qa-review-hidden-text
Reviews text hidden in DOM attributes — placeholders, alt text, title tooltips, aria-labels, input values, and select options — for spelling, grammar, word choice, and untranslated keys. Catches what
SKILL.mdskills/qa-review-hidden-text/SKILL.md
GitHubWhat it checks
A lot of user-facing text never appears in document.body.innerText — it lives in DOM attributes. The companion skill qa-review-content skips all of this. This skill closes the gap.
Specifically reviews text in:
| Bucket | Source | Why it matters |
|---|---|---|
placeholders |
<input placeholder="..."> |
Most common hidden-typo location |
altText |
<img alt="..."> |
Accessibility audit fails; screen-reader announces the typo |
titles |
<element title="..."> |
Hover tooltip text |
ariaLabels |
<element aria-label="..."> |
Screen-reader-only labels — sighted users never see the typo |
valueAttrs |
<input type="submit" value="..."> / <button> value attribute |
Default button labels |
optionLabels |
<select> <option> |
Dropdown labels — only visible when dropdown is open |
Same issue categories as qa-review-content:
| issueType | severity | What it catches |
|---|---|---|
spellingError |
medium | Misspelled words in any attribute |
grammarError |
medium | Grammar mistakes in tooltip / placeholder copy |
wordChoice |
medium | your/you're, their/there, its/it's |
placeholderText |
high | "TODO", "Enter text", "placeholder" left in production |
untranslatedKey |
high | common.button.save, {{var}} in alt/title/aria-label |
htmlEntityLiteral |
medium | &, < rendered in an alt tag |
encodingMojibake |
medium | Broken UTF-8 in attributes |
attributeEmpty |
medium | Empty alt/aria-label/title where text is expected |
attributeRedundant |
low | Alt text duplicating visible caption (a11y noise) |
Orchestrator flow
Interactive Sonnet skill. Same pattern as qa-review-content.
Run
probe.extractAttributeText— collects all attribute text into buckets.Self-skip the cell if every bucket is empty (no attributes worth reviewing).
Pass the buckets + cell context to a Sonnet sub-agent:
Agent( subagent_type: "general-purpose", model = "sonnet", prompt = `You are reviewing text hidden in DOM attributes for content quality issues. Cell context: route: {cell.route} viewport: {cell.viewport} browser: {cell.browser} Extracted attribute text (each item has source / text / locator): PLACEHOLDERS: {placeholdersJson} ALT TEXT: {altTextJson} TITLE ATTRIBUTES: {titlesJson} ARIA-LABELS: {ariaLabelsJson} INPUT VALUE ATTRS: {valueAttrsJson} SELECT OPTIONS: {optionLabelsJson} Find any issues from these categories (only emit if you're confident): - spellingError - grammarError - wordChoice - placeholderText (TODO / placeholder / Lorem ipsum / "Enter text here" generic) - untranslatedKey (common.foo, {{var}}, [error.x], __MSG_*) - htmlEntityLiteral (&, <, ) - encodingMojibake (é, ’ — broken UTF-8) - attributeEmpty (empty alt on a non-decorative image, empty aria-label on a button) - attributeRedundant (alt text duplicates visible caption word-for-word) Rules: - Do NOT flag product names, brand names, or technical acronyms as misspellings. - Do NOT flag intentional casing like iPhone, macOS, GitHub. - For "attributeEmpty", only flag images that look content-bearing (>100px wide), not pure decorations. - Limit to 15 findings per cell. - Output ONLY a valid JSON array — no prose, no markdown. Return JSON: [ { "issueType": "<one of the categories above>", "severity": "low" | "medium" | "high", "source": "placeholder" | "altText" | "title" | "ariaLabel" | "valueAttr" | "option", "snippet": "<the offending text, max 80 chars>", "suggestion": "<corrected text, max 80 chars, or null>", "locator": "<selector or descriptor of the element, max 80 chars>" } ] If nothing is wrong, return [].` )Convert each finding into the standard issue record:
{ skill: "qa-review-hidden-text", issueType: <finding.issueType>, severity: <finding.severity>, selector: <finding.locator>, description: `[${finding.source}] "${finding.snippet}"` + (finding.suggestion ? ` (suggestion: "${finding.suggestion}")` : ''), bbox: null }Append to issues. No screenshot annotation — attribute findings don't have bboxes.
