Week 2 sucked.
Like, for real. Not the fun kind of suck where you’re building cool stuff and learning fast. More like I broke everything, lost access to tools I depended on, burned money, and seriously wondered what the hell I’m doing with my life. That said, it was probably one of the most useful weeks so far, but I don’t want to do it again. Week 2 was not really about building. It was about learning what kind of systems you need if you want to build with AI without constantly depending on it 24-7.
The Wake-Up Call: Dependency Is a Hidden Risk
On April 4, I lost access to my Claude Max subscription, because, apparently, OpenClaw is banned by Anthropic (the guys who own Claude). I DID have about a 24-hour lead time before I needed to transition off of Anthropic, but I don’t read many emails, and I missed it. When I did lose access, I panicked. This is because I built a dependency on a single model, Opus 4.6, while I was still learning how all of this worked. It’s THE BEST model out there, but I’ve had it explained to me like learning to drive in a Ferrari and then acting shocked when someone takes the keys. But I did PAY for the Ferrari, right? Er, okay, well I rented it at least.
What I considered
- Trying to bypass restrictions
- Pushing harder to regain access
- Pretending this was somehow a stable foundation
What I decided
- Not worth risking a ban
- Not worth building on shaky ground
- Not worth pretending dependency is the same thing as a system
The lesson here is that if your whole workflow depends on one model, you do not have a system. You have a single point of failure. So, I switched to ChatGPT Pro. And I wasn’t happy about it.
ChatGPT Pro Reality Check
At first, it felt broken.
- Responses taking 5 to 6 minutes
- Sessions freezing
- “Something went wrong” errors
It is amazing how impatient you get once you get used to near-instant responses from a top-tier model. At one point I genuinely thought ChatGPT forgot about me. Weird feeling, indeed. Turns out my session context was overloaded and I tried to have the model send me updates every 30-90 seconds, but that’s not how these models work. Large language models do not have little internal alarm clocks. They do not “check in” on their own. They run, and they respond.
From this I learned to ask for an estimate before the work starts, not fake status updates during the work.
The Website Disaster
I had a solid site. It worked.
- Booking flows
- Modals (website pop-ups)
- Schedules
- Notifications
- Sandboxed Stripe
Then I did the thing that always seems harmless right before it is not. I cleared the context with /new. You do that because you do not want the AI dragging around days of old context while trying to solve a fresh problem. Then I asked Ava, my AI, to make a minor update to the site. Total destruction.
- Broken HTML and CSS wrappers
- Dead links
- Missing modals
- Wrong schedule
- Hours of work gone in minutes
Why? Because the AI read its notes, assumed the site still matched those notes, and started “fixing” things that were not broken. The lesson here is that after any context reset, run a read-only audit before making changes. Very painful lesson.
No Snapshots = No Mercy
At the time, I had:
- No Git
- No backups
- No rollback path
So when things broke, there was no undo. Just me sitting there going “please fix what you broke.” And I feel like I have to say please with ChatGPT because that model gives me the corporate creeps. But, with ChatGPT I did fix things.
So, I set up:
- Git
- database backups with
mysqldump - a site snapshot workflow
Now I can save state, roll back quickly, and stop reverse-engineering my own mistakes from memory. Version control is a MUST HAVE. If you don’t have it, you’re not building; you’re gambling.
Root Access: Yeah, That Was Dumb
At one point I gave Ava full root access to my server. Which meant she could, in theory:
- modify anything
- delete anything
- break everything
Now, to be fair, she did not delete everything, but she broke a lot of shit. So, I removed her root access and her a regular user. To be clear, she could’ve still destroyed the website as she did, but this removes a particular layer or stress from my brain since she could’ve also destroyed my server. And yes, the reason I gave her root access was so that she could install and configure the web stack, server security (ironically), and SSL.
OpenClaw Was Blocking Me, Not the Model
I kept seeing that something went wrong and so I blamed the model. But the real issue The real issue was OpenClaw security was blocking command where I was getting strange errors like “long SQL commands looked obfuscated,” or approval mode set to block instead of ask.
So I changed the approval behavior to ask: on-miss. Funnily enough, I had to use Claude Code from another machine to help fix the environment Ava was running in. At this point I’m starting to wonder why in the hell I’m even using OpenClaw when I can just use Claude Code…?
Gemini: A Fast Way to Burn Money
I switched to Gemini for stability. It worked. It also burned through around $10 to $11 in API credits in about 90 minutes.That is a very efficient way to learn a billing lesson. I learned that you don’t save money with Gemini API pay-as-you-go. And that’s when I decided to stop using Gemini, but AFTER it fixed the website errors where all my images were turned into text. So, at least Gemini helped fix ChatGPT’s major error.
If I Had to Boil Week 2 Down
- Never depend on one model
- Always have version control and backups
- Limit AI permissions aggressively
- Verify before modifying
- Understand your tools before blaming them
- Very frustrating
It hurt, but I’m learning…
New to the tech terms? Check out the Glossary for plain-English definitions.
Leave a Reply