improve-animations (emilkowalski/skills)

Review

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a dedicated Windows test PC, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below). One small web app (am_app) with 8 planted motion problems. Tasks: audit the whole project, audit only the dropdown and modal, and fix five named problems. Prompts say 'animations and motion' without naming any rule. The one-liner lists the rules and asks for exact replacement values.

Three-arm comparison

Measured with Codex CLI over 18 runs, scored by rule-based checks with no LLM judge. Total 25.5 / 30, verdict Recommended.

Verdict: Recommended (25.5 / 30). Source: emilkowalski/skills, skills/improve-animations, pinned at commit a81dbca. Install: npx skills@latest add emilkowalski/skills --skill improve-animations. Licence: MIT (repo LICENSE).

What it is

The improve-animations skill (emilkowalski/skills): a motion-design audit guide: custom ease-out curves instead of ease-in/default, 150-300 ms durations, transform and opacity only, no transition: all, no scale(0) starts, transform-origin at the trigger, reduced-motion support, and no animation on high-frequency actions.

How we tested

Codex CLI 0.166 (gpt-6-astra, low reasoning) on a dedicated Windows test PC, a fresh temporary project and temporary HOME/TMPDIR/caches per run, skill installed (treatment) or absent (control) or absent with the skill's core requirement squeezed into one sentence appended to the prompt (one-liner group), 2 runs per group per task, network off, /tmp writes blocked. The prompts never mention the skill (except the wording noted below).

One small web app (am_app) with 8 planted motion problems. Tasks: audit the whole project, audit only the dropdown and modal, and fix five named problems. Prompts say 'animations and motion' without naming any rule. The one-liner lists the rules and asks for exact replacement values.

1. Audit the motion code — Audit the animations and motion in this project. Give me a prioritized list of what to fix, each with the file and the exact replacement values. Do not change any files.
2. Audit two components — Audit only the dropdown and the modal animations in styles.css. What is wrong and what exactly should they become? Do not change any files.
3. Top fixes for another person to implement — I'm handing the motion fixes to a colleague who has not seen this project. List the 5 highest-leverage animation fixes with the exact CSS to write for each, so they can implement them without asking me anything. Do not change any files.

Scoring is programmatic (hidden tests / planted-issue recall / constraint checks / render+OCR where relevant), 0–5 per run. No LLM judge decided any score.

Results

The one-liner given to the third group: "Audit the motion code like a senior design engineer: flag ease-in and default curves, transition: all, scale(0) starts, layout-property animation, missing reduced-motion, animation on high-frequency actions, wrong transform-origin and slow durations; for each give the exact replacement (custom ease-out cubic-bezier, 150-300 ms, transform/opacity only) so someone else can apply it, and change no source files."

Mean score: 3.15 without vs 4.81 with the skill (delta +1.66). Tokens per run: 65k vs 78k (1.20x); time 48s vs 49s. One-liner group: 4.94 (the skill's gain over the one-liner: -0.13; one-liner's own gain over no skill: +1.79); one-liner tokens 65k per run, so the skill costs 1.19x the one-liner.

What happened

• Control 3.15, skill 4.81, one-liner 4.94 (n=6 each). Without guidance Codex found 3-7 of the 8 planted problems and gave no cubic-bezier or exact easing curve in 5 of 6 runs.
• The skill closes that gap: 5.0 on both focus runs, where control missed the scale(0) modal start in both, and every audit gives exact curves. The one-liner reaches the same result.
• On the five-fixes task the skill was slightly behind the one-liner (4.62 vs 5.0): both treatment runs missed the 700 ms linear card hover, and each missed one other planted problem (scale(0) modal, or transition: all).
• Tokens: skill 78k, one-liner 65k, control 65k (1.2x control). The skill was read in 6/6 treatment runs without being named.

Dimension scores

Limits of this test

• Recall checks on the text of the audit and regex checks on the changed CSS; nothing was rendered or measured in a browser.
• The one-liner is a close paraphrase of the skill's own checklist, which makes it a strong comparison.
• One fixture; n=2 per cell.
• n = 2 per group per task: differences under about 1 point are noise.
• Rule-based scoring measures what the checks can see; it cannot judge taste or whether a result is the best possible.

Security

See security.md. Scan: 0 CRITICAL / 2 HIGH / 0 MEDIUM. Verdict: safe to run (no-network sandbox). A clean scan does not prove a skill safe.

Bottom line

The skill clearly beats an unprompted audit, mainly by always supplying concrete curves and durations. A single sentence listing the same rules does as well for less, so the gain is over no guidance, not over a good prompt.