← All notes

At 4 pm I message the agent so that at 9 pm I have a fresh limit

Sławek Łukjanow · · Not by Prompt Alone · episode 13 · Czytaj po polsku

At 4 pm I send the agent some really tiny prompt. Not because I have time to work right then. Just to get the clock ticking.

My real working time is in the evening. Let's say 8 to 10 pm. And within an hour I can burn through the whole limit meant for a five-hour window. If I only started at 8 pm, I'd get the next budget at one in the morning. Great timing, if I'm not planning to sleep. Except I am 😉

If I start the window at 4 pm, it resets at 9 pm. During the first hour of the evening session I use what's left of one window, and during the second I already have the next one. Two hours of my time can cover the budgets of two five-hour windows.

It's a small detail. But it changed how I plan my work with agents.

Five hours doesn't mean five hours of work

It's easy to read "5h limit" as if the agent had fuel for five hours of continuous work. In fact five hours is the length of the window in which I can use a certain budget. A hard task, lots of context, or a little wink at Fable 5.2, and you just watch Claude snort your entire limit.

Claude's documentation says it plainly: the window starts with your first message and resets five hours later. That's exactly what I use by sending a small command earlier.

That's how it works in Claude on the Pro, Max and Ultra plans. I'm not generalising it to Codex: this year OpenAI changed the rules of the five-hour window several times, and switched it off for some plans.

In my example evening it looks like this:

  • 4:00 pm - a tiny prompt starts the countdown.
  • 8:00-9:00 pm - I work with the agent, using what's left of the first window.
  • 9:00 pm - the session limit resets.
  • 9:00-10:00 pm - I keep working, with a fresh budget.

This doesn't give me a bigger weekly limit. I just move the reset to a moment when I can actually use it. And of course this only makes sense on a plan with that mechanism. Not all tools and subscriptions count it the same way.

I haven't measured savings in money or the number of extra tokens. The benefit is more down-to-earth: at 9 pm I don't have to stop working just because I started the countdown only an hour earlier.

And then I go to sleep

The evening session is one part of the puzzle. The second starts when I'm no longer at the computer.

I've written before about the Mac mini as a base for agents. Earlier my laptop could fall asleep with the lid closed, and all the work fell asleep with it. The agent supposedly had a task, but for an hour nothing happened. A machine available around the clock removed that obstacle.

Which left a more interesting question: what should the agent do when I'm not there?

I can leave it a vague idea and in the morning read a long list of assumptions it made up on its own. Or I can work the task out with it beforehand, so that at night it has something to execute. In practice a lot of the work moves to preparation. We need to agree on what we want, what the result should be and what we'll be able to check.

That sounds less spectacular than "I went to sleep and woke up with an app". But it describes much better how we work now.

Requirements by day, PRs by night

At ANSLAN we're currently working with one of our clients on a new project. During the day we refine the requirements together, with the agents. Talking to people and deciding what should be built happen when we can make those decisions together.

The point is to prepare the work well enough that overnight Claude Code and Codex can each put out anywhere from a few to a dozen or so pull requests. A PR is a proposed change waiting to be checked. The number of such proposals alone doesn't tell you whether the client got what they need.

That's why the next stage is in the morning. With fresh limits, the agents review each other's code. They set up a test environment, run e2e tests - meaning they go through whole scenarios of how the system works - and hand us a result to test with the client.

So people come back to something they can run and check. Only then can you see whether a requirement that seemed clear yesterday produces the right effect on screen. And whether we forgot a case the client will think of immediately.

An example from one of the recent nights. In the morning a report from Codex was waiting for us: "I prepared and pushed 6 PRs [...] Locally green: 121 backend tests, 22 shared, 22 E2E and the web build." Lovely! All green 🙂

And then one of our people checked what had actually been built and replied briefly: "It absolutely must not be possible to book a resource that isn't in a ready state".

During the day the requirements had seemed well defined. The agent did what it read from them, the tests confirmed it... and it still let through something a person who knows the business would never let through. No test checks a rule nobody thought of writing down.

More freedom at night

And here's our conscious choice. At night we give the agents more freedom than during the day.

During the day, when an agent has a doubt, it can ask and get an answer shortly. At night nobody answers. If it stopped on a trivial question, it would stand still until morning, and the whole night of work with it. I'd rather it made a reasonable assumption and kept going.

As long as a decision isn't a "one-way door" - a door you can't come back through - the cost of a mistake is small. A quick fix in the morning. And the gain is huge - a night of work instead of a night of waiting for an answer.

Why a second agent does the checking

In the episode about Agile ZOO I described this in practice. Codex built the workshop version, and in the morning Claude didn't just read the code, it played a session with many participants. There were loads of green tests. Even so, that kind of check turned up problems with simultaneous clicks and with a player coming back after closing the tab.

That example explains well why our rhythm includes mutual review. The author of a change already has their own solution in their head. A second agent can come in with different questions and check something the first one didn't anticipate.

It's no guarantee that two models will find every bug. They may well share the same wrong assumption. That's why the process includes both running the system and going back to people and testing with the client. Code can do exactly what we wrote down, and we can still be unhappy with the result.

Agile ZOO also had a night-time slip-up with a release. I described it along with the fixes. It's an important lesson, but I don't want it to define night work as a whole. In the morning I want a result to check, not mandatory cleanup after surprises.

Guessing all night would cost the most

The tip with the tiny 4 pm prompt is useful because it fits my day. But setting the clock right won't prepare the work for me.

Freedom at night has its limits, though. If I leave the agent a task with an important unanswered question - one that can't easily be undone in the morning - a renewed limit may simply let it build on a wrong assumption for longer. Before I hand it the night, I'd rather spend time on the requirements. And before we hand a result to the client, we need a review and a working test environment.

This rhythm combines several different kinds of time: people's time, when we can talk; agents' time, when they can execute prepared tasks; and the moments when limits reset. Not everything has to happen while I'm sitting in front of the screen.

And that's probably the nicest change. I can finish my evening and the prepared work still moves forward. In the morning we come back to concrete changes and questions we can discuss with the client. We still have plenty to do, we just start from a different place 🙂


What this means for your team: check whether the moment your limits reset matches the hours you actually work. And prepare overnight tasks so that in the morning you can judge the result - with review, tests and a return to the people who define the requirements. At ANSLAN this is how we organise the work of people with Claude Code and Codex on a real client project.