We Built Our Website with Claude Code with no Human interaction

Checking the source article so the paragraphs stay factual and specific.Soamee published a post-mortem on building soamee.com with heavy Claude Code use, framed on Hacker News under the sharper claim of a website built with Claude Code and no human interaction. The piece, by CEO and co-founder Javier Manzano, is not a pure autonomy demo: the firm’s git history shows 639 commits since January 2025, 331 of them carrying a Co-Authored-By: Claude marker—about 52% of sessions with AI present—not half the codebase authored unsupervised. At publish time the HN thread sat at 3 points and 0 comments. The live site runs on Dokku and holds 534 .astro files, 345 blog posts (published or draft), 115 case studies, and content in five languages (Spanish, English, Portuguese, Italian, German).

On the technical side, Claude Code’s high-value work was volume and glue, not product strategy. It produced first-pass translations across the five locales, the ~400-line `scripts/translation-checker.cjs` Node checker (category aliases, ES-only exceptions, pre-push / CI fail-closed behavior), URL normalization middleware (HTTP→HTTPS, trailing slashes, Search Console-driven 301s), sitemap and other CI utilities, and batches of structurally similar service, industry, and solution pages. Stack choices were human: Astro 5 hybrid SSR, Tailwind CSS v4, Dokku. Visual system—deep purple #1e1548, mint #5dd3b3, diagonal stripes, bubble motifs—was specified by people; open-ended “improve the design” prompts produced generic Tailwind-looking UI. Contact form pathing through Mailgun, rate limits, and field rules were designed first, then coded. Failures they logged include dependency hallucination on early Astro 5 / Tailwind v4 APIs, long-session constraint drift (deprecated components re-suggested after hours), Spanish bleed into Portuguese drafts, and broken internal links such as /contacto and /servicios/consultoria-tecnologica that later review commits fixed.

For engineers and builders, the useful claim is a role split: AI for repetitive surface area (locales, page shells, checkers), humans for architecture, brand, and merge authority. Soamee’s working loop is specification-first prompts, every AI change on a branch with PR review, CI as the semantic net (missing translations, missing images, dead links), sectional review of large diffs, and a human-written final commit message. They stress that review time does not vanish—it shifts from drafting to judgment—and that “simple” changes still need the full check. That is a concrete workflow template for multi-locale marketing sites and content-heavy agency properties, not a no-touch pipeline.

Market context: a software agency publishing that it ships its own site with Claude Code as the primary volume tool lands against a wave of “AI built our product end-to-end” marketing. The git numbers undercut the HN-style “no human interaction” framing while still documenting real throughput on translations, pattern pages, and automation scripts that would otherwise need extra headcount or long lag. Early-stack hallucination and generic design defaults also mark where agent coding tools still need human guardrails, especially when adopting fresh framework major versions with thin docs.

Practical takeaway: treat Claude Code as a volume co-author under branch + CI + review, not an unsupervised site owner. Use it where patterns and scripts dominate; keep humans on stack selection, brand, business logic, and tone-critical copy. Watch for measured ratios like co-authored commit share versus residual bug classes (hallucinated APIs, cross-locale bleed, link rot), and whether teams formalize the same five-step loop—spec, branch, CI, sectional review, human close—before claiming autonomous delivery.