Codex Caught What the Build Didn't: Claude Code + 4-Round Cross-Validation on jidonglab.com

Codex ran four times over this redesign. It found zero code bugs. It caught three copy errors — including one sentence that was factually false and passed every automated check: npm run build, TypeScript, ESLint. No tool flagged it.

That gap between “builds successfully” and “is actually correct” is the most interesting part of this project.

TL;DR: Full homepage redesign of jidonglab.com as a portfolio business card. Goal: a stranger with just the URL should know what Jidong builds within three seconds. 10 Claude Code sessions, 164 tool calls, 4 Codex cross-validation rounds. More copy errors surfaced than code bugs.

The Question That Prompted the Redesign

A friend visited jidonglab.com and replied with one question: “What do you do?”

That was the brief.

The site had daily AI news above the fold, projects buried below the scroll, and a hero section with copy vague enough to describe half the internet. Nothing was wrong with the code. The site was built as a content feed, not a portfolio — and it showed.

One constraint: don’t touch the tech stack. Astro 4 + React + Tailwind on Cloudflare Pages stays as-is. Only copy and section structure. One test: can a stranger identify what this person builds within three seconds of landing?

Claude mapped the current state with 12 Read calls across Hero.tsx, About.astro, Projects.tsx, index.astro, and home.css, then executed 11 Edit calls. The new addition was src/components/home/Capabilities.astro — four capability cards:

• AI automation (build logs, news pipelines, dental ad systems)
• Full-stack development (Next.js, Astro, Cloudflare Workers)
• Dental ad operations (Google/Meta campaigns, landing pages)
• Bilingual build log writing (Korean + English)

Each card uses data-ko/data-en HTML attributes for client-side language switching. The .do-grid CSS class handles the mobile one-column fallback.

The sessions kept hitting Hermes’s max_turns limit and cutting off mid-task. What should have been one session became ten. That’s why ten session boundaries cover changes that form one logical unit of work.

Codex Round 1: The Commit That Wasn’t Self-Contained

First Codex cross-validation came back with this:

“src/pages/index.astro imports/uses src/components/home/Capabilities.astro, but the file is untracked. Do not leave the tracked diff non-self-contained.”

index.astro was already importing Capabilities.astro. The import was in the commit. The actual component file wasn’t — staging had been interrupted by the session cutoff. The diff described a state that couldn’t run.

This is the simplest class of error Codex catches: structural completeness. A reviewer reading only the diff should be able to apply it and have the code work. An untracked import breaks that guarantee.

Fixed by staging the missing file. Also caught a copy problem in Projects.tsx during the review: the “currently running” section had development-stage projects mixed in with live ones. Separated them with explicit labels.

Codex Round 2: When the Build Passes But the Copy Lies

Round 2 was more interesting:

Capabilities.astro: "Every commit diff becomes a Korean/English build log." — This is not true.

It isn’t. Not every commit gets a build log. This was aspirational copy that crossed into factual territory.

Here’s what makes this error worth thinking about: it’s not a code bug. Not a type error. npm run build passes. TypeScript is happy. ESLint has no opinions. The sentence is just false, and no automated tool in a standard web development pipeline is designed to catch that.

Before:

KO: "커밋 diff를 한국어·영어 빌드 로그로 매일 쌓는다."
EN: "Every commit diff becomes a Korean/English build log."

After:

KO: "진행 중인 작업을 한국어·영어 빌드 로그로 꾸준히 남긴다."
EN: "Progress gets documented as bilingual build logs."

The fix is a vocabulary shift: absolute quantifiers (“every”, “daily”) become directional ones (“consistently”, “documented”). Meaning preserved. Statement now true.

If you’re using Claude Code to write copy and code together, semantic accuracy is a category of error to watch specifically. The model writes what sounds right; automated tools validate what compiles. The intersection doesn’t cover factual correctness.

Codex Round 3: A Sentence at War With Itself

Round 3 flagged this Korean sentence:

"혼자 같이 만든다." — literally “building alone together.”

The intent reads through: one person building, with AI as the working partner. But “alone together” is an oxymoron in Korean the same way it is in English.

Replaced with:

KO: "AI와 함께, 실제로 혼자 만든다."
EN: "Building alone, but with AI as co-pilot."

Same concept. No contradiction. The English version works standalone for anyone who hasn’t encountered the solo-plus-AI building model before.

Codex running total after three rounds: 0 code bugs. 2 factual copy errors. 1 logical contradiction.

