Today is day 2,158 of my daily running streak, and day 12 of a 30-day experiment. The experiment is to take a business concept through to launch, and hopefully profit, in 30 days, with AI doing 90% of the workload.

Claude Code usage limits shaped the whole of today. So that is what I want to write about.

What happened today

It was a productive day. We built 12 modules in a first-version format. Four of them were the operations side of the software. Think of it as the CRM and the day-to-day running of the business. You can pull reports off the customer base. You can look at the assets the business owns. The business here is a property business, so we can look at performance by property and by unit. All of it now sits in screens a person can read and use.

Building was not the slow part. Deciding what I wanted was. If you do not know the end result before you start, the tool cannot get you there. After the plan, the time goes on testing and fixing bugs. That is where most of my hours went.

Claude Code usage limits

I am doing most of this build with Claude Code, on the Max plan at $100 a month. There is a $200 plan with roughly four times that allowance. I am keeping costs down because I am bootstrapping this launch. The subscription also already sits in the operating costs of another business of mine, so I am borrowing it.

The limit that matters is the five-hour window. You get a set amount of usage in each window. When you run out, you wait for the reset.

Here is what I have found on my plan:

- Two coding sessions at once works. - Three or four sessions at once uses up my allowance before the window resets. - On the larger plan, three to five sessions should be fine. I have not tested that, so treat it as an expectation.

The pull towards running more sessions is strong. You give an instruction, then you wait for the output. Starting another session fills the wait. I did try four sessions on four modules today. The lesson is that the limit decides your pace, not your enthusiasm.

My working pattern is one session building something new and one session fixing errors. Fixing is the heavier job on usage. The tool has to get back up to speed on what it built, then work out where the fault is.

Sync status and health checks

I am not a software developer. What I am learning is that most software sits on top of a database. The risk is that what you see on screen is not what is stored. You can do a lot of work, and if it has not synced, you can lose it.

So every page now has a refresh button and a live sync status. I can see whether my data has reached the database.

The second change is a health check on every page. When you set up automations, you tend to forget them. You only find out they are not working when you need them. Now I can run a check on a page and get a list of what is set up and whether each item works. I screen grab the result and give it to Claude Code. It fixes problems faster and uses fewer tokens. That was the most useful change of the day.

The cost of AI inside the product

If you put AI inside the software, you need a budget for tokens. I pay for the subscription that builds the product. Running the AI features for the end user costs extra. The amount depends on how heavily it is used and which level of AI you choose.

Used carefully, I estimate it is around 50 times cheaper than human labour. That is my own estimate from this experiment, not a measured benchmark. It is still an overhead, and you need to plan for it.

Where we are

Day 12, still on track. I know the limits of the tools better than I did a fortnight ago, and that stops them slowing me down.

The link to the running

I had not planned a link to the streak when I started writing, but it is there. Limits are not the enemy of progress. On the road I run at a pace I can repeat tomorrow. The streak is at 2,158 consecutive days. I have covered 16,570.30km of the 40,075km lap of the world, which leaves 23,504.70km. That distance only gets covered by days I can repeat. Running three sessions at once in the AI build and hitting the cap is the same mistake as going out too hard on a run. You lose the next day.

I run barefoot-style, one day at a time, to raise £1M for children's causes. The experiment and the running share one rule for me. Respect the limit, keep the system simple, and turn up tomorrow.

If you are a business owner thinking about AI, start with the plan, understand your usage window, and build checks into every page. If you want to follow the 30 days, or the run, subscribe and come back tomorrow.

Watch the full episode: https://youtu.be/ahii_-DqQL0