# Evaluate a block, prompt or document

*Get a score and suggestions for a saved block, prompt or Markdown document, or check it for private details before you share it.*

Online version: https://promptmason.ai/docs/evaluate

Evaluate reads a saved block, prompt or Markdown document and gives you notes on it. There are two kinds:

- **Grade**: a score out of 10, what works, and what to improve.
- **Privacy**: a check for personal, financial or confidential details before you paste the text into a public AI.

Evaluate never changes the item and doesn't save anything. It's on every plan.

## Before you start

Evaluate needs the same two things as AI assist: the **AI assist** switch on in **Settings → Features**, and the local model downloaded or an AI provider set up. Until then, the **Evaluate** button doesn't appear. See [Set up AI](https://promptmason.ai/docs/ai-setup).

## Where to find Evaluate

The **Evaluate** button shows when you're viewing an item, not editing it:

- **Blocks:** select a block.
- **Prompts:** select a static prompt, a dynamic prompt or a Markdown document.

    ![A block's pane with an Evaluate button.](https://promptmason.ai/images/docs/evaluate-button-on-block-pane.webp)

While you edit, the editor shows **AI assist** instead (see [Improve text with AI assist](https://promptmason.ai/docs/ai-assist)).

## Grade an item

1. Select the block, prompt or document on the **Blocks** or **Prompts** screen.
2. Click **Evaluate**.
3. Under **Evaluation**, **Grade** is already chosen.
4. Under **Model**, check the provider and model. Choose another if you want.
5. Click **Run**, or press Enter.

    ![The Evaluate box with Grade chosen, a Model list and a Run button.](https://promptmason.ai/images/docs/evaluate-grade-popover.webp)

A panel opens at the top of the item. It shows "[Provider] is grading…" while the notes arrive, then "Grade from [Provider]".

### Reading a grade

The model is asked to answer in this order:

- The first line is the score, for example **Score: 7/10**.
- The next line starts **What works:** and says what's good about it.
- Then come two to four suggestions, one per line, saying what to change.
- For a prompt made of blocks, each suggestion starts with the type of the block it's about, for example "Constraints: no output length is stated". Change that block to act on it.

    ![The Grade panel showing a score out of 10, what works and suggestions.](https://promptmason.ai/images/docs/evaluate-grade-result-panel.webp)

To act on a suggestion, edit the item yourself, or open it for editing and paste the suggestion into **What should change? (optional)** in AI assist.

## Check for private details

1. Select the block, prompt or document.
2. Click **Evaluate**.
3. Under **Evaluation**, click **Privacy**.
4. Check **Model**. If the local model is downloaded, it's chosen for you, so the text stays on this computer. If you choose a cloud provider, a line under it says **Sends the text to** followed by the provider's name.
5. Click **Run**.

The panel shows "[Provider] is checking…", then "Privacy check from [Provider]".

### Reading a privacy check

The model is asked to answer in this order:

- The first line is a score, for example **Score: 8/10**. **10/10** means nothing sensitive was found; a lower score means more, or more serious, findings.
- The next line starts **What works:** and says what's already safe.
- Then there's one line for each finding, up to five: what kind of detail it is, the value with most of it hidden (for example "Sam R***"), and how to fix it.
- If nothing was found, the last line says **Nothing sensitive found.**

    ![The Privacy check panel with a score and findings with most of each value hidden.](https://promptmason.ai/images/docs/evaluate-privacy-result-panel.webp)

In a block or prompt, the usual fix is to replace the detail with a [variable](https://promptmason.ai/docs/variables), such as {{client name}}. PromptMason fills in a variable only when you copy or run the prompt, so the value isn't kept in the text. In a Markdown document, remove the detail or describe it in general words.

An AI model can miss things. Read the text yourself before you share it.

## Stop, copy or close the notes

- **Stop** ends the notes early. The panel then says **You stopped this early, so it is only part of a grade.** (or **a privacy check.**)
- **Copy notes** copies the notes to the clipboard. The button shows **Copied**.
- **Close** closes the panel.

## What gets evaluated

- The **saved** item. Save your changes before you evaluate.
- Variables stay as written, such as {{client name}}. Values are never filled in.
- Hidden notes are left out.
- A dynamic prompt is read as a whole, block by block, in order.
- A Markdown document is read exactly as stored.

With a cloud provider, the item's text and Evaluate's instructions go to that provider when you click **Run**: for **Grade**, also its name and, for a prompt made of blocks, each block's type and name. The **Before you use AI assist** window asks first, until you tick its **Don't show this again** box (see [Before text is sent to a provider](https://promptmason.ai/docs/ai-assist#before-text-is-sent-to-a-provider)). With the local model, none of your text leaves your computer. See [What AI features send, and where](https://promptmason.ai/docs/ai-privacy).

To change what Evaluate asks the model, edit the **Grade** and **Privacy check** entries in [the AI instructions](https://promptmason.ai/docs/ai-instructions).

## If something goes wrong

| Message | What to do |
|---|---|
| **This block has no text to evaluate yet.** (or prompt, or document) | Add text, save, then evaluate. |
| **This prompt is too long to evaluate: AI assist works on text up to** [N] **characters.** | Shorten the item. For a prompt made of blocks, you can evaluate its library blocks one at a time on the **Blocks** screen. |
| A message about a provider or the local model | Evaluate shows the same messages as AI assist. See [If something goes wrong](https://promptmason.ai/docs/ai-assist#if-something-goes-wrong) and [Free, private AI on your PC](https://promptmason.ai/docs/local-model#messages-and-what-to-do). |

If your library is paused, Evaluate is paused too: the paused message shows and nothing is sent. See [When your library is paused](https://promptmason.ai/docs/paused-library).

## Want a hand?

Paste this prompt into Claude, ChatGPT or another AI assistant, together with this guide, to be walked through it step by step:

> I want to use Evaluate in PromptMason to check one of my prompts. Explain the difference between “Grade” and “Privacy”, help me run one, and help me act on the notes it gives me.