Token waste patterns
The 26 patterns the analyser looks for, with the reasoning behind each one. Ordered by how much money they typically move, largest first.
Every entry is checked automatically when you run a prompt through the analyser.
Caching
Caching is a configuration change rather than a rewrite, and it routinely moves a bill by more than every text edit combined.
- high
A cacheable prompt, uncached
Reading a cached prefix costs a tenth of sending it again. This prompt qualifies.
- high
Per-request values above your static content
A cache prefix ends at the first byte that changes. Anything after it pays full price forever.
- medium
A one-hour cache for a five-minute workload
Longer cache lifetimes cost more to write. They only pay back if you actually reuse them.
Structure
What the prompt is made of, and how much of it earns its place. These need judgement, so the optimizer reports them rather than applying them.
- high
The same rule, stated twice
auto-fixablePrompts grow by accretion. Instructions get re-added rather than edited.
- high
More examples than the model needs
Few-shot examples are the most expensive content in most prompts, and they have a ceiling.
- high
Encoded data pasted into the prompt
Base64 and data URIs tokenize terribly — roughly one token per two or three characters.
- medium
A long list of things not to do
Prohibitions accumulate one incident at a time and are rarely removed.
Tools & schemas
Tool definitions and output formats. Both are re-sent on every request, and both are commonly larger than the system prompt they accompany.
- medium
JSON indented for a human reader
auto-fixablePretty-printing JSON in a prompt adds 20-30% tokens for whitespace nobody reads.
- high
A JSON schema written out in English
Describing your output shape in prose costs more than declaring it, and works less well.
- high
What your tools cost before anyone calls one
Tool schemas are re-sent on every request, plus a provider system prompt you never see.
- medium
Tool schemas carrying dead weight
Long descriptions, JSON Schema boilerplate, and enums with a hundred values.
Model choice
Choices made outside the prompt that change what it costs: which model, which tokenizer, which price tier, and whether the bill is really about input at all.
- high
A newer model at a lower price
Some model upgrades cost less on both input and output than the model you're currently using.
- high
Your prompt crossed a pricing cliff
Some models double their input price above a prompt-size threshold.
- medium
This model counts your prompt differently
The same text is not the same number of tokens on every model. Upgrades can raise your bill silently.
- medium
Running out of room
A prompt filling most of the context window leaves nothing for the conversation.
- medium
Output is most of this bill
When output is most of the bill, shortening the prompt barely moves the total.
Formatting
Whitespace, punctuation and emphasis. Individually trivial, collectively real, and always safe to fix automatically.
- medium
Markdown tables padded for alignment
auto-fixableColumn alignment is whitespace, and whitespace is tokens.
- low
Shouting is billed by the letter
auto-fixableALL-CAPS words split into more tokens than the same word in title case.
- medium
Smart quotes from a word processor
auto-fixableCurly quotes and em dashes cost several tokens each. Their ASCII twins cost one.
- low
ASCII art in a paid channel
auto-fixableRows of equals signs are for humans reading a terminal, not for the model.
- low
Whitespace you are paying to store
auto-fixableTrailing spaces and stacked blank lines are billable and invisible.
Filler
Words that do not change the output. Usually the easiest tokens to cut, and a sign the prompt has never been audited.
- medium
A role that describes nothing
auto-fixable"You are a helpful AI assistant" tells the model nothing it did not already assume.
- low
Prompt folklore
auto-fixableTips, deep breaths, and other emotional-appeal phrasing, copied from 2023-era prompting advice and never removed.
- low
Chat pleasantries in a system prompt
auto-fixableSign-offs written for a human conversation, left in a machine instruction.
- low
Politeness aimed at a billing meter
auto-fixablePlease, thank you, and "I would like you to" cost tokens on every request forever.
- low
Long ways of saying short things
auto-fixable"Due to the fact that" is four tokens. "Because" is one.