Case study: EvenBetter - Curated agent skills for iOS UX and accessibility
An open framework of agent skills that teaches AI agents to plan, audit and repair SwiftUI apps against Apple's HIG and WCAG 2.2. It began as a Mackenzie graduation thesis and became a paper accepted at JIIHC 2026 in Talca, Chile.
- Project
- EvenBetter
- Year
- Service
- Research, Agent Skills, Accessibility
Research
Code that compiles can still shut people out
- TCC · Mackenzie
- JIIHC 2026 · Talca
- Apple HIG and WCAG 2.2
- SwiftUI
- Agent Skills
- Claude Code and Codex
- Open source
Path
From thesis to workshop
From a Mackenzie TCC
Started as a graduation thesis at the Faculty of Computing and Informatics of Universidade Presbiteriana Mackenzie, by Marcello Gonzatto Birkan and Daniela Brazolin Flauto, advised by Prof. Dr. Valéria Farinazzo Martins.
A CLI, then a Mac app
The first two versions were a command-line tool and a macOS app. Both were hard to adopt, and every full analysis needed an expensive end-to-end LLM run.
Skills inside the agent
As skills, the knowledge runs in Claude Code or Codex with the login, permissions and tools the developer already has. Plan, analyze, validate and repair are separate skills, each with its own output.
A study with ten developers
Ten iOS developers spent about 30 minutes each exploring the site, planning an app, auditing a project and filling in the SUS. Mackenzie's research ethics committee approved the study.
Accepted at JIIHC 2026
Accepted at the 8th Workshop on Intelligent Interactive Systems (8WoSI²) of JIIHC 2026. The talk is on the program for 10 November 2026 in Talca, Chile.
Motivation & proof
Why curated skills, and what we measured
Of the million most-visited home pages had detectable WCAG failures (WebAIM Million, 2026)
Of screen-reader errors in apps found by a widely used checker (ScreenAudit, CHI 2025)
Of the same errors found by an LLM-driven audit. The work starts from that gap
Use or plan to use AI, while trust in its accuracy stays low (Stack Overflow Developer Survey, 2025)
Planning, design, audit, accessibility, performance and system. The version in the paper had 17
From Apple's HIG, Apple's accessibility documentation and WCAG 2.2, across seven domains
Ten iOS developers, median 81.25. It measures how usable the flow felt; the study did not measure whether the findings are right
All ten participants found the report's 55 findings: 15 critical and 40 high
Artifact
The living framework
Product · home
The research
Four steps
The corpus
Install in a minute
Stack & links
Open procedures for agents you already use
Technologies
- Agent Skills
- SKILL.md
- Claude Code
- Codex
- SwiftUI
- VoiceOver
- WCAG 2.2
- Apple HIG
- SUS
- Design Science
See the product at evenbetter.mellofattura.com, the method and results on the research page, the open framework on GitHub, the presentation slides and the JIIHC 2026 program.