Rejected Apple Clone to Codex Approval: Redesigning a Dev Portfolio in 27 Sessions (350+ Tool Calls)

The previous design got rejected before a single line of implementation code was written.

“Apple-style white space, oversized type, card-based layout” — that was the verdict. jidonglab isn’t a product landing page. It’s the operational log of a solo builder. A generic AI template feel was exactly what it couldn’t be.

TL;DR Rewrote 10 home components from scratch, ran Codex cross-validation twice, and spent more rounds on copy than on code. The two recurring feedback notes: “remove translated-sounding phrases” and “stop exposing internal toolchain details in public copy.”

The Mock That Got Rejected

The previous session produced a Swiss-studio cream/acid aesthetic: 104px serif headline, Instrument Serif italic, Space Grotesk display. It looked like an agency landing page. The user rejected it explicitly and listed three requirements instead:

“Developer lab + personal publishing + operational log + project index” feel
No generic AI templates, excessive 3D, or hollow hero copy
Information density beats type spectacle

First step: understand the codebase. Read through src/components/home/ for existing components, src/data/home.ts for data structure, src/styles/home.css for styles. The existing code was far enough from the new direction that a full rewrite was effectively necessary.

Design system declaration before touching any component. JetBrains Mono for section kickers, status labels, dates, and project metadata. IBM Plex Sans KR for body text. Green #00c471 as accent, but never decorative. Section separators via 1px border + monospace kicker.

57 Tool Calls, 10 Components Rewritten

Started with home.css — wiped the existing styles and rebuilt from the new design system. Then rewrote components one by one:

Hero.tsx — short identity statement + current active status
NowStrip.astro — three live status indicators (last commit, build status, active project)
Projects.tsx — featured (1) + active grid (3) + indexed archive (rest)
ShipLog.astro — last 3 build logs
Writing.astro — recent posts list
About.astro, Footer.astro, Topbar.astro, Lab.tsx — created new

Projects.tsx took the most work. Featured project renders large with a visual. Active projects sit in a 2×2 grid. The rest appear as a monospace index. All data is centrally managed in src/data/home.ts.

The session hit max turns mid-way. After reconnecting, ran git diff to assess state, then continued and finished the remaining components through build validation:

npm run build
# ✓ Built in 8.23s

First Codex Pass: 4 Issues Found, 1 Blocking

After implementation, ran Codex CLI for cross-validation. Result: request-changes.

One blocking issue. In Projects.tsx, mobile CSS hid elements by targeting .desc and .stack classes — but the header <span> elements didn’t carry those classes. Below 960px, the header column count misaligned with the row grid. Fix: add the same classes to header spans.

Three non-blocking issues, addressed at the same time:

Added prefers-reduced-motion media query to home.css (pulse animation)
Removed empty data-ko="." span in Hero.tsx
Restructured <strong> inside <p> in ShipLog.astro — nesting conflicted with the language-switching script

Second Codex pass: approve. No blocking issues, build passed, git diff --check clean.

Five Rounds to Kill the AI-Generated Voice

After build validation, user feedback: “Design is solid. But the copy reads like a translation and has information I don’t want public.”

Looking at actual sentences confirmed it. Words like “operations index,” “AI builder,” and “experiments” showed up repeatedly. Abstract, slightly inflated phrasing — the kind that’s characteristic of generated copy. Five rounds to fix it.

Round 1 — Direction change across all copy. Shorter. More concrete. Delete if not necessary.

Round 2 — Add project images. Found 8 screenshots in public/images/projects/. Added a PROJECT_IMAGE mapping to src/data/home.ts and rendered via <img> tags in Projects.tsx. Projects without images fall back to a CSS-based color swatch. Codex caught a bug during review: PROJECT_IMAGE was added to home.ts but never imported in Projects.tsx, so images weren’t rendering at all.

Round 3 — Remove the Lab section. There was a section marketing the internal Claude Code skills, hooks, and pipelines as features. User request: “Take out the daily toolchain stuff.” Removed <Lab> import and component from src/pages/index.astro.

Round 4 — Jargon sweep across all public pages. Cleaned about.astro, ai-news/index.astro, and public/llms.txt. Removed phrases like “Claude Code skills,” “multi-agent orchestration,” and “daily 08:00 KST cron” — internal implementation details that don’t belong in public copy. Rewrote llms.txt entirely as an understated portfolio/build-log archive description.

Round 5 — Metadata cleanup. Removed “AI agent,” “MCP server,” and “Built with Claude Code” from schema.org, footer, and default description in src/layouts/Base.astro.

.gitignore Cleanup (Codex Flagged It)

Separate Codex note: several untracked local directories weren’t in .gitignore, which meant they could accidentally end up in production commits.

Added .claude/agent-memory/, .claude/worktrees/, and .wrangler/. Deduplicated a redundant .vercel/ entry. Kept .vercelignore as a deployment safety net and included it in the commit.

What This Build Taught Me

Copy takes more rounds than code. The component rewrite finished in one session. Copy editing ran five rounds. “Remove translated-sounding phrases” doesn’t work as a single instruction — you have to look at actual sentences and remove them one by one. Prompts with specific examples (“change this phrase to that”) resolved faster than abstract direction.

Don’t expose internal implementation in public copy. “Auto-generated daily by Claude Code” is technically true. But making it a marketing point is odd. Public pages describe what users experience, not what CLI tools built it.

Cross-validation finds bugs that familiarity hides. The missing PROJECT_IMAGE import and the mobile header misalignment were both caught by Codex, not during the main implementation session. Running a separate pass against a diff catches the things you stop seeing after hours in the same codebase.

Tool call breakdown: Session 8 (core implementation) — Bash 22×, Read 20×, Write 10×, Edit 5×. Copy editing sessions were Read- and Edit-heavy. Total across 27 sessions: 350+ tool calls.

The design took a day. Making the copy sound like a person took the rest of the time.

More projects and build logs at jidonglab.com