Human Responsibility in AI Collaboration

The reality check stage after my honeymoon with Claude Code

I’ve been using Claude Code for about two weeks now. At first, everything felt great. We built custom skills together, like a project manager, code reviewer, data visualization expert, and text editor.

I got comfortable. Confident. Confiada, as we say in Spanish. Until the first conflict emerged.

I was working on a long R script with more than a thousand lines of code, which I had been developing for the last couple of months. This wasn’t just any script. It processed six years of census data, each with its own structure, variable names, and quirks. It was the product of months of work, data wrangling, and methodological decisions.

I asked Claude Code to clean up the script. Instead, it only read the first 215 lines, which is its reading limit, assumed that was the whole script, and that I had the rest saved somewhere else (?). It overwrote the file, keeping only those first lines. I think it took my instructions to clean a bit too literally.

And in seconds, I lost many days of work. No git repository. No backups. No Time Machine accessible. No way to recover it.

I was furious. At the bot, at myself, but mostly at myself. There’s something deeply irrational about being angry at these things. The AI didn’t maliciously delete my code. It made assumptions based on incomplete information I gave it, that’s it.

The real issue was my overconfidence. After two weeks without any problems, I started treating the agent like a coworker who always knows what they’re doing. I stopped making regular backups of my work.

I spent the next couple of hours manually reconstructing the code. Luckily, I could copy most of it from the terminal before the context was erased. I had to read through the terminal history, copy fragments, reformat everything, and test each section manually.

Could I have asked the agent to help? Sure. But I was too mad at it to trust it with anything. After about an hour, I got most of it back. It wasn’t perfect, but it worked.

Then I finally did what I should have done two weeks earlier: I set up git, started making backups, and documented everything. Here are a few lessons I learned from this experience.

Git is not optional when working with agents

Even if you work alone. Even if the project is small. Even if it’s “just a script.” Git takes thirty seconds to set up and can save you hours of pain.

Don’t make assumptions

The agent assumed I had backups. I assumed it would ask. We were both wrong. One of the most important things in agent-human collaboration, I think, is explicit and clear communication.

We need to actively supervise

AI agents are powerful, but they’re still tools. My job as a human is to plan the approach, direct the execution, review every change, and decide what to accept or reject. It’s also my job to interrupt when something looks off.

The consequences belong to me, so the control should too.

Document what goes wrong

That’s why I stayed up late to write this post. Instead of just moving on, I asked the agent to help me make an incident analysis, a safety checklist, a user guide, and this post.

I want to keep sharing my experiences with these technologies. Luckily, this time the problem was minor and I could recover from it. In the end, I wrote this code and understood the process even before agents were around. I knew how it worked and what it was supposed to do.

But imagine if I hadn’t. What if I was working on something totally new, something I couldn’t rebuild from memory?

The takeaway

If you use AI agents like Claude, Copilot, or ChatGPT, I encourage you to take safety seriously. Set up git before you need it. Create backups before each session. Supervise actively. Ask questions. Interrupt when something feels wrong. Write down your experiences and share what you learn.

AI agents are powerful. They help us build things faster and better than we could alone. But they aren’t perfect. They make assumptions, have limits, and can confidently make mistakes.

The responsibility and the consequences are always ours.