The Prompt Is An Economy
The model reuses the prompt from the front and stops at the first change. Part order sets the bill and the speed. Put the steady parts first and the new message last, and most of a turn comes back free.
I read the briefing card as setup. It is the biggest line on the bill, and it goes out every turn. The work is to pay for it once.
Prompt economy
Every word you hand the agent costs a coin. The big briefing card, who it is, the rules, the history, is almost the same every turn.
So you do not buy it twice. The old card sits on the table like coins already counted. You pay for the one new line you added today.
Spend new words where they earn their keep. Reuse everything else. Get the order wrong and you pay full price on every turn.
The savings are large. The same prompt costs full price on turn one and 16 percent of that on turn two.
One early change undoes that. Step through the panel below. The bill climbs back to 84 percent after one line near the front changes.
01Every turn re-sends the same expensive briefing
The model forgets between calls, so the system rebuilds and sends the card every turn. It carries who the agent is, the rules, the tools, and your project notes. That card is most of the words. One line is new each turn, your latest message.
Full price on every turn puts most of the bill on words you already sent. The system charges you for that card once, then reuses it.
02Reuse is a prefix, which makes order the lever
Reuse needs an exact match. Start at the first word and run forward while the words match, letter for letter. The first word that differs ends the run. You pay full price for everything after it.
That rule is why order saves money. The parts that stay the same go first, the card and the tool list, inside the free run. Your newest message goes last, where it cannot break the reuse in front of it. The chat grows at the end on purpose. Flip the front to a ticking clock below and the bill goes from one line to five.
03This explains choices that look odd from outside
Four standing choices in the system come from that rule alone. Move a part up in the panel below and watch the free share change. Each choice buys back a slice of the same free run.
- The tool list keeps the same order every turn. Shuffle it and the match breaks at the first reordered word, and you pay for the rest again.
- Rare tools stay out of the card, and the agent looks them up by search. That keeps the front small and steady.
- When old text has to go, the system protects the front first. Touching the front is the costliest edit available.
- The card stays identical turn after turn. Holding it still buys a steady agent and a smaller bill.
04The trap is a small change in the wrong place
One high-churn part near the front is enough to break the run. Three parts do this most often: a live clock, a self-shuffling tool list, and a counter that ticks up. Each writes a different first word, so reuse stops there.
Nothing behind it changed, and you pay for all of it again. Read your own prompt from the top and mark the first part that changed since the last turn. If you cannot find the mark, the cost and the speed of the next turn are both guesses.
The rule
Where the first change sits sets the bill. A tick at the front costs more than a fresh paragraph at the back. Sort the parts by how often each one changes, steadiest first. The bill drops and you cut nothing.
05Speed rides on the same mechanism
A reused front is a front the model does not read again, so the turn returns sooner. The habit that keeps the bill small keeps a long chat quick. When an agent turns slow and costly halfway through, look for a change near the front.
The cost
This economy pays for stillness and charges for change. It fights any feature that wants fresh data up front. One tick near the front multiplies the bill for a whole chat, and nothing throws an error. The share of the prompt that comes back free is the only reading that shows the slip.
Decision
My prompt layout is fixed now, and nothing that ticks lives in the steady part. Measure the front on a long chat: count how much of it came back free. Watch that number when the agent slows. The next note prices the speed a reader feels.