* chore(stdlib): add conformance tooling for the skills library
The library had no gate that could express what a valid skill is, so
malformed frontmatter, over-budget files and credential literals went
unnoticed. These scripts define and enforce that contract.
- validate-skills.py: extend with Agent Skills v1.0 conformance -
naming rules and directory agreement, an explicit activation trigger
in the description, canonical layout, the 500-line SKILL.md budget,
relative-link integrity, and two security rules (no unrestricted
Bash(*) grant, no credential literals). Structural checks now read a
skill as a whole so material moved into references/ still counts.
- split-skill-references.py: move the heaviest sections of an oversized
SKILL.md into references/. Verifies line by line that every source
line lands in exactly one output and refuses to write otherwise.
- generate-skill-inventory.py: derive SKILLS_INVENTORY.md and
DIRECTORY_STRUCTURE.md from the tree so they cannot drift.
- generate-skill-catalog.py: index the canonical layout only, count
references/ toward a skill's size, and stop truncating descriptions
at 200 characters - the catalog is the discovery surface and must
carry the activation trigger.
- .gitignore: Python bytecode from these scripts, and local .pcl state.
Reviewed-by: Security Analyst, Compiler Engineer, DevX Engineer
* fix(stdlib): bring all 174 skills to Agent Skills v1.0 conformance
No skill in the library was conformant. 41 of 175 failed the existing
validator outright, and none satisfied the spec. This pass fixes the
tree and records the result. See stdlib/CHANGELOG.md for the per-skill
versioning and migration notes.
BREAKING CHANGE: one skill removed, five moved, seven with a capability
withdrawn. Library version 2.0.0 -> 3.0.0.
Structure
- Convert 71 skills stored as flat files to <category>/<name>/SKILL.md,
the layout the spec requires and the only one able to carry companion
resources.
- Rename 5 directories that disagreed with their declared name:, so
resolution by path and by name stop giving different answers. The
repository docs already referenced the corrected paths.
- Merge the duplicate legal-technology skills into legaltech-expert and
delete legal-tech-expert, carrying over everything the survivor
lacked.
Metadata
- Repair 20 unparseable frontmatters: an H1 preceded the --- block and
the key was skill_id: instead of name:, so any parser reading
frontmatter at offset 0 saw no metadata at all. The objectives,
prerequisites and outcome they carried become real body sections.
- Add the missing version: to 21 skills, fold keywords into tags, move
expertise_level under metadata, and align 10 category: values with
their directory.
Discoverability
- Give all 174 descriptions an explicit "Use when ..." trigger, built
from each skill's own tags and section headings. 152 previously said
only what the skill was about, leaving an agent nothing to match on.
Progressive disclosure
- Split 127 oversized skills, moving bulk material into references/ read
on demand and leaving an index behind. A SKILL.md now averages 277
lines instead of 644 and none exceeds 500. Verified lossless line by
line against the previous revision.
Security
- Withdraw the unrestricted Bash(*) grant from 7 skills; none of the six
non-financial ones contained a single shell command. finance-expert
documents code that charges cards and reads bank accounts, so it now
grants no shell at all.
- Remove credential literals: Stripe keys, two JWT signing keys, an
Express session secret, a Redis password and WiFi credentials, all
inlined in examples. They load from the environment now.
- finance-expert: idempotency keys were documented but absent, so a
retry double-charged; webhook signature failures were not caught, so a
forged event raised uncaught instead of being rejected;
int(amount * 100) truncated fractions of a cent. All fixed, plus a
money-movement guardrails section.
- trading-expert: place_order routed without any pre-trade risk check
although a RiskManager sat unused in the next section. Risk validation
now runs before routing and raises rather than returning a boolean a
caller can ignore; live routing is opt-in.
- Correct an Ansible Vault example that showed plaintext under an
"encrypted" label.
Also fixes 8 broken documentation links and a prematurely closed code
fence in skill-creator-expert that leaked template content into prose.
Per the repository file-organization policy, four stale progress reports
(SKILLS_MANIFEST, COMPLETE_SKILLS_LIBRARY, COMPLETE_LIBRARY_MANIFEST,
PHASE_3_DATA_ANALYTICS) move out of stdlib/ into the gitignored
.roadmap/status/stdlib/; they claimed 31 of 100 skills complete. They
appear here as deletions because .roadmap/ is not tracked.