Probe (browser_evaluate)
// probe.extractAttributeText
() => {
const selFor = el => {
const id = el.id ? `#${el.id}` : '';
const name = el.name ? `[name="${el.name}"]` : '';
return (el.tagName.toLowerCase() + id + name).slice(0, 100);
};
const placeholders = [];
const altText = [];
const titles = [];
const ariaLabels = [];
const valueAttrs = [];
const optionLabels = [];
// Placeholders
for (const el of document.querySelectorAll('input[placeholder], textarea[placeholder]')) {
if (placeholders.length >= 25) break;
const t = el.getAttribute('placeholder');
if (t && t.trim()) placeholders.push({ text: t.slice(0, 200), locator: selFor(el), inputType: el.type || 'text' });
}
// Alt text — include empty alts on visible images (potential a11y issues)
for (const img of document.querySelectorAll('img')) {
if (altText.length >= 25) break;
const r = img.getBoundingClientRect();
if (r.width === 0 && r.height === 0) continue;
const alt = img.getAttribute('alt');
const src = (img.src || img.currentSrc || '').split('/').pop().slice(0, 30);
altText.push({
text: alt === null ? null : alt.slice(0, 200),
locator: selFor(img) + `[src=…${src}]`,
decorative: img.getAttribute('aria-hidden') === 'true' || img.getAttribute('role') === 'presentation',
widthPx: Math.round(r.width)
});
}
// Title attributes (skip on html/body/script/style)
const SKIP_TITLE_TAGS = new Set(['HTML', 'BODY', 'SCRIPT', 'STYLE', 'LINK', 'META']);
for (const el of document.querySelectorAll('[title]')) {
if (titles.length >= 20) break;
if (SKIP_TITLE_TAGS.has(el.tagName)) continue;
const t = el.getAttribute('title');
if (t && t.trim()) titles.push({ text: t.slice(0, 200), locator: selFor(el) });
}
// Aria-labels
for (const el of document.querySelectorAll('[aria-label]')) {
if (ariaLabels.length >= 25) break;
const t = el.getAttribute('aria-label');
if (t && t.trim()) ariaLabels.push({ text: t.slice(0, 200), locator: selFor(el) });
}
// Input/button value attributes
for (const el of document.querySelectorAll('input[type="submit"], input[type="button"], input[type="reset"]')) {
if (valueAttrs.length >= 15) break;
if (el.value && el.value.trim()) valueAttrs.push({ text: el.value.slice(0, 200), locator: selFor(el) });
}
// Select <option> labels
for (const opt of document.querySelectorAll('select option')) {
if (optionLabels.length >= 30) break;
const t = (opt.innerText || opt.value || '').trim();
if (!t) continue;
const select = opt.closest('select');
optionLabels.push({
text: t.slice(0, 200),
locator: select ? selFor(select) + ` option[value="${(opt.value || '').slice(0,30)}"]` : 'option'
});
}
const totalCount = placeholders.length + altText.length + titles.length +
ariaLabels.length + valueAttrs.length + optionLabels.length;
return { totalCount, placeholders, altText, titles, ariaLabels, valueAttrs, optionLabels };
}
Cell-skipping rules
- Self-skip if
totalCount === 0— no attribute text on this page. - Honors the same
[grammar].critical_routesand[grammar].max_cellsconfig asqa-review-content— both skills share the scope.
Notes
- Same Sonnet pricing as
qa-review-content. Per-cell input is smaller (attribute text is shorter than body text), so this skill is cheaper per cell. - No DOM mutation, no
data-argus-*attributes, no cleanup needed. - Findings have a
sourcetag (placeholder/altText/title/ariaLabel/valueAttr/option) prepended to the description so the bug report makes it obvious which attribute is broken.