Claude Code Changed How I Ship Software

There's a debate right now about "vibe coders" — people who let AI write all their code without understanding what's happening. Traditionalists say real developers will outlast the prompt jockeys. AI maximalists say coding knowledge is about to be irrelevant.

Both sides are wrong. I have the numbers to prove it.

What I Actually Do

I run multiple senior-level dev workloads. Not by working 16-hour days — by using Claude Code to multiply my output in ways I couldn't have imagined a year ago.

The workflow: Claude Code instances running in tmux panes, each working on a separate GitHub issue, using git worktrees so nothing conflicts. I review, steer, and merge.

Last month I shipped an entire admin dashboard — 30 pull requests across 9 worker batches in 2.5 days. Real-time log viewers, cron job management, SSE live updates, a memory browser, disk usage dashboards, full PWA layer.

30 PRs. 2.5 days. Me and a bunch of AI agents.

The Workflow

Here's what the "vibe coding" critics miss: I'm not copy-pasting output and hoping for the best.

Triage and plan. I review the GitHub board, group related issues, decide which can run in parallel without stepping on each other. This is pure engineering judgment. No AI involved.

Spin up agents. Each Claude Code instance gets a specific issue, a project context file (a CLAUDE.md with architecture decisions, conventions, and hard boundaries), and clear constraints. Separate git worktrees, zero merge hell.

Review ruthlessly. This is where the engineering happens. In the dashboard batch alone I caught and fixed 15+ security issues — shell injection vulnerabilities, missing path traversal guards, input validation gaps. More on this below.

Merge and iterate. Clean PRs go in. Broken ones get bounced back with specific feedback. Some issues took 3 attempts before they were right.

AI Agents Are Confidently Insecure

This is the part that should make every "vibe coder" nervous.

In one batch, three separate PRs had shell injection vulnerabilities. The code worked. The tests passed. It would've been a disaster in production.

// What Claude Code wrote (looks fine, right?)
const result = execSync(`git log --oneline ${branch}`);
 
// What I replaced it with
const result = execFileSync('git', ['log', '--oneline', branch]);

If I didn't have 10+ years of knowing what execSync with string interpolation means, those vulnerabilities ship. That's the real argument for engineering experience in the AI age: not that you need to write every line, but that you need to catch the lines that could destroy you.

The Honest Numbers

I'm going to be specific because most AI productivity content is either breathless hype or hand-wavy dismissal.

Agent success rate: ~85%. About 1 in 7 attempts fails outright — wrong approach, scope creep, or hallucinated APIs that don't exist.

Review overhead: 5-15 minutes per PR. That's 2.5-7.5 hours of review for 30 PRs. Not free.

Real throughput: 4-5x. Not 100x. Not 10x. Four to five times more productive than working solo. That's still transformative, but let's not pretend it's magic.

Cost: $50-80 for a heavy batch week. Less than a junior dev's hourly rate for one afternoon.

The highest-ROI investment isn't a better model — it's a better CLAUDE.md. A well-written project context file (architecture, conventions, boundaries) changes the quality of everything an agent produces. I update mine multiple times a week.

What Actually Changed

Before Claude Code, I was the bottleneck on everything. Multiple workloads meant constant context-switching. Side projects moved at a crawl because there were never enough hours.

Now the gap between "this should exist" and "this exists" is radically smaller. This blog? Built by Claude Code in a single session — MDX rendering, RSS feeds, OG image generation, syntax highlighting — while I focused on other work.

I still make every architecture decision. I still review every line that goes to production. I still catch the security issues AI introduces. But translating decisions into code is no longer the bottleneck. Thinking is the bottleneck, which is how it should've always been.

What the Debate Gets Wrong

The "vibe coding" discourse frames this as binary: either you're a real developer or you're an AI parrot.

I am a real developer. I also let AI write most of my first drafts. These aren't contradictions — they're the same workflow writers, architects, and designers have used forever. The skilled professional sets direction, reviews output, catches what the tool misses.

The developers who'll struggle aren't the ones using AI or avoiding it. They're the ones who can't tell the difference between code that works and code that's correct.

If you're going to try this: start with 2 agents, not 4. Use git worktrees. Write a CLAUDE.md before you write a single prompt. Never skip security review — automated tests won't catch what AI introduces. And track your actual numbers, because "AI makes me faster" isn't useful. "4x faster with 15% failure rate" is a real decision-making input.

The future isn't developers vs. AI. It's developers with AI, being honest about what that actually looks like.