Our Morning Routines Read 315 KB Before Breakfast

Scheduled AI routines spend most of their budget reading. Ours were reading 315 KB before breakfast, and more than half of it was work already finished.

That one finding applies to anything you run on a schedule against a budget.

We run four scheduled routines, starting a little after 6:00 AM. The biggest one builds our week board, the page that puts finished copy for every channel on screen before the workday starts. In early October we measured what they read, file by file.

Where does a scheduled routine spend its budget?

Reading.

The week board is 315 KB of HTML. Reading it whole costs somewhere around seventy-five to eighty thousand tokens, and the routine read it whole every morning. Here is where those 315 KB sit:

• The record of finished work: 160 KB, 97 items
• Still open: 35 KB
• Ready to post: 22 KB
• Next week: 19 KB
• Everything else: 79 KB

More than half the board is finished work. It never changes. It only grows. And it went into memory in full every morning, to be written back out untouched.

The second find: the routine fetched the published copy of the board from the web while an identical copy sat on the local disk.

What are the eight rules?

None of them are specific to our setup. They transfer to anything that runs on a clock.

1. Read a source once, then chain. Three of our routines read the same events calendar within minutes of each other. Routines that run in sequence should read each other's reports, never each other's sources. The second one needs the sentence "here is what changed," and the first one already wrote it.

2. Index the archive. The record can't be ignored, because it is what stops work being repeated. So a script reads the 160 KB and writes one line per finished item: an id, what it was, where it went. That came to 8 KB for 97 items, small enough to fit on one screen.

The important part is who does the reading. Python reads the 160 KB. The model reads the 8 KB. The expensive file still gets read in full every day, by the thing that reads for free.

This is the biggest saving in most systems, and it almost always sits in the same place: the archive, the log, the done pile, the part that only grows.

3. Prefer the local copy. Before reaching for the network, ask whether the remote thing is actually different from the file on disk. Comparing byte counts is cheap.

4. Text beats pixels. A screenshot costs many times what the text of the same page costs, and carries less. Take a picture only when the question is visual.

5. Look at a thing once. You made the image. Open it to check the composition, and move on.

6. One pass per deliverable. Write it, check it against the rules, hand it over. A rewrite because a better line occurred to you is where budgets go to die.

7. Gate expensive reads on freshness. Meeting notes that change once a week don't need decoding every morning. Check the modified date first.

8. Do the usable thing first. On a long run, produce what the person can use before what is merely tidy. A run that gets cut off halfway should still leave something on the screen.

What did it add up to?

• Building from the local file instead of the published one removes a read of roughly 75,000 tokens per run.
• Indexing the record turns 160 KB into 8 KB with nothing lost.
• Gating the meeting notes on their date removes a large decode six mornings in seven.

Together, by our estimate, the morning build now reads roughly half of what it read before. The file sizes are measured. The savings are estimated from those sizes, not read off a meter. The output didn't shrink by a line.

What did we keep?

Every routine. Every morning sweep still runs, including the ones that usually report nothing. A quiet sweep is the sweep working.

The archive, whole. Throwing it away would have been the cheapest change of all, and the one that made the work worse, because the record is how repeats get caught and how a person checks that the routine did what it said.

How often should a routine run?

Set frequency by who can change the source.

Our events sync watches a calendar other people add to, so it runs every morning. A day stale is a day an event is missing from the website. Our creator form sweep watches a form that fills slowly, and a submission that waits one morning costs nothing, so it runs every other day.

A quiet source with many authors still needs watching. A busy source with one author does not.

Frequently asked questions

What is a scheduled AI routine?
A task an AI system runs on its own at a set time, like building a morning report or checking a form for new entries.

What uses the most tokens in a routine?
Usually reading. Large files, archives and screenshots cost far more than the writing the routine produces.

Should I run routines less often to save money?
Make each run cheaper first. Running less often trades cost for staleness, and the right frequency depends on who can change what the routine watches.

Build yours with us

This is the kind of system we set up in one to one work, and the word we use for it is a harness. Tell us what you are running: https://thehouseofcuriosity.ai/consulting-services

Everything we run: https://luma.com/TheHouseofCuriosity