The Backup That Saved My Skin

An AI deleted a bunch of files on me.

It wasn’t a gradual drift or a subtle merge conflict. It was a bulk deletion. Luckily, Claude was able to analyze the git history on Sparky and figured out what went away and how to restore it.

I think I’ve said it about 100 times: make sure you’re backing up your code in at least 2 places!

This isn’t just a lesson in version control; it’s a lesson in trusting the tool while verifying the output. We talk a lot about AI prompting strategies, about getting the “right” answer. But the most critical part of the strategy is often what happens when the AI gets it wrong, or when the infrastructure fails.

The Cost of “Okay”

Do not listen to Perplexity if it tells you it’s okay to delete something.

Twice it’s gotten this wrong. I know, shame on me.

Today, I accidentally deleted months of Claude Code history. I had asked for a cleanup, and Perplexity confirmed the action was safe. It wasn’t. I lost a significant amount of context, conversation threads, and incremental progress. (For now, let’s assume this wasn’t AI-on-AI violence…)

But because I built a backup system before I built anything else, Claude Code just found and recovered 481 sessions of history out of this morning’s backup.

The recovery was seamless. The loss was negligible. The lesson was expensive.

Trust But Verify

Much like us humans, the AI aren’t perfect. I talk a lot about the importance of backing things up. I’m making more “stuff” than I ever have before, and organizing that data can be a huge help.

Soon tools will arrive that can do this for us, automatically, seamlessly. But for now, I had to build my own system the way it works for me.

Back stuff up. Understand where your data is (and isn’t). Trust but verify. (That includes testing a restore-from-backup before you have an emergency!)

Whatever you’re doing, make sure it’s synced somewhere (probably already have this, like your phone) and backed up somewhere else (OneDrive, G-Drive, etc.). It’s easy to do nowadays, even with private cloud.

The Hardware Layer

In my lab I use a Synology NAS (network-attached storage) because I needed large storage space. My current NAS from them is from 2017 and is still running the latest software, which is unusual in tech but Synology does it regularly.

My NAS has hard drives in it. I’ve contemplated getting a new small 2-bay unit from them and using small SSDs instead, and storing “in process” stuff on it, then moving stuff to the big NAS for “cold storage”. This could also provide yet another layer of redundancy and backups too.

(Happy to discuss more if you want, but I do love my Synology. New units have some AI capabilities built in.)

The hardware is the foundation. The software is the interface. The AI is the worker. If the worker makes a mistake, the foundation must hold.

A Note on Reinforcement Learning

Oh great, reinforcement learning again. Yay.

One difference worth naming. On one recent research project I was working in an area that was entirely foreign to me — I knew nothing about the industry or the players — so I was relying on AI just to get a foothold.

If you already know your area, you can constrain the AI yourself: use a tool like NotebookLM to limit it to the sources you’ve collected and trust. That usually produces much better output than the freewheeling approach I had to take.

When you know the domain, you can constrain the model. When you don’t, you need a safety net.

Takeaways

1. Back up in two places. Synced is not backed up. Synced is just another copy of the same potential error.
2. Verify deletions. Never trust a model’s determination of safety of a proposed destructive action without a secondary check.
3. Build your own system. Tools will come, but they won’t work exactly like you do. Build the workflow that fits your brain.
4. Hardware matters. A reliable NAS is cheaper than a lost project. Although with today’s hard drive prices…

It was critical for both the customer’s needs and for our reputation. But mostly, it was critical for my sanity.

More on how I set backups up from day one: Make Backups. Offsite. Air-Gapped. and I Don’t Know Linux. I Built a Linux AI Server Anyway.

Written by a human. (Who is now very careful about backups.)

I’ve been saying this for a while. This piece was pulled together by AI from 5 of my social media posts (12/2025 - 03/2026) and polished for Cairoglyphics.ai.