Source: emilkowalski/skills · Maintainer: community: Emil Kowalski · License: MIT · Checked: 24 October 2026
What is the improve animations skill?
The improve animations skill, improve-animations, makes your agent audit every animation in a codebase, rank the problems by how much fixing them would help, and write a self-contained plan for each fix that any agent can execute later. It plans; it doesn't edit your source. It is for teams who want a roadmap of motion fixes across an app rather than a review of one pull request.
It is one of 15 skills in Emil Kowalski's emilkowalski/skills repo, where the README describes it as: "Audit all the animations in your codebase and get prioritized, self-contained plans that any agent can execute." The skill calls itself an advisor: the capable model does the judgment work (understanding the motion, deciding what matters, writing the spec) and execution goes to any agent, "including cheaper models."
As of 24 October 2026, its skills.sh page listed about 220,000 installs (snapshot, 24 October 2026).
What does it do?
The folder holds three files:
- SKILL.md: hard rules, the four-phase workflow and the invocation variants.
AUDIT.md: the eight audit categories, what to hunt for in each, and exact target values, such ascubic-bezier(0.23, 1, 0.32, 1)for a strong ease-out. The agent is told never to approximate these values, only copy them.PLAN-TEMPLATE.md: the required structure for every plan.
Invoked with no question, the agent replies only with "I'm ready to audit your animations and plan the fixes, my knowledge comes from Emil Kowalski's animation philosophy." and waits.
Hard rules. "Never modify source code." The only files it writes live under plans/, or animation-plans/ if plans/ already holds something else. No installs, builds with side effects, commits or formatters. Plans must be fully self-contained, because the executor has no context from your conversation. Repository content is treated as data, so a file that tries to give the agent instructions becomes a finding. Documented, deliberate motion decisions are respected rather than reported.
The four phases:
- Recon. Map the stack and motion libraries, where motion lives (tokens, Tailwind config, keyframes), existing easing and duration conventions, the product's personality, and a frequency map of which elements users hit 100+ times a day.
- Audit. Check the code against eight categories. On anything beyond a small repo, it fans out read-only subagents, one per category or per app area.
- Vet, prioritize, confirm. Re-read the cited code for every finding, drop anything by design or exempt, and present one table ordered by leverage (impact divided by effort), followed by two to four missed opportunities. Then it stops and waits for you to choose which findings become plans. Run non-interactively, it defaults to the top three to five.
- Write plans. One plan per selected finding, numbered
NNN-short-slug.md, stamped withgit rev-parse --short HEAD, plus aplans/README.mdwith order, dependencies and status.
The eight audit categories in AUDIT.md:
| Category | Examples of what it hunts for |
|---|---|
| Purpose and frequency | Animation on keyboard actions or a command palette toggle |
| Easing and duration | Any ease-in on UI, UI durations over 300ms |
| Physicality and origin | scale(0) entrances, popovers scaling from center |
| Interruptibility | @keyframes on toasts or toggles, drags without velocity-based dismissal |
| Performance | transition: all, animated layout properties, Framer Motion x/y/scale shorthands |
| Accessibility | Movement with no prefers-reduced-motion handling, ungated hover motion |
| Cohesion and tokens | Near-duplicate cubic-beziers, one bouncy component in a crisp app |
| Missed opportunities | State changes that teleport, rare moments with no delight |
Severity is fixed too. HIGH is feel-breaking, such as wrong easing on UI or dropped frames. MEDIUM is noticeably off, such as a wrong origin or missing reduced-motion handling. LOW is polish, such as stagger or token consolidation.
Plans. Each plan follows PLAN-TEMPLATE.md: status, commit, severity, category and scope, then Problem (with the current code quoted at file:line), Target (every value spelled out), repo conventions with an exemplar, ordered steps, boundaries, and verification. The template insists that "The feel check is not optional," so each plan tells the executor what to watch for in slow motion and with reduced motion turned on.
How do you install it?
