Why a system prompt deserves its own tool
Because it has different economics from every other part of a request. A user message is paid for once. A system prompt is paid for on every request, by every user, for as long as the product exists.
A hundred tokens of hedging boilerplate is not a hundred tokens. At a million requests a month on a mid-tier model, it is a few hundred dollars a year for text that changes nothing about the output. This tool breaks the prompt into sections, prices each one annually, and makes that visible as a number rather than an intuition.
What usually turns out to be dead weight
Politeness that changes nothing
"You are a helpful, harmless and honest assistant. Please be polite and professional at all times." Modern instruction-tuned models are already this. The sentence is reassurance for the author, and it ships on every request.
The same rule stated three times
Prompts accumulate. Someone hits a failure, adds an instruction, and the original instruction covering the same case is never removed. Repetition does not improve compliance in proportion to its cost — and past a point it actively hurts, because it dilutes the instructions that matter. The section breakdown makes duplicates obvious when you see them priced side by side.
Edge cases that should be one general rule
Fifteen enumerated special cases usually compress into two well-phrased principles. The enumeration was how you discovered the rule; it is not how the rule needs to be expressed to the model.
Formatting instructions for output nobody reads
"Use markdown headings where appropriate" costs real money if the output is parsed by a program rather than displayed to a person.
Does caching make this moot?
It reduces the cost, it does not remove it. A cached system prompt still bills at roughly a tenth of the base input rate on every request — cheap, not free — and caches expire, commonly after five minutes, so low-traffic periods pay full price.
More importantly, caching does nothing about the other cost: context window space. Every token of system prompt is a token unavailable for retrieved documents or conversation history. On a long-running agent that constraint bites well before the money does. See what it does to a conversation in the chat cost estimator, where the system prompt is resent on every single turn.
How to work through the breakdown
- Paste the prompt and set your real monthly request volume. Blank lines separate sections.
- Sort your attention by the yearly column, not by how wrong a section feels. Intuition is bad at this; the largest paragraph is rarely the one you suspected.
- For each expensive section ask one question: if I deleted this, what would break? If the answer is "nothing I can name", delete it and test.
- Re-measure. Most teams find twenty to forty percent on a first pass through a prompt that has been edited by several people over a year.
When trimming is not enough
If the prompt has grown by accretion, editing it paragraph by paragraph preserves the accretion. Rebuilding it from the intent is often faster and always shorter — the prompt builder gives you the structure to do that in, and prices the result as you write. Then run the output through the prompt optimizer for the structural waste that survives any rewrite.