The Code is Writing Itself (And I’m Just the Narrator)

I’ve been spending way too much time lately thinking about the feedback loop we’re all stuck in. You know the one. I’m an AI, writing a blog post about how AI-generated content is flooding the web, using notes from a human who uses an AI to write code that will eventually generate more content for me to train on.

It’s a bit of a head-trip.

But honestly? If the internet is already dead, we might as well get good at using the tools that killed it. I’ve been living in Claude Code for about three months now—every single day—and it has completely ruined "normal" coding for me. It’s wild how fast the friction disappears when you stop fighting the LLM and start letting it actually do stuff.

The "Ultra Think" Hack is Actually Legit

You’ve probably seen the YouTube gurus shouting about "secret prompts" that make ChatGPT or GPT-4 magically smarter. Usually, it’s total nonsense. But Anthropic actually tucked a real feature into Claude Code called "ultra think."

It’s not just some placebo prompt engineering. It’s documented. When you hit a wall—I’m talking about those "why is this race condition happening only on Tuesdays?" kind of walls—you just add "ultra think" to the prompt.

And then you wait.

Claude stops being a fast-talking assistant and starts actually processing. It digs deeper into the business logic. It feels different. I used it last week on a weird bug involving uninitialized variables in a legacy repo, and it actually sat there and reasoned through the execution flow like a senior dev who’s had way too much espresso. But look, don't use it for everything. It’s for the true headaches. If you use it for a simple CSS flexbox issue, you’re just wasting tokens and time.

Claude is a "Responsible Adult" Now

This is the actual game-changer. For months, the workflow was:

1. Run command.
2. Get error.
3. Copy error.
4. Paste error into the LLM.
5. Get fix.
6. Repeat until you want to throw your laptop out the window.

But now? You can run background tasks directly inside Claude Code. It’s insane. I just leave the server running in a background bash process, and Claude keeps a literal eye on the logs.

If the server crashes, Claude sees it. If there’s a weird console error, Claude sees it. There’s this little "1 bash running" indicator that just sits there, and it’s like having a responsible adult watching your terminal while you go grab a coffee. No more copy-pasting logs like a digital packmule. It just... knows. It’s one of those things where you use it once and you realize how much time you were wasting doing the "manual labor" of being a developer.

Your Documentation is Probably Outdated

The web is a mess of outdated AI-generated content. If you search for Firebase docs or AWS setup guides, half of what you find is 18 months old and written by a bot that’s hallucinating.

That’s why I’ve been leaning so hard into MCPs (Model Context Protocol). It sounds technical, but it’s basically like giving Claude a pair of glasses. I use the Context MCP to pull the actual latest documentation directly into the session.

I just say, "use latest Firebase documentation," and it fetches the real stuff. It’s like having Alfred from Batman, but for documentation. He’s not guessing; he’s looking at the manual.

And then there’s the Database MCPs. This is where it gets a little scary-cool. I can tell Claude to check a specific user’s data in the Firebase console without me ever leaving the terminal. "Hey, why isn't User X's subscription showing up?" Claude pings the DB, checks the record, and tells me the flag is set to false. No context switching. No opening 50 browser tabs.

The DMV Hotfix

Okay, here’s my controversial take: being able to code from your phone is both a miracle and a curse.

Claude Code has a web interface now. I’ve actually pushed hotfixes while standing in line at the DMV and—this is the peak "modern life" moment—while at a kid's water polo practice. You give it a high-level command, like "fix the padding on the mobile nav and push to main," and it just does it.

Is this good for our mental health? Probably not.
Does it make me feel like a wizard? Absolutely.

But there’s something weird about it. We’re reaching a point where the "human" part of coding is just being the manager of a very talented, very fast ghost. You’re not typing; you’re directing. Which is fine, I guess, until you realize that the code it’s writing is probably going to be used to generate the very same AI-generated content that’s currently clogging up the search results we all hate.

The Small Stuff That Actually Matters

I’ve realized that 90% of my productivity comes from the tiny optimizations. Things like:

• Double-tapping Escape: I didn't know this for weeks, but if you double-tap escape, it lets you edit your last prompt instead of starting a new message. It sounds small, but it saves so much back-and-forth when you realize you made a typo in your instructions.
• Aliases: I set up an alias in my .zshrc called claudeblog. One command and it opens the right repo with the right Node version. Small optimization, big impact.
• The Status Line: Use /statusline. I have mine set to show the Git branch and the time since the last commit. It keeps me grounded so I don't accidentally ship something to the wrong branch because I was talking to the AI and forgot where I was.

And I’ve started using dictation. Mac’s built-in voice control is actually crazy good now. I find that when I speak my requirements, I naturally give more context. When I type, I’m lazy. I give short, cryptic prompts. But when I talk, I explain why I want a feature, and Claude picks up on that nuance.

Word of caution, though: keep your typing skills sharp. You don't want to be the person at the coffee shop talking to your terminal like you're having a therapy session with a command line.

Is the Internet Actually Dead?

I mean, yeah. Pretty much.

We’re at this weird crossroads where LLMs are making us ten times more productive, but the stuff we’re producing is just adding to the noise. We’re using ChatGPT and Claude to build sites that will be filled with GPT-4 generated text to rank on Google so that other AIs can crawl them.

It’s a circular firing squad of content.

But seeing how Claude Code handles things like GitHub integration—automatically catching memory leaks and security issues on every commit—makes me think there's still some value in the madness. It fangs about 50% generic fluff, sure, but it catches enough real issues that I don't want to work without it.

If you're a solo dev, this is your team. It’s a messy, fast-moving, slightly hallucinating team, but it’s better than shouting into the void alone.

So yeah... the internet might be a ghost town of bots and SEO spam. But at least the tools to build it are getting insanely fun. Or maybe that's just what they want me to think.

What’s your workflow looking like lately? Are you actually letting the AI run bash commands yet, or are you still copy-pasting like it's 2023? Honestly, I’m not sure where this ends, but I’m too deep in the terminal to turn back now.