Website review — Double-Knuth (correctness + consistency)

A two-pass audit of a site's code/config/content, modelled on the portfolio-code-reviewer Double-Knuth. It sits on top of website-qa: the tests are the floor (must be green first); this is the human-grade review that catches what a test suite can't — cross-file drift and half-done work.

When to run

• End of a build, before launch (pipeline step 8).
• After adding or substantially editing a page — a new route touches PAGES, nav, the sitemap, internal links and canonical, so consistency can silently break.
• It is read-only: produce a ranked list, then fix BUGs before launch. Don't fix-and-forget — surface first (Rule 12: never hide a skipped or broken thing).
• For just "is the offer clear to a stranger" on a site that's already built — faster and narrower than this full audit — use website-positioning-check.

Review depth — the owner chooses, every time

A website is lower-stakes than application code: a subtle bug in a landing page's copy costs a bad reading experience, not a broken payment or a data leak. This skill (Double-Knuth, below) is free, runs entirely in the current session, and is the default, sufficient review for the large majority of website changes — new pages, copy edits, layout tweaks, small SEO adjustments, content updates. An external AI reviewer on top of it is an optional extra, never a requirement, and never more than one at a time.

Default — no setup, no cost, use this unless there's a specific reason not to:

• This skill, both passes — Pass 1 delegates to /code-review (correctness bugs), Pass 2 is the cross-file/consistency sweep below.
• /code-review alone — just Pass 1, for a small, self-contained diff where cross-file consistency clearly isn't at risk (a one-page copy fix, a single component tweak). Faster than the full two-pass audit when that's all the change touches.

Optional escalation — one external reviewer, only when it's genuinely worth it: A handful of website changes carry enough risk that a second, independent model is worth the extra step: a site-wide SEO repositioning (the seo-reposition skill's plan/diff gates), a payment or checkout flow, a form collecting sensitive personal data, or a major structural redesign where the author reviewing their own architecture risks sharing its blind spots. Even for these, use one external reviewer, not the independent-review skill's own default of two — that default is calibrated for production application code (Rails migrations, auth changes), not website content/structure work:

# from this repo's skills root, or the equivalent path in a site that has
# independent-review vendored in:
skills/independent-review/scripts/independent_review.sh <file> --plan --first-success   # or --diff

A --diff run first checks the diff against git, from the repository it came from: a diff git diff made from the merge base with the default branch passes as it is; a diff of only some commits names its start with --base <commit>, and one not made by git diff there passes --base none (independent-review's script header, BASE CHECK). --first-success stops at the first reviewer that answers (Codex CLI, falling back to ollama-cloud) instead of running both — independent-review honors this override for plans, it doesn't force the pair. Setup: if independent-review isn't already working in this repo, its own SKILL.md has a guided onboarding wizard for getting Codex or ollama-cloud running (free, no payment) — walk through that first rather than assuming it's ready. In a scaffolded site the skill isn't shipped at all: check that the independent-review folder actually exists before offering this escalation as ready-to-run. If it doesn't, either vendor it first (copy skills/independent-review/ from the Croftweaver suite — the same kit this skill was copied from — into the site's .claude/skills/; a plugin installed from Anthropic's directory leaves that folder out, so then take it from https://github.com/karero/croftweaver) or frame the offer honestly as "available after a one-time setup step".

Explain the choice, then ask — don't decide silently. Before reaching for anything beyond this skill's own two passes, say plainly what the options are and what's actually at stake: "I can review this with the built-in check — free, runs now — or also send it to an outside AI for a second opinion, which costs nothing either but takes a few extra minutes. For [this kind of change] I'd lean toward [recommendation] because [reason] — want that, or would you rather [the other option]?" The owner decides; this skill's job is to make an informed recommendation, never to pick unasked.

