Ernie
AI Playbook

Start With Zero Risk

Afraid AI coding means worse code quality? Start with tools that review code, not generate it. Here's the adoption path that never risks your codebase.

2 of 6

You're Not Going to Break Anything

Let's name the fear: "AI coding" sounds like letting a machine write your production code. You've seen the demos — someone types a vague prompt, gets a React app, ships it. That's vibecoding. It looks fast. It produces fragile, unreviewed, architecturally incoherent code. And you're right to want nothing to do with it.

But here's the thing: AI coding tools aren't just code generators. The most valuable ones don't write code at all. They review it. And starting there means you improve your code quality from day one — with zero risk of degradation.

Step 1: Start With Code Review

The safest first step is an AI code reviewer. It reads your PRs, flags real issues, and never touches your source code. You stay in full control.

CodeRabbit

This is the tool I'd recommend first to anyone. It runs on every PR automatically and catches things human reviewers miss — not just style nits, but actual bugs, missed edge cases, and architectural violations.

  • What it does: Reviews every PR, leaves inline comments, learns your codebase conventions
  • What it doesn't do: Write code, modify files, or merge anything
  • Cost: Free for open source, $15/seat/mo for private repos
  • Setup: Connect your GitHub repo, add a .coderabbit.yaml with path-specific instructions, done

The ROI is obvious on any team where PR reviews are a bottleneck. Even solo, it's a second pair of eyes that never gets tired.

Codex (Review Mode)

OpenAI's Codex can also review code when used through ChatGPT. It's a decent complement to CodeRabbit — good for asking "explain this code" or "what could go wrong here" — but it's not purpose-built for PR review. CodeRabbit is significantly better at catching real issues in the context of a diff.

Use both if you have a ChatGPT subscription anyway. But if you're picking one, pick CodeRabbit.

What you'll notice after two weeks: PR review cycles get shorter. You catch things earlier. Your existing code quality goes up, not down. That's the point — you haven't generated a single line of AI code, and you're already getting value.

Step 2: Add Completions

Once you're comfortable with AI reviewing your code, add inline completions. This is still not "AI writing your code" — it suggests the next line, you accept or reject with a keystroke.

GitHub Copilot ($10/mo) is the best value here. Unlimited completions, always on, never in the way. You don't switch it on and off — it's ambient. Most developers accept 30–40% of suggestions and find the rest easy to dismiss.

This is low-threat, high-productivity. You're still writing every line — just faster.

Step 3: Go Agentic

By now you've been using AI-reviewed and AI-assisted code for weeks. You trust the feedback loop. You've seen that quality didn't suffer — it improved. Now you're ready for the big step.

Claude Code Pro ($20/mo) is the best agentic coding tool right now. It reads your entire codebase, executes commands, creates PRs, and understands your project's architecture when you give it a good CLAUDE.md file. Multi-file refactors, debugging sessions, new features — this is where the real productivity unlock happens.

The key: you're not going in blind. You already have CodeRabbit reviewing everything the agent produces. The safety net is already in place.

See the developer guide for my full daily workflow, CLAUDE.md patterns, and when to stop trusting the output.

Track Your Progress

Don't try to evaluate everything at once. Here's a concrete timeline:

Weeks 1–2: Code Review Only

  • Set up CodeRabbit on your most active repos
  • Check: Are PR reviews faster? Are you catching things you used to miss?
  • Signal it's working: You merge with more confidence, not less

Weeks 3–4: Add Completions

  • Enable Copilot in your editor
  • Check: Are you accepting >30% of suggestions? Does your typing flow feel faster?
  • Signal it's working: You stop noticing it's there — it just feels like a faster keyboard

Month 2: Go Agentic

  • Start Claude Code on real tasks — a refactor, a bug fix, a new feature
  • Check: Time a task with and without. Compare the diffs
  • Signal it's working: You spend more time reviewing and less time typing. Quality stays the same or improves

For Teams: Maturity Levels

If you're thinking bigger, the feedback loop article defines four maturity levels — from "vibes" (Level 0) to a self-tightening organism where agents, linters, CI, and observability reinforce each other (Level 3). Most teams should aim for Level 2: architecture encoded as lint rules, not just documented in a CLAUDE.md.

What to Buy

Ready to spend? The tools and budget guide breaks down every option by price tier — from $0/mo to $60+/seat for teams.

Up next

AI-Native Dev Workflows

The multi-agent setup I actually use. CLAUDE.md patterns, Cursor rules, CodeRabbit config, and when to stop trusting the output.

Continue reading