
2026-09-24 · macOS (Apple Silicon)
last verified
2026-09-24
depth
HANDS-ON
sponsorship
none, ever
// share this review
// feedback
was this review helpful?

Don't install your next skill blind.

Every week: the shortlist of skills worth installing β€” and the ones to skip β€” from 100+ hands-on tests. No spam, no affiliate links.