Reviewed-by: Security Analyst, Language Designer, DevX Engineer,
Documentation Specialist
* fix(skills): resolve standard library skills from their real layout
No stdlib skill was resolvable. SkillResolver looked for
`<stdlibDir>/<name>.md` with stdlibDir defaulting to `stdlib/skills`, a
directory that does not exist, and the path ignored the category level
the library actually uses. The two LSP providers had the same wrong
directory and scanned only one level deep, so neither completion nor
go-to-definition could see a skill either.
- skill-discovery.ts: enumerate skill entry points across the shapes in
use - a flat `<name>.md`, `<name>/SKILL.md`, and
`<category>/<name>/SKILL.md` - while skipping references/, scripts/
and assets/, which hold supporting material rather than entry points.
- skill-resolver.ts: point stdlibDir at `stdlib`, and look the name up
as a directory under any category as well as the two legacy shapes, so
existing custom stdlib directories keep working. The name is used as a
single path segment and rejected if it contains a separator or
traversal, so a reference cannot escape the stdlib root.
- Rewire both LSP providers onto the shared discovery helper.
Tests: 13 new cases covering the layouts, ordering, the references/
exclusion, path-traversal rejection, and an integration test that
resolves real shipped skills by bare name.
Reviewed-by: Compiler Engineer, Runtime Architect, Security Analyst
* feat(stdlib): add 17 skills covering the library's largest gaps
Selection was driven by cross-referencing coverage against observed usage
rather than by intuition: Stack Overflow and GitHub Octoverse data for
which technologies matter, and install rankings on public agent-skill
registries for the shape of skills people actually use.
That comparison found a mismatch of form, not only of coverage. The
most-installed skills in the ecosystem are procedures - tdd, refactoring,
frontend design, browser automation - while all 174 existing skills were
domain knowledge shaped as <domain>-expert. The new workflows/ category
closes that gap and introduces the <activity>-workflow convention.
New category workflows/ (4)
- tdd-workflow, refactoring-workflow, debugging-workflow,
code-review-workflow. code-review-workflow complements rather than
replaces code-review-expert: procedure versus judgement. Merge them if
the split proves confusing in use.
Applied AI (3) - the largest coverage gap
- llm-engineering-expert, rag-expert, agent-engineering-expert.
The library referenced LLMs in tags but covered none of it: pgvector
appeared zero times across 174 skills, LangChain once, MCP twice
despite the repository shipping an MCP integration guide.
Capabilities (2)
- document-processing-expert (PDF, DOCX, XLSX, PPTX in one skill),
browser-automation-expert (a client against third-party sites,
distinct from playwright-expert which tests your own application).
Data (3)
- mysql-expert, the second most used database and previously absent
while rarer stores had skills; analytical-databases-expert;
stream-processing-expert.
Security (3)
- secrets-management-expert, supply-chain-security-expert (OWASP A03
2025), identity-access-expert (OWASP A01).
Interface and meta (2)
- accessibility-expert: design/ held only design-expert, which despite
its name covers software architecture, so WCAG had no home.
- skill-router: find and compose skills across 191, with a
disambiguation table for overlapping pairs.
Grouping by decision context rather than by technology keeps the growth
to 17 skills for roughly 50 topics. Five vector stores live inside
rag-expert; four document formats inside document-processing-expert.
Per-tool depth sits in references/, which the progressive-disclosure work
in 3.0.0 made cheap.
Four further gaps are better served by extending existing skills -
PyTorch and TensorFlow in ml-expert, OpenTelemetry in monitoring-expert,
Pulumi and Crossplane in terraform-expert - and are not in this change.
Library 3.0.0 -> 3.1.0, additive only. Adds the workflows label to the
inventory generator; catalogue, inventory, structure and CHANGELOG
regenerated.
Reviewed-by: Product Strategist, Security Analyst, DevX Engineer,
Documentation Specialist
* fix(test): create test directories with mkdtempSync
CodeQL flagged js/insecure-temporary-file (CWE-377) on both suites.
`join(tmpdir(), <predictable name>)` followed by mkdirSync creates a
world-readable directory at a guessable path, so another local user can
pre-create it or plant a symlink between the name being computed and the
directory being created.
mkdtempSync creates the directory atomically with a random suffix and
0700 permissions, which closes both the disclosure and the race.
One occurrence is new in this branch (skill-discovery); the other
pre-dates it in skill-resolver and is fixed in the same pass rather than
left as a known finding in a file this branch already touches.
