Skip to content
[ d ] dassi / docs

Why repeat runs cost less

The first time dassi does a job, it pays full price. Every time after, most of that cost is already gone.

What a run costs, first time versus every time after Two bars. The first run is made of three parts: reading everything, working out the site, and doing the actual work. On every run after, the first two shrink to slivers and the rest of the bar is gone; only the actual work stays the same size. First run Reading everything Working out the site The actual work Every run after you no longer pay for this part The actual work — same job, same price Context dassi already sent — reused, not reprocessed Figuring out the page — replaced by a shortcut dassi wrote itself
Illustrative. How much you actually save depends on the job, the site, and which model provider you're pointed at.

It doesn’t re-read what it already sent

Section titled “It doesn’t re-read what it already sent”
The repeated part of the prompt is kept identical so it can be reused Two runs side by side. Each is a bar made of a large repeated part and a small part for today's request. In run two the repeated part is identical and marked reused; only the small part is new. The date is shown living in the small part, not the repeated part. Run 1 instructions · what you taught it · your rules today's request Run 2 same bytes — reused, not reprocessed today's request anything that changes — even the date — lives here →

Most of what dassi sends is the same on every run — the instructions, what you taught it, your rules. Anything that runs a model can reuse a repeated part instead of chewing through it again, on one condition: it has to arrive identical to last time.

So dassi keeps it identical, and keeps everything that varies out of it — right down to the date. Whoever runs the model — a big provider, your own machine, a server you rent — the reuse is the same trick. The only difference is what you get back: money off the bill (on Claude, up to 90% off the reused part), or the time not spent reprocessing.

It doesn’t work out the same website twice

Section titled “It doesn’t work out the same website twice”
First visit reads the page and writes a shortcut; later visits call the shortcut Two rows. First visit: the page flows to reading the whole page, which flows to the job done, and a shortcut is written below. Later visits: the page flows straight through the shortcut to the job done, skipping the reading step. First visit the page read the whole page job done writes a shortcut Every visit after the page calls the shortcut job done

The other half of the cost is figuring out the page — which button is the right button, where the table is, what the form wants.

As dassi works a site, it writes itself a shortcut: a small piece of code that knows where things are. Next time, it calls the shortcut instead of reading the page again.

What this means

Section titled “What this means”

The two compound: a weekly job reuses most of what it sent and skips most of what it read. The first run pays for the thinking; repeats shouldn’t — so a job gets cheaper the longer you keep it.

What it doesn’t do

Section titled “What it doesn’t do”
  • New work costs what new work costs. This is a discount on repetition. A job dassi has never done, on a site it has never seen, pays full price.
  • How much it’s worth varies. A hosted model turns it into money off a bill; your own hardware turns it into time. Either way the size of the win tracks how much of the job actually repeats.
  • A redesign resets the shortcut. When a site changes shape, the shortcut dassi wrote for it can stop fitting, and it goes back to working the page out directly until it has learned the new one.
Section titled “Related”