A11y LLM Evaluation Harness and Dataset is a test suite for measuring how well language models generate accessible HTML. It runs agent-generated pages in a browser and evaluates them with automated accessibility checks.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add microsoft/a11y-llm-eval --skill building-accessible-uigit clone --depth 1 https://github.com/microsoft/a11y-llm-evalWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/microsoft/a11y-llm-eval/building-accessible-ui)<a href="https://agentmods.dev/skills/microsoft/a11y-llm-eval/building-accessible-ui"><img src="https://agentmods.dev/badge/skills/microsoft/a11y-llm-eval/building-accessible-ui.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00112 | $0.02435 |
| Opus 5 | $0.00056 | $0.01218 |
| Sonnet 5 | $0.00022 | $0.00487 |
| Haiku 4.5 | $0.00011 | $0.00244 |
Grade A, and why
building-accessible-ui scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 101 lines — stays where its author put it; the contents beside it link to each section on GitHub.
building-accessible-ui
Checklist for producing and reviewing accessible UIs. Each rule leads with the platform-agnostic principle and, where relevant, the Web (HTML + ARIA + CSS) implementation. Apply the web guidance only when the output is web.
Detailed rationale lives in references/; widget-specific guidance in components/. Open a file only when it's relevant to the current task. Do not preload. Every file opened and every line a tool prints stays in context — don't re-read.
Accessibility constitution
Ground rules. Use them to resolve conflicts and decide how much custom work is justified. The checklist below is their mechanical application.
1. Accessibility is a core outcome
A UI inaccessible to realistic users is not "done". Treat accessibility as a first-class criterion alongside correctness, performance, and security — not a finishing step. When scope must be cut, record the gap explicitly. Never claim output is "fully accessible"; state what was addressed and known limitations.
2. Build for real people
Evaluate designs against these personas; if a decision breaks one, justify it and offer an alternative:
- Screen reader — landmarks, headings, accessible name/role/state, reading order.
- Keyboard-only — Tab/arrows/Enter/Space/Escape, visible focus, no traps.
- Low-vision — zoom, reflow, contrast, Forced Colors.
- Cognitive — plain language, clear labels, actionable errors, forgiving interactions.
- Deaf / hard of hearing — captions/transcripts; no sound-only cues.
- Motor / voice / switch — large hit targets, named controls, no precise/timed gestures.
- Situational — sunlight, one-handed, noisy, flaky network.
3. Implementation priority
Use the highest option that fits:
- Existing accessible component in this codebase / design system.
- A component library.
- Native platform semantics (
<button>,<a href>,<input>,<label>,<fieldset>/<legend>,<dialog>,<details>,<nav>,<main>, headings). - Native element + minimum necessary ARIA (
aria-describedby,aria-expanded,aria-current, etc.). - Fully custom ARIA widget — only when nothing above fits, and only if you implement the APG keyboard, focus, and state behavior end-to-end.
What ships with it
21 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- AUTHORING.md 5.1 KB
- components/checkbox-group.md 4.4 KB
- components/disclosure-widget.md 3.4 KB
- components/forms.md 3.7 KB
- components/modal-dialog.md 6.7 KB
- components/page-layout.md 4.3 KB
- components/radio-button-group.md 4.4 KB
- components/README.md 762 B
- components/single-checkbox.md 2.2 KB
- references/contrast-forced-colors.md 2.9 KB
- references/images-graphics.md 2.4 KB
- references/keyboard-focus.md 3.9 KB
- references/navigation.md 2.2 KB
- references/reflow.md 2.1 KB
- references/specs-documentation.example.md 6.3 KB
- references/specs-documentation.md 5.9 KB
- references/specs-documentation.template.md 3.5 KB
- references/status-messages.md 4.3 KB
- references/structure.md 2.3 KB
- references/tables-grids.md 2.7 KB
- references/testing.md 5.7 KB
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 8d ago First seen · 101 lines · 112 tokens per session scan A 2d8fff1426dc
building-accessible-ui is a skill published in the GitHub repository microsoft/a11y-llm-eval (59 stars, last pushed 4mo ago), licensed MIT. It adds 112 tokens to every session and 2,435 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
chakra-ui-builder
Build responsive, accessible UI components and layouts using Chakra UI v3, install or configure Chakra UI in new and existing projects, and design scalable themes using tokens, semantic tokens, recipes, and slot recipes. Use this skill whenever a user asks to build, create, or generate any UI component, page, form…
visual-ralph
Visual Ralph orchestration for frontend UI from generated references, static references, or live URL targets, using $ralph with built-in visual verdict and pixel-diff evidence until the implementation matches and leaves a reproducible design system.
frontend-visual-qa
Audits already-rendered web, landing-page, HTML deck/slide, browser tool/game, dashboard/admin, design-system, and desktop UIs using real-browser or native-app journeys, inspected screenshots, DOM geometry, responsive or projection viewports, and a bundled Playwright sweep. Use after UI implementation to find…
prototype-web
A clickable, high-fidelity web product prototype with navigation, a hero section, feature cards, steps, social proof, and optional pricing. It is designed to resemble a finished landing page while remaining a prototype.
animation-principles
Apply animation principles — easing, staging, follow-through — to one specific UI motion. Use when tuning how an animation feels. For product-wide duration and easing tokens use motion-system (design-systems); for a full interaction spec use micro-interaction-spec.
refactoring-ui
Audit and fix visual hierarchy, spacing, color, and depth in web UIs. Use when the user mentions "my UI looks off" (or amateur/unprofessional), "fix the design", "Tailwind styling", "color palette", "visual hierarchy", "design system", "spacing scale", or "component styling". Also trigger when building consistent…