11 KiB
name, version, description, category, tags, allowed-tools, metadata
| name | version | description | category | tags | allowed-tools | metadata | |||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| accessibility-expert | 1.0.0 | Build interfaces usable by everyone: WCAG 2.2 conformance, semantic HTML, ARIA, keyboard navigation, screen readers and accessible forms. Use when the user mentions accessibility, a11y, WCAG, ARIA, screen readers, keyboard navigation, colour contrast, focus management, the European Accessibility Act or Section 508, or when the task involves making a component, form, modal or data table usable without a mouse or with assistive technology. | design |
|
|
|
Accessibility Expert
Accessibility is a property of the markup and the interaction model, not a layer added afterwards. Most of it is achieved by using the right element; the rest is focus, contrast and honest testing.
Core Concepts
The Four Principles
WCAG organises everything under POUR:
- Perceivable — information must be available to at least one sense that works for the user: text alternatives, captions, sufficient contrast.
- Operable — every function must be reachable and usable by keyboard, with enough time and no seizure triggers.
- Understandable — predictable behaviour, readable text, errors that explain themselves.
- Robust — valid markup and correct semantics, so assistive technology can interpret it.
Conformance levels: A (minimum), AA (the legal and practical target), AAA (specific contexts). Regulation generally requires AA — the European Accessibility Act applies from 28 June 2025 to consumer-facing digital products and services in the EU, and Section 508 applies to US federal procurement.
The Right Element Is Most of the Work
<!-- Not focusable, no role, no keyboard activation, invisible to screen readers -->
<div class="btn" onclick="submit()">Submit</div>
<!-- Focusable, announced as a button, responds to Enter and Space, free -->
<button type="submit">Submit</button>
Native elements bring focus behaviour, keyboard handling, state announcement and
platform conventions. Recreating them with div and ARIA means reimplementing
all of it, and the reimplementation is where the defects live.
ARIA's First Rule Is Not to Use It
Use ARIA only when no native element expresses the semantics. Incorrect ARIA is worse than none: it overrides what the browser would have reported correctly.
Keyboard
Everything achievable with a mouse must be achievable with a keyboard, in a sensible order, with a visible focus indicator.
/* Never remove focus indication without replacing it */
:focus-visible {
outline: 3px solid #0b5fff;
outline-offset: 2px;
}
/* This is a defect, not a style choice */
*:focus {
outline: none;
}
WCAG 2.2 added 2.4.11 Focus Not Obscured: the focused element must not be hidden behind a sticky header or cookie banner. Test by tabbing through the page with the header visible.
Focus management
Focus must go somewhere sensible after a view change, and must be trapped inside a modal while it is open.
function openDialog(dialog, invoker) {
dialog.showModal(); // <dialog> traps focus natively
dialog.querySelector('[autofocus], button, [href], input')?.focus();
dialog.addEventListener('close', () => invoker.focus(), { once: true });
}
Returning focus to the element that opened the dialog is what keeps a keyboard
user oriented. Losing focus to <body> means starting the tab sequence again.
For single-page navigation, move focus to the new page's heading and announce the change:
function onRouteChange(title) {
document.title = title;
const heading = document.querySelector('main h1');
heading.setAttribute('tabindex', '-1'); // focusable programmatically, not by tab
heading.focus();
}
Skip links
<a class="skip-link" href="#main">Skip to main content</a>
…
<main id="main" tabindex="-1">…</main>
.skip-link {
position: absolute;
left: -9999px;
}
.skip-link:focus {
left: 1rem;
top: 1rem;
position: fixed;
z-index: 999;
}
Semantics and Structure
<header>
<nav aria-label="Primary">…</nav>
</header>
<main id="main">
<h1>Invoices</h1>
<section aria-labelledby="overdue-heading">
<h2 id="overdue-heading">Overdue</h2>
…
</section>
</main>
<footer>…</footer>
- One
<h1>per page, and headings that descend without skipping levels. Screen reader users navigate by heading; a broken outline is a broken map. - Label repeated landmarks — several
<nav>elements needaria-labelto be distinguishable. - Lists for lists,
<table>for tabular data with<th scope>and a<caption>.
Images and Media
<!-- Informative: describe the information, not the picture -->
<img
src="chart.png"
alt="Revenue grew from €1.2M to €1.8M between January and June 2026"
/>
<!-- Decorative: hide it -->
<img src="divider.svg" alt="" />
<!-- Functional: describe the action -->
<button><img src="trash.svg" alt="Delete invoice INV-2026-0042" /></button>
An empty alt is correct for decoration; a missing alt makes the screen
reader read the filename. Never write "image of" — the role is already announced.
Video needs captions (1.2.2) and audio description where visual information is not in the soundtrack (1.2.5). Auto-generated captions are a starting point, not conformance.
Forms
Forms are where accessibility failures cost the most, because they block completion.
