2 of 4
I run a crew of AI assistants on one monthly plan. At 12:12 one Saturday morning, one of them started a big job: 943 requests, one after another. It used almost half of the plan's five-hour allowance in one burst, and every other job had to wait until 1:56 that afternoon.
The job was useful. The problem was the timing: it spent what should have lasted the morning, all at once.
If you pay a flat monthly fee for an AI plan, you can lose money in two ways. Spend too fast and your work stops in the middle of the day. Spend too slowly and the hours you paid for disappear at the reset. Most people only notice the first one. The second one is quieter.
So three questions matter. How fast can you safely spend? Why does the obvious fix make it worse? And what goes wrong on the first day of each week?
Most people's first move is to buy the bigger plan. But a bigger allowance doesn't spend itself any more evenly. One big job still eats the short window, and the extra hours you paid for still vanish at the reset. You pay more to lose the same way.
The second move is to check your usage now and then. That fails exactly when it matters: the big job runs at night, when nobody is looking.
What works is a rule you can check in one glance.
Picture a line that starts at zero when your weekly allowance resets and rises to 90% by the next reset. Each day, look at how much you've used.
It's the same rule a sensible household uses for a monthly budget: check the balance against the calendar, not against how you feel. But a budget has one limit, and an AI plan has two.
Most plans also have a short window, often five hours, with its own cap. Hit it and work stops at once, whatever the week looks like. So I never start new work once that window is 70% used. That leaves room to finish what's already running.
That's the whole method. But it broke on the first day of every week, and the fix surprised me.
If your assistants run in Claude Code, a free tool I published, paceline, does the checking. Before a new job starts, it answers LAUNCH or WAIT, and says when to try again. That Saturday the week stood at 53% used against a line of 46%, so the answer was WAIT until 1:56 pm.
Setting it up takes about five minutes:
Install it: cargo install --git https://github.com/mackenziebowes/paceline
Add one line to your Claude Code settings so it can read your usage:
{ "statusLine": { "type": "command", "command": "paceline statusline" } }
Tell your assistant, in its instructions: "Before you start a new job, run paceline gate. On WAIT, don't start it."
If you don't use Claude Code, the rule still works by hand. Check your usage page at lunch, compare it to the straight line, and decide.
Right after a reset, the line sits near zero. Use a few percent and you're "over" it, and the rule says wait. That's wrong: nothing is scarce yet. So for the first day after a reset, I ignore the line and only watch the five-hour cap.
Five hours turned out to be too short: it cost more than it saved. An assistant that waits four hours forgets what it was doing, and getting it back up to speed uses more of the plan than the wait saved. A full day worked.
If your plan keeps running out mid-week, the cause is usually one big job started without a check, like mine. If it never runs out, you're probably paying for more than you use, and a smaller plan may do.
If you'd like help working out which one you are, write to me: mackenzie@mackenziebowes.com or @_mackenziebowes.
Next article →