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

# Testing prompt changes

> Re-classify real mail with your edited or proposed prompts and see exactly which classifications would change, before anything goes live.

## Overview

Changing an intent prompt affects every request assessed against it. Testing lets you see that impact on real mail first.

Axon re-classifies a batch of your actual requests twice — once with your live prompts, once with the changed ones — and shows you every classification that would move.

Nothing is saved and nothing goes live. Testing is a dry run.

> [!NOTE]
>
> Testing prompt changes is part of **Axon Premium**. See [Included vs. Axon Premium](/knowledge-base/axon/overview.md#included-vs-axon-premium).

## Before you start

You need **Editor** access to Axon, and at least one unsaved prompt change — the test runs against your draft, so there has to be something to compare.

## Two ways in

| Where | Button | What it tests |
| :- | :- | :- |
| Intent tree toolbar | **Test changed prompts** | Prompt edits you've made yourself but not yet saved. |
| Suggestion review dialog | **Test proposed prompts** | A proposed prompt from a [suggestion set](/knowledge-base/axon/prompt-suggestions.md), including any edits you've made to it. |

Both open the same test dialog and work the same way.

> [!NOTE]
>
> **Test changed prompts** stays greyed out until you've actually changed a prompt, and while a save is in progress.

## Run a test

**1. Choose which mail to test against**

Pick **Last mails** to use your most recent requests, or **Pick mails from Workstation** to choose specific ones yourself.

Use recent mail for a general read on the change. Hand-pick when you're targeting a particular problem — the requests that were being misclassified in the first place.

**2. Set the sample size**

For recent mail, choose **Last 100**, **Last 250**, **Last 500**, or enter a custom number up to 500.

Larger samples give a more reliable picture but take longer to run. 100 is a good starting point.

If you're picking mails from Workstation, you can select up to **50**.

**3. Run it**

Axon classifies every mail in the sample twice — once with your live prompts, once with the changed ones — and compares the results.

## Read the results

A summary appears at the top:

* **{n} of {total} mails changed intent** — how many classifications the change would move.
* **{n} unchanged** — how many would stay exactly as they are.
* **{n} failed** — mails that couldn't be classified in both runs. These are excluded from the results.

> [!NOTE]
>
> A failure means Axon couldn't complete both runs for that mail — a missing mail file, an error, or a timeout. It isn't a verdict on your prompt, and the mail simply doesn't count toward the comparison.

Below the summary, each changed mail is listed with:

* The **subject**, **sender**, and **date**.
* The intent it has today, and the intent it would move to.
* The **change in confidence**, shown as a percentage.

You also get a per-intent breakdown of how many mails would move **in** to each intent and how many would move **out** — useful for spotting a change that fixes one intent by quietly draining another.

## What the results tell you

A good change moves the mail you expected to move, and leaves the rest alone.

> [!WARNING]
>
> A high number of changed classifications isn't automatically good. Check *which* mails moved — a broad prompt can sweep up requests that were already being classified correctly.

Work through the changed mails and ask whether each move is one you wanted. If the answer is mostly no, go back and narrow the wording, then test again. You can run as many tests as you like before saving.

Confidence changes are worth a look too. A classification that moved to the right intent but with low confidence may need the prompt to be more specific, not just redirected.

## After testing

Testing changes nothing on its own. When you're happy with what you see:

* For your own edits — close the dialog and **save the intent tree**.
* For a proposed prompt — accept the set, then save the intent tree. See [Prompt suggestions](/knowledge-base/axon/prompt-suggestions.md).

If you're not happy, discard your changes and nothing is affected.

## What's next

* [Prompt suggestions](/knowledge-base/axon/prompt-suggestions.md) — review a proposed prompt change.
* [Auto-tuned intent descriptions](/knowledge-base/axon/auto-tuned-intent-descriptions) — how the whole tuning loop fits together.
* [Create and train custom intents](/knowledge-base/axon/create-custom-intents.md) — write prompts for your own intents.