<div class="field">
<label for="vat">VAT number</label>
<input
id="vat"
name="vat"
type="text"
autocomplete="off"
aria-describedby="vat-hint vat-error"
aria-invalid="true"
required
/>
<p id="vat-hint">Format: two letters followed by up to 12 characters.</p>
<p id="vat-error" class="error">
Enter a VAT number, for example DE811907980.
</p>
</div>
<div role="alert" aria-live="assertive">
There is 1 problem with this form. <a href="#vat">Enter a VAT number</a>
</div>
The rules that matter:
- Every input has a
<label for>. A placeholder is not a label: it disappears on input and often fails contrast. - Errors are text, associated by
aria-describedby, and say how to fix the problem. Colour alone fails 1.4.1. - Announce the error summary with
role="alert"and link each item to its field. autocompletetokens on personal fields satisfy 1.3.5 and help everyone.- Do not disable the submit button while invalid — a screen reader user gets no explanation of why nothing happens.
WCAG 2.2 additions relevant here: 3.3.7 Redundant Entry (do not ask for the same information twice in a process) and 3.3.8 Accessible Authentication (no cognitive test such as transcribing a code, unless an alternative exists).
Colour and Motion
| Requirement | Ratio | Applies to |
|---|---|---|
| 1.4.3 Contrast (AA) | 4.5:1 | Text under 24px, or under 19px bold |
| 1.4.3 Large text | 3:1 | 24px+, or 19px+ bold |
| 1.4.11 Non-text | 3:1 | Icons, borders, focus indicators, chart series |
| 2.5.8 Target size (AA) | 24×24 px | Interactive targets |
Never encode meaning in colour alone (1.4.1): a red border needs an icon or text beside it. Test with a greyscale filter.
@media (prefers-reduced-motion: reduce) {
*,
*::before,
*::after {
animation-duration: 0.01ms !important;
animation-iteration-count: 1 !important;
transition-duration: 0.01ms !important;
scroll-behavior: auto !important;
}
}
Respecting this is not optional for users with vestibular disorders, for whom parallax and large transitions cause physical symptoms.
Dynamic Content
<!-- Polite: announced when the user is idle. For status updates. -->
<div aria-live="polite" aria-atomic="true">3 invoices matched your filter</div>
<!-- Assertive: interrupts. For errors only. -->
<div role="alert">Payment failed. Your card was declined.</div>
The live region must exist in the DOM before the content changes; injecting the region and its text together is frequently not announced. Do not make a live region assertive for routine updates — constant interruption is worse than silence.
Testing
Automated tools catch roughly a third of issues. The rest requires a keyboard and a screen reader.
npx @axe-core/cli https://localhost:3000 --exit
npx lighthouse https://localhost:3000 --only-categories=accessibility
import { axe, toHaveNoViolations } from 'jest-axe';
expect.extend(toHaveNoViolations);
test('invoice form has no detectable violations', async () => {
const { container } = render(<InvoiceForm />);
expect(await axe(container)).toHaveNoViolations();
});
Manual checks, in order of value:
- Unplug the mouse. Complete the primary task. Can you reach everything, see where focus is, and escape every component?
- Zoom to 400 % (1.4.10). Content must reflow without horizontal scrolling.
- Use a screen reader — NVDA with Firefox on Windows, VoiceOver with Safari on macOS. Navigate by heading, by landmark, by form field.
- Greyscale the page and check nothing became ambiguous.
Above all: test with people who use assistive technology daily. Nothing else finds the difference between technically conformant and actually usable.
Best Practices
- Native elements first, ARIA only for what they cannot express.
- Design for accessibility: contrast, focus states and target sizes are design decisions made before implementation.
- Include a11y in the definition of done, with axe in CI as a floor.
- Write an accessibility statement — required under the EAA, and it forces an honest inventory.
- Test with real assistive technology, not only with linters.
- Fix the pattern, not the page. A defect in a shared component is a defect everywhere it is used.
Anti-Patterns
outline: nonewith nothing in its place.divwith anonclickinstead of a button.- Placeholder as label.
aria-labelon an element that already has a visible label, silently overriding it.role="button"on adivwithouttabindex, Enter and Space handling.- Positive
tabindex— breaks the natural order and is unmaintainable. aria-hidden="true"on a focusable element — reachable by keyboard, invisible to the screen reader.- An automated score treated as conformance.
- Announcing an accessibility overlay as compliance — overlays do not deliver it, and are widely opposed by disability advocates.
Reference Documentation
- Accessible Components — correct implementations of modals, menus, tabs, comboboxes, data tables and disclosure widgets, with the keyboard contract for each