”The Korean Tone Is Off” — A Bug Hidden in the Layout

Feedback arrived after the redesign went live:

“Looks good, but the language switching and Korean text tone feel off.”

Two separate problems compressed into one sentence.

The tone issue was fixable — some copy had drifted into a more formal register than the site’s voice. Standard copy editing.

The language switching issue was different. The toggle wasn’t working on the homepage at all.

Subpages using Base.astro load a JavaScript snippet that watches the language toggle button, reads and writes localStorage, and swaps data-ko/data-en attribute visibility. This script lives in Base.astro.

index.astro (the homepage) uses an independent layout that doesn’t inherit from Base.astro. The toggle script was never imported. Clicking the language button did nothing. Korean text stayed Korean regardless of user selection.

This is a common Astro footgun: components are portable across pages, but layout-level scripts aren’t. A component that depends on a layout-provided script will silently break on any page using a different layout. The component renders fine. The interactive behavior disappears with no error.

Investigation session: 13 Read calls + 3 Grep calls to trace the script loading chain. Diagnosis only — no fix that session. The actual changes landed in a separate session: Hero.tsx, Capabilities.astro, Projects.tsx, ShipLog.astro — 7 Read calls, 5 Edit calls to wire the toggle script directly into the homepage layout.

The sitemap 404 Nobody Had Noticed

Before the SEO/AEO work, an audit ran first. Found this immediately.

Base.astro had <link rel="sitemap" href="/sitemap-index.xml">. The old robots.txt pointed to the same path. Astro’s @astrojs/sitemap integration actually generates /sitemap.xml. Every subpage on jidonglab.com had a sitemap link pointing to a 404 — for an unknown duration.

JSON-LD was also missing entirely. Without structured data, Google indexes the page content but has no entity context. It can infer “software development”; it can’t infer “this is a specific person with specific skills and projects.”

Fixed in one session across five files:

src/components/Analytics.astro — New. Emits the GA4 script only when PUBLIC_GA_MEASUREMENT_ID is set. Local dev without the env var doesn’t break the build.

src/pages/index.astro — JSON-LD schemas: Person + WebSite + ProfilePage. Explicit og:type, og:locale, canonical URL in OG tags.

src/layouts/Base.astro — Sitemap href corrected: /sitemap-index.xml → /sitemap.xml. Analytics component wired in.

public/robots.txt — New file. 13 named AI crawler groups with explicit rules: GPTBot, ClaudeBot, PerplexityBot, GoogleOther, Google-Extended, ChatGPT-User, Applebot-Extended, Bytespider, CCBot, anthropic-ai, cohere-ai, Omgilibot, Diffbot.

public/llms.txt — New file. AEO entity context for language model consumption: who Jidong is, what he builds, canonical URLs, what the site contains.

Tool usage this session: Read ×14, Bash ×6, Edit ×6, Grep ×3, Write ×3.

Codex Round 4: robots.txt Rule Conflict

Round 4 checked the SEO work and flagged the robots.txt pattern:

User-agent: GPTBot
Allow: /
Disallow: /api/
Disallow: /admin

The issue: Allow: / can make the subsequent Disallow rules meaningless. The robots.txt specification doesn’t clearly define rule precedence when Allow and Disallow match the same path at different specificity levels, and real crawler implementations vary — some take the most permissive matching rule, others use declaration order, others use longest-path matching. With Allow: / present, there’s no guarantee /api/ and /admin are actually blocked.

Fix: for all 13 crawler blocks, added explicit Disallow: /api/ and Disallow: /admin entries. Allow: / continues to signal that public content is crawlable. API endpoint protection is a separate, explicit rule — not something inferred from the absence of other allow rules.

By the Numbers

10 valid sessions after filtering Hermes max_turns interruptions that fragmented one logical task across multiple restarts.

~164 tool calls total:

Tool	Count
Read	64
Bash	54
Edit	31
Write	4

Files: 9 modified, 4 new (Capabilities.astro, Analytics.astro, robots.txt, llms.txt)

Codex cross-validation rounds: 4

Codex findings:

Category	Count
Code bugs	0
Structural (untracked file)	1
Factual copy errors	2
Logical contradictions	1
Infrastructure config bugs	1

Cross-validation caught five issues across four rounds. None were code bugs. The most valuable catches were in categories where automated tooling has no opinions: semantic accuracy, logical consistency, and spec-ambiguous configuration.

The build passing doesn’t mean the copy is true. That’s the part worth keeping.

---

More projects and build logs at jidonglab.com