What 90 codebases and “you ran out of credits” and “memory full” taught me about adapting my work to be more effective

Three days ago I ran out of Claude Code credits. And I can tell you, that really suc*ed. I was in the middle of building and refining my projects and I just realized how dependent I was on these tools to build me my products.

I had started coding with AI 18 months ago and while I was getting better each day, I was still far away from being able to write code myself as a non-coder by training.

But here we were, I ran out of my AI coding credits and the timer showed “they will be available again in 3 days” - 3 Days? That is a damn long time. And I did not had that time.

Action 1: Buying Claude Code credits

My first action was to buy Claude Credits. I thought, hey I spend $217 per month on Claude Code, that is little compared to what I do - that should be ok to spend a bit more. So I spent $100 for extra credits.

I felt good - $100 is half of $200, so that should be good. But it wasn’t. A day later these $100 were gone and the “you are out of credits” hit my codebases again.

I tried it one more time as I thought I made a mistake somehow and booked the $100 into the wrong Claude environment - API credits for the codebases themselves instead of for the coding - but no. 30 (!!!!) minutes later, the #200 were gone. And I was pissed beyond. Tried to communicate with Claude support, but were only able to speak with an AI agent that was not helpful at all and did not wanted to forward me to a real human being.

For $217/month and another $300 in the whim of a second, I did not like that.

But I knew that I had two options: Either take a break, do other stuff or find another solution. And I decided for that other solution.

Action 2: Asking Gemini for a better solution

I asked Gemini for another solution, after even Claude.ai did not work anymore. And voila - we had 3 options:

• Run on your own Hardware and a self-hosted model

• Run on Cline

• Run on Continue, another AI Coding Agent

Interestingly, it did not suggested Jules, one of its own lab agents orAntigravity for that task. But well, here we were.

I did some research into buying Nvidia GPUs or the new Mac Studio M5 Max or Ultra, but after the costs were $2500-$5500+, I decided to skip that option.

Thus, only Cline or Continue were left. And Cursor, but I used Cursor before Claude Code and had a rip-off by them of enormous amounts of tokens being priced on my credit card that I had to re-order, that I wanted to avoid that issue for that 3-day bridge of tokens.

Action 3: Running DeepSeek on Cline

My next action was to install both Cline and Continue into VSCode and to code via them. As Cline was the first I installed, I went with it. It was a bit of a hassle. As I first created an account and bought the ClinePass (no, not ClassPass). This Pass turned out to not include the Deepseek Pro Model but only less powerful ones in their $9.99 pass. But I wanted a strong alternative model, and thus, I stopped the subscription again and selected Deepseek Pro 1.4 into Cline’s terminal window. You need to get an API key, paste it into Cline in VSCode and ready you are.

And voilá, it worked.

At least - kind of.

Here comes what I learned throughout that experiment and why - after all - I am happy to be back at Claude and still - make some major changes to my current setup.

What I Learned

Lesson 1: Have free memory capacity on your computer

First carefully, then more intensly, I used Cline + Deepseek on 4 of my code bases. Not in a chat window but in the terminal. And I left these terminals open when I went to bed. The next morning I tried to copy-paste images to my computer and got an ugly message “you don’t have enough memory left”.

I was irritated but not surprised as I managed that before already. And yet, I wasn’t sure where it came from specifically.

Claude was back at that time and ran an analysis for me on what was going on: the 4 open terminal windows had used and thus blocked over 40 GB of memory on my computer and I was left with 0.1% of it. Because it kept running over night, it accumulated these massive amounts of it and I basically ran out of it completely.

Learning: Never let Cline windows open over night. And have enough memory on your computer before using it. Or if that is not feasible, use another coding agent.

Lesson 2: Don’t keep your .venv, and .node and cache and downloaded files without usage on your computer when not in active use

While Claude ran its analysis, it also found a massive amount of GB on my computer that included libraries, download files, and cache files that I did not use in a long while. And it told me- you know you have 88 codebases on your computer, right?

I was shocked but not surprised. I knew I had a lot, but had cleaned them up a while ago and moved several of them to Github only, but apparently I created many more again since that moment.

