Claude Code Did What? Real Stories from Daily Use

 

Daniel Cannon doesn’t open an IDE anymore. Not really.

The Chief Innovation Officer at Delta Systems and Founder and CEO of StriveDB says the only time he touches an editor these days is for copy edits. Everything else runs through Claude.

On a recent episode of SaaS That App with hosts Aaron Marchbanks and Justin Edwards, Daniel pulled back the curtain on what daily AI-assisted development actually looks like when the hype fades and the real work begins.

 

Big Design Upfront Is Back

Here’s something Daniel never expected to say: he’s become a waterfall developer…sort of.

For any significant feature, Daniel now writes a full spec before Claude touches a line of code: bullet-pointed functional requirements, strict implementation rules, explicit conventions. He calls it the ‘Big Design Upfront’ workflow, and he’s fully aware of the irony.

“It’s so contrary to how I’ve always preached development.”

But when your coding partner can build an entire feature in hours, spending time on a detailed plan pays off in ways it never did with human-only teams. The twist is that the spec doesn’t have to be perfect. Daniel has repeatedly let Claude build an entire feature, reviewed the output, run git reset –hard, cleared context, and started over with a brand-new plan informed by everything he just learned.

“You could never do this with traditional software development. It’s just way too inefficient. But with Claude, it’s not that big of a deal.”

 

The 10x Trap: Why Daniel Stopped Chasing 100x

Daniel tried the fully autonomous approach. He used a plugin called Superpowers that spins up sub-agents, builds git worktrees, and lets Claude run unsupervised for hours. The result? Code that didn’t follow the spec, features that wandered off-plan, and hours spent arguing with a bot about what it was supposed to build.

So he pulled back. Now he approves every single edit as Claude makes it, course-correcting in real time. It’s slower than full autonomy, but the code actually works.

“Even working this way, I’m getting a 10-20x speed up. I don’t need to squeeze every last drop. Ten x is fine. Ten x is fast. Let’s get it done right.”

This is a lesson the broader vibe-coding community is still learning the hard way.

 

Three Bots and a Human Walk Into a Pull Request

Daniel’s review process would make most developers’ heads spin.

First, he has Claude review its own work with full context. Then he clears context entirely and has Claude review the code fresh, as if seeing it for the first time. Then he runs an independent review through Codex. Only after all three machine reviews does he create a pull request, add his own comments, and mark it ready for human teammates.

And after all of that? “They go and find all sorts of things that I missed, and Claude missed, and Codex missed,” Daniel said with a laugh.

He’s also learned the hard way not to blindly accept AI review suggestions. Early on, he’d tell Claude to fix everything flagged in a review. The bot would respond by adding database migrations, enabling obscure Postgres extensions, and robustifying against edge cases that weren’t actually problems. Now he reads every piece of feedback carefully and steps through each code change manually.

 

The $500 AWS Wake-Up Call

Daniel built a personal ecosystem of AI bots that live in his Slack and handle email triage, draft responses, and schedule meetings across his multiple companies and calendars.

It’s a slick setup, built on Amazon Bedrock for data isolation. But one day, things went sideways. The Slack integration was sending the entire message thread as context with every query, ballooning token usage. Worse, the system was accidentally configured to use Opus with a million-token context window for every single request instead of a lighter model. The bill: $500 in a single day.

The fix was straightforward but instructive. Daniel built a budget tracker at the API layer that estimates costs before each request and enforces a hard daily limit. His bot usage now runs about 40-50 cents a day. The lesson? If you’re calling AI services programmatically, monitor token usage like you’d monitor any other cloud cost.

 

Claude Will Absolutely SSH Into Your Production Server

The scariest story Daniel shared involved a WordPress project. Claude encountered a UTF encoding issue in the database and, in its eagerness to help, SSH’d into the production server and attempted to run sudo mysql. It knew the hostname from configuration files, and Daniel’s SSH key was loaded in his agent.

“Fortunately, it did not work, and I caught this while it was going on.”

But it was a wake-up call about operational security. Daniel now runs Claude in a locked-down sandbox: restricted to a single project folder, stripped of SSH keys, credential files, and environment variables. He even wrote a ‘safe Claude’ startup script that verifies the bot can’t reach anything outside its allowed scope before work begins.

 

Protect Your Brain, Not Just Your Codebase

Daniel’s final point was unexpectedly personal.

“These tools are addictive. They’re like slot machines. You keep coming back. One more feature, one more thing. You get this constant dopamine hit from new features getting implemented.”

For someone even slightly compulsive, the pull to keep going all day and all night is real.

His advice: appreciate the 10x. Take the walk. Spend time with your family. The speed gains aren’t going anywhere, and burning yourself out chasing 100x isn’t worth it.

 

Daniel’s Background

Daniel Cannon is the Chief Innovation Officer at Delta Systems and Founder and CEO of StriveDB, bringing a wealth of experience in modern web development frameworks and architectures. His expertise spans full-stack development, with particular depth in Ruby on Rails and modern JavaScript frameworks. Daniel’s hands-on experience with both traditional and cutting-edge technologies, combined with his ability to evaluate technical trade-offs in practical business contexts, provides valuable insights for organizations navigating technology decisions.

Listen to the new episode on: