> ## Documentation Index
> Fetch the complete documentation index at: https://docs.corepillar.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Optimise

> The Optimise switch in the ModelHub composer removes greetings, sign-offs and wordy phrases from your message before it is sent, so you pay for fewer tokens without changing what you asked.

People write to an AI the way they write to a colleague, with greetings, apologies and thanks. Every one of those words is billed. **Optimise** strips them out before your message is sent.

## Turning it on

The **Optimise** switch (with a sparkle icon) sits in the composer next to the model picker.

* It is **off** by default.
* When on, it applies to every message you send until you turn it off.
* It turns off again when you reload the page or switch project.

## What it changes

Optimise only removes or shortens fixed phrases. It never asks an AI to summarise or rewrite your message, because an AI "shortening" a prompt tends to drop the one constraint that mattered.

| Change | Example |
| - | - |
| Request preambles | "Could you please", "I was wondering if you could", "Would it be possible to" are removed |
| Sign-offs | "Thanks in advance", "Much appreciated", "Cheers" are removed |
| Opening greetings | "Hi there!", "Hello, hope you're well" at the start of a message are removed |
| Telling the AI what it is | "As an AI language model," is removed |
| Closing hedges | "if that makes sense", "if you don't mind" are removed |
| Bare courtesy words | "please" and "kindly" followed by more text are removed |
| Wordy phrases | "in order to" becomes "to", "prior to" becomes "before", "due to the fact that" becomes "because" |
| Spacing | Runs of three or more blank lines become one blank line; trailing spaces and runs of three or more spaces are tidied |

The complete rules are listed in [How prompt tidying works](/modelhub/how-it-works/prompt-tidying).

### What it leaves alone

* **Code.** Anything inside a fenced code block is sent exactly as written.
* **Words that carry instructions.** "just", "simply", "briefly", "only" and "exactly" are never removed. "Just the total" is a different request from "the total".
* **Attachments.** Optimise works on your typed message only. Documents are sent as extracted.
* **Short, plain messages.** A terse message often has nothing to remove, and is sent unchanged.

## Examples

| You type | Sent to the model |
| - | - |
| Could you please summarise the Q3 refund policy? Thanks in advance! | summarise the Q3 refund policy? |
| Hi there! Hope you're well. I was wondering if you could review this contract. Much appreciated. | review this contract. |
| As an AI language model, can you explain OAuth? Thank you so much. | explain OAuth? |
| In order to ship prior to Friday, we need to comply with the policy. | to ship before Friday, we need to follow the policy. |
| Just the total, please. | Just the total, please. *(unchanged)* |

## Seeing what it did

Your message bubble first shows what you typed, then switches to the trimmed and masked version once the server has processed it. That is exactly what the model received.

After the answer, the report card shows **N tokens trimmed** when Optimise saved anything. Open **Details** and the **Usage** section lists each change, for example "removed request preamble", "removed sign-off" or "shortened 3 wordy phrases".

The token count is an estimate at about four characters per token, comparing your original text with the trimmed text. Your actual bill uses the provider's reported token counts.

## How much it saves

It depends entirely on how you write. On deliberately wordy business English, CorePillar measured about a 37% reduction across a six-message sample (45 of roughly 121 tokens). On a terse, well-written prompt it saves nothing, which is the correct result.

Optimise runs before routing and privacy screening, so the trimmed text is what gets scored and masked. Removing filler can lower a message's difficulty score slightly, because length is one of the routing signals.

Your totals over time are on the [AI Usage](/modelhub/usage) page under **Routing & savings**, as **Prompt optimisation**.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.