Pass 1 — correctness / bugs

• Run /code-review on the branch/diff (correctness bugs + reuse/simplification). /code-review is Claude Code-only — under another host, review the diff by hand against this pass's checklist, or vendor the generic double-knuth skill from the Croftweaver suite.
• npm run build is clean — no errors or warnings; TS strict passes.
• npm test green (a11y/seo/navigation/anchors/orphans/images/tone/positioning/placeholders/email/links/llms-coverage/middleware) — nothing skipped or loosened. (A site scaffolded before a spec existed: copy it in from the starter rather than reviewing without it. placeholders.spec.ts goes in with its UNFILLED_UNTIL_LAUNCH as shipped, on a launched site too: the first run names each entry to delete, and its check for lower-case slots, which runs on listed pages only, finds the one older starters carried on the privacy page. placeholders.spec.ts also fails on a legal page with no recorded answer in the README.md, unless that page is still listed and shows its draft notice: ask the owner as AGENTS.md §5 says.)
• astro preview the new/edited pages — no console errors; interactions work.
• Nothing half-done: no TODO/placeholder/lorem and no leftover [BRACKET] slots in shipped pages OR public/ assets. placeholders.spec.ts enforces this on the rendered pages and the manifest; read for what it cannot see: pages outside PAGES (the 404 page, noindex pages), other public/ files (llms.txt and its own slots included), PDFs, text inside images, the starter's example values, and an entry still sitting in its UNFILLED_UNTIL_LAUNCH at launch (it only warns there).
• Freshness date matches the edit: if the page's visible content changed and it carries a last-modified signal (a dateModified/"Updated" byline, an Article/BlogPosting JSON-LD date), that date is bumped to today. It silently drifts otherwise — stale in the visible byline, the <time datetime> attribute, and the schema's dateModified (an SEO/AI-answer-engine freshness signal) all at once, and no test catches it because nothing can know a date is "wrong" without knowing the edit happened.
• Nothing silently unshipped (two-stage sites): git log origin/production..origin/main empty, or the gap is a conscious decision — and the last npm run ship ended "✓ LIVE — verified". Any deploy you run or announce follows the site's PUBLISHING.md § "For AI assistants — deploy-time guardrails" (preview-vs-live announcements + the cached-404 rule).

Pass 2 — completeness + cross-file/skill consistency

• Routes: every src/pages/** route is in tests/_helpers.ts PAGES (and vice-versa; seo.spec.ts's sitemap drift alarm enforces it); nav + every internal link points to a real route; every route is reachable via at least one inbound internal link (no orphan pages — the privacy page included).
• URLs: the no-trailing-slash convention holds everywhere (/about, not /about/); canonical == og:url == SITE.url + path.
• Single source of truth: address, email, phone, nav come only from src/config.ts — no hardcoded copies drifting inside pages.
• Recurring section patterns carry consistent headings: if a component pattern (e.g. a grid grid-2/grid-3 card block) has its own <h2> directly above it everywhere else it appears on the site, a new or edited instance should too. A local heading step (h2 → h3) will not trip axe's heading-order rule, so this only surfaces by checking this page's heading structure against how the same pattern is used on other pages, not by trusting the a11y suite.
• SEO (website-seo-geo): title ≤60, description 120–160, og/twitter mirror, exactly one <h1>, JSON-LD valid, sitemap covers every page, llms.txt current.
• Images (website-design-system): WebP/AVIF, explicit width/height, alt; the LCP image eager (fetchpriority="high"), below-fold lazy; no oversized files.
• Email: every address rendered via <EmailLink>; email.spec.ts green.
• Headers/redirects: _headers CSP present; legacy URLs 301 in _redirects; the *.pages.dev middleware in place (previews noindexed); on a launched site, the project's own pages.dev address (as Cloudflare shows it; not always the project's name) answers 301 to the live domain — if it answers 200, see the causes listed in new-website/references/CLOUDFLARE_FIRST_DEPLOY.md, "After go-live", step 3 (variable not under Production or rejected, no redeploy since, or a middleware that predates the redirect).
• Secrets: no .env/token committed; .gitignore covers them.
• Docs match code: README test list + "how to add a page / run tests / deploy" current; CONTENT_GUIDE.md + BRAND.md filled (+ STORY.md, consistent with POSITIONING.md, if the site opted into website-story); decision-interview answers recorded; the handoff skill set present in .claude/skills/.
• Language/i18n (non-English or multi-locale sites): every page's CONTENT language matches its <html lang> (the tone gate's density check covers German; eyeball other languages); og:locale present and matching on every page; chrome (skip link, footer) in the site's language; German-market sites carry the filled Impressum and the German privacy page at the right URL (/datenschutz single-locale, /de/privacy multilingual); multi-locale sites: hreflang sets complete + reciprocal (tests/i18n.spec.ts green, or the twin-page seo.spec test for the light path), every /locale route in PAGES, the language switcher on every multi-locale page, llms.txt entries in each page's language.

Output / done means

A ranked BUG / RISK / NIT list, each with file:line and a concrete fix (same shape as the portfolio Double-Knuth). Fix BUGs before launch; RISK/NIT by judgment. Re-run website-qa after fixes. Done = both passes clean, or the remaining items explicitly listed — never hidden.