The result, I decided to clean up my codebases and archive all those that I would not need anymore. I created an _archive folder and pushed them to Github/their latest versions and saved their env files (API Keys, database connectors and co) to another folder for all of those. That reduced my blocked memory massively on my computer and made it - surprisingly - also faster again.

Learning: From time to time it is super helpful to analyze your computer memory use and to make sure that clutter-files like those .venv and .node ones are deleted- they can be re-installed easily

Lesson 3: Deepseek was not as effective as I hoped it would be

Deepseek felt like going back to the early days of AI coding. It showed you all the files it changed in the terminal window one by one with the red and green marks for accept the change or not. It had way less abilities to do code for you and you were doing manual actions, and bash and curl commands and similar on a much more regular basis. It sounded amazingly confident but still messed up big time- it just would not tell you that and I did not spend too much time to improve its performance in these three days. It was performing well though where it had to copy parts of Claude’s code or use the same mechanism for a different agent. But cross-codebase? It was a desaster and it did a rather bad job- or more precisely made it worse than before.

Learning: When using less-capable models for coding, make sure to properly prompt them for that expected behavior and/or don’t use them and/or use them only for smaller changes within one or two coding files.

Lesson 4: Going back to the roots - a welcoming change of perspective

And yet, while from a productivity perspective, Deepseek did not convince me, on a calming perspective it did. Everything slows down. You are coding slower, you have to figure out more for yourself, you look at what you produce more carefully. Those actions helped me to see coding again from the early beginnings and to at least a bit better, understand the hours and frustrations a coder without any AI had to endure to build those systems and products in the early days. I also spent countless hours with early AI agents and those were frustrating at times. I self-taught myself, but I had AI agents already - so it was a special frustration. One however, that till that date helps me to better understand what is going on when I let AI code features for me. The deeper meaning of what the individual pieces of code mean, how they are structured, which files a codebase have and how to stick them together.

Learning: Sometimes it is good to step a step back and look how far you have come and to understand that those early painful moments were helpful for the journey ahead. Being thankful for what they have taught me.

Lesson 5: Build a better system to burn less tokens and code more effectively

Nevertheless, I am back coding with Claude Code, but I also learned that I am not willing to spend more than $217per month if it does not necessarily need to be. So what to do about it? I knew I needed a better system. One, that reduces my token amount drastically and maybe - well I did it now already - moving down from coding mainly with Fable 5.1 to Opus 5.5.

My previous approach with coding was rather flexible and interactive and chat-based. And while I will keep it that way, as it is the most fun to work with, it had one major issue: I would code in one session-window for hours. And thus the re-loading of information that was necessary within that session got bigger and bigger, eating more tokens with each new message.

To reduce that, I now work on creating a clean system -one that has all the relevant rules, decisions, sessions and features per codebase. All documented and updated automatically, so that when I start a new session, it is all already there. And each agent does not need to start from scratch. And the documentation - it lives per codebase, instead of in my Notion. Notion had the benefit of having it all across code bases within one folder, and to have a history of what I did. All in one table, easy to search and structure. And it was external -in case something would break.

But it was also a lot of copy-paste and a system that at some point was bound to be running out of it’s capacity. And here we are now - one Claude across all development folders within. Automated rules for each of them and a documentation of what is going on - where are they deployed, where are they running, all its services and so on.

So that when I open a new session window, Claude knows what to do and can start right away. The system is in test mode right now - and once I have the first results I will share them with you.

Learning: Sometimes it is time to change your running system. Not because you necessarily feel excited about it - but because it is necessary and you know - deep down - it will save you not only costs, but also save resources and make you more effective. If that time has come - don’t hesitate but go that step.

Some last notes

If you wonder how much Deepseek did cost me? - $1.5 for 2 days. Clearly a difference to Claude code per API. But know the trade-offs.

If you did not understand all the words in that article, let me know. Then I try to improve that next time or explain them to you. It is surprising how fast you take on an industry’s vocabulary once you dive deeper into it. And you forget that you had no idea about these before neither.

A Hope

That article is for you when you run into similar issues and you wonder - where is the light at the end of the tunnel among all these aspects of code and changes and models. And where do I even start with building those products I care about.