> ## 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.

# Auto-tuned intent descriptions

> Axon learns from your real mail and proposes improvements to your intent prompts, so you can tune classification on evidence instead of guesswork.

## Overview

Intent prompts describe what each intent means. When classification goes wrong, the fix is usually a change to one of those prompts — but working out *which* prompt and *what* change has always meant reading through misclassified mail and making a judgement call.

Auto-tuned intent descriptions close that loop. Axon watches how your requests are actually classified, finds the patterns your prompts don't cover yet, and proposes specific wording changes — along with the mail it learned from and the accuracy improvement it expects.

Nothing changes on its own. Every proposal is yours to review, edit, test, and accept.

> [!NOTE]
>
> Auto-tuned intent descriptions are part of **Axon Premium**. There's nothing to switch on — if your organisation has Premium, it's already running. See [Included vs. Axon Premium](/knowledge-base/axon/overview.md#included-vs-axon-premium).

## What you get

Three things work together on every intent.

| | What it does |
| :- | :- |
| **[Prompt suggestions](/knowledge-base/axon/prompt-suggestions.md)** | Proposed prompt updates, based on how your real mail was classified. |
| **[Prompt testing](/knowledge-base/axon/testing-prompt-changes.md)** | Re-classify real mail with a proposed prompt — or your own edit — and see exactly what would change, before anything goes live. |
| **Prompt history** | Every saved version of an intent prompt, who changed it, and what changed. See [Prompt history](#prompt-history) below. |

## How suggestions build up

Suggestions are grounded in your own corrected mail, so they need evidence before they can say anything useful.

**1. Axon collects examples**

Every time someone corrects or confirms an intent in Flow, Axon keeps that mail as an example for the intent. This happens continuously, as part of real-time learning.

**2. The intent reaches 50 examples**

Once an intent has gathered **50 examples**, it has enough evidence for Axon to reason about, and suggestions unlock for that intent.

**3. Axon proposes a change**

Axon looks for patterns your prompts don't cover yet and creates a suggestion set for you to review.

### You don't have to wait for the monthly run

Axon reviews your intents on a monthly cycle, but that isn't the only way suggestions appear. An intent unlocks as soon as it has gathered 50 examples — so if your team has been correcting heavily, suggestions can show up well before the month is out.

In practice: the busier an intent is, the sooner you'll see a proposal for it.

> [!NOTE]
>
> If your classifications have been stable, there's nothing to suggest and no set is created. Organisations with few corrections may not see suggestions for a while — that isn't a fault, it means Axon has no evidence that anything needs changing.

## When to use it

Reach for suggestions when Axon doesn't feel like it's improving — when the same kind of request keeps landing in the wrong place even though your team has been correcting it.

That's the signal that the problem is the intent's **definition** rather than one sender's phrasing. Examples alone can't fix a definition; a prompt change can.

It's also worth checking after you've worked through a backlog of misclassified requests, or after adding intents that overlap with ones you already have.

> [!TIP]
>
> Use the intent chart in [Analytics](/knowledge-base/analytics/overview.md) to see how your intents are performing today. It's the quickest way to spot which intents are worth tuning.

## Prompt history

Every time an intent prompt is saved, Axon keeps the previous version. Open an intent and look next to the **Instruction Prompt** label — if the intent has earlier versions, you'll see a **Prompt history** button with the number of versions.

Click it to open the history. On the left is the list of saved versions, newest first. Select one and the right-hand side shows how it differs from the prompt that's live today.

Each version tells you:

| | What it shows |
| :- | :- |
| **Who saved it** | The name of the person who saved it, or **Axon learnings** when the version came from an accepted suggestion. |
| **When** | The date it was saved. |
| **Source** | **Manual edit** for a prompt someone wrote, or **From learnings** for one that came from a suggestion. |
| **Summary** | A short description of what changed, when one is available. |

The newest version is marked **Latest**.

### Revert to an earlier version

Select the version you want and click **Revert to this version**. This brings the old wording back into the prompt editor.

> [!WARNING]
>
> Reverting loads the earlier prompt into your draft — it doesn't save it. You still need to **save the intent tree** to apply it, which also means you can test the reverted wording first.

Reverting requires **Editor** access to Axon, and you can't revert to the version that's already live.

## What's next

* [Prompt suggestions](/knowledge-base/axon/prompt-suggestions.md) — review, edit, and accept a proposed set.
* [Testing prompt changes](/knowledge-base/axon/testing-prompt-changes.md) — see the impact before anything goes live.