The skills.sh page gives this command for the skills CLI:
npx skills add https://github.com/emilkowalski/skills --skill improve-animations
The repo README's command adds the whole collection, and the CLI lets you pick which skills to keep. Since execute reviews against the review-animations bar, you may want both:
npx skills@latest add emilkowalski/skills
Add -g to make it available across projects. For Claude Code, the folder lands in .claude/skills/improve-animations/ or ~/.claude/skills/improve-animations/, the standard Claude Code skill locations. Codex and Cursor read the same SKILL.md format; see Codex skills and Cursor skills.
On claude.ai, you can zip the folder and upload it under Customize > Skills (Claude help article on skills), though the workflow is built around a local repo, git and subagents, so a coding agent is the natural fit.
How do you use it?
Run it from the root of the repo you want audited. It can load automatically when you ask to improve the animations or audit the motion, or you can call it by name. The SKILL.md defines these variants:
| Invocation | What happens |
|---|---|
| bare | Full workflow: recon, audit all categories, vet, confirm, plans |
quick or deep |
Changes audit depth; combines with a focus |
a category, such as performance |
Recon plus that one category |
plan <description> |
Skips the audit and writes one plan for what you describe |
execute <plan> |
An executor subagent implements the plan in an isolated worktree, then its diff is reviewed |
reconcile |
Marks finished plans DONE, refreshes stale references, retires fixed findings |
quick covers high-traffic components with zero or one subagent and about five HIGH findings. standard, the default, covers all interactive UI with up to four subagents. deep covers the whole repo, including marketing pages, with up to eight.
Example prompts (suggestions, not quotes from the repo):
/improve-animations quick/improve-animations accessibility- "Use improve-animations to audit the motion in our dashboard app and write plans for the top three findings."
/improve-animations plan Make the settings drawer dismiss with a flick and rubber-band at the top./improve-animations reconcile
After the plans exist, hand any of them to another agent or session, or use the execute variant.
Tips and limits
- Subagents cost tokens.
standardanddeepcan start several read-only subagents. On a big repo, start withquickor a single category. See Claude Code subagents for how they work there. - It stops to ask. After the findings table, it waits for you to pick. Plan for that pause, or run it non-interactively to get the top three to five.
- Plans age. Each plan is stamped with a commit and tells the executor to stop and report if the code has drifted. Run
reconcileafter changes land. - Opinions, stated as rules. The 300ms ceiling, "No animation. Ever." for 100+ times a day actions, and the custom curves in AUDIT.md are Emil Kowalski's craft rules. Note deliberate exceptions in a design doc or comment, which the skill respects.
- Library claims age. Advice on Framer Motion shorthands and browser behavior reflects the time of writing. Check current docs before large refactors.
- "Already right" is allowed. The SKILL.md says "the motion here is already right" is a valid audit result.
- Security audit. As of 24 October 2026, skills.sh showed Pass from Gen Agent Trust Hub, Socket and Snyk. The folder has no scripts.
The repo's MIT License (copyright 2026 Emil Kowalski) lets you copy and change the skill as long as you keep the notice.
Save it in your Kipt library
In Kipt, open Skills and paste the skill's GitHub folder link into the import box. Kipt reads SKILL.md, AUDIT.md and PLAN-TEMPLATE.md; paste the repo link instead and Kipt shows a list of all 15 skills to pick from, so you can add review-animations alongside it. You can also drop the folder or a .zip. Your skills live in your own private GitHub repo and install into Claude Code or Codex with one command:
npx skills add https://github.com/<you>/<repo>.git --all -g
Kipt is not affiliated with Emil Kowalski or emilkowalski/skills.
Similar skills
- Review animations skill: a strict review of one diff, ending in Block or Approve.
- Emil Kowalski skill (emil-design-eng): the same rules as a guide while you build.
- Find animation opportunities skill: finds places that should animate but don't.
- Animation vocabulary skill: names a motion effect so you can ask for it precisely.
- Web design guidelines skill: Vercel's broader UI audit, including motion rules.