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

# AI guardrails: control what Dear's AI can write

> Guardrails let you define syntax and framing patterns that Dear will automatically rewrite or remove from AI suggestions and rewrites.

Guardrails are rules that run after Dear generates text. When a suggestion or rewrite contains a blocked pattern, Dear silently replaces it with your preferred alternative — or removes it entirely if you don't specify one. You stay in control of the specific phrases, punctuation, and rhetorical moves the AI is allowed to use.

The guardrails system is called **Blacklist** in AI preferences. It's **on by default**.

## Built-in rules

Dear ships with three built-in guardrails. All three are enabled by default, with scopes chosen to reflect typical writing preferences.

### Dash (em dash → double dash)

|                   |                                                 |
| ----------------- | ----------------------------------------------- |
| **Type**          | Syntax                                          |
| **Blocked**       | `—` (em dash, U+2014) and `–` (en dash, U+2013) |
| **Replace with**  | `--` (double dash)                              |
| **Default scope** | Everywhere                                      |
| **Default state** | On                                              |

Many writers prefer the typewriter convention of `--` over typeset em dashes in prose. This rule catches both em and en dashes in AI output.

### Framing A ("It's not X, it's Y")

|                   |                                                   |
| ----------------- | ------------------------------------------------- |
| **Type**          | Framing                                           |
| **Blocked**       | Constructions matching `"It's not [X], it's [Y]"` |
| **Replace with**  | State Y directly                                  |
| **Default scope** | Everywhere                                        |
| **Default state** | On                                                |

The "it's not X, it's Y" pattern is a common AI rhetorical move that adds weight without adding meaning. This rule strips the setup and keeps the claim.

### Framing B ("In today's fast-paced world")

|                   |                                     |
| ----------------- | ----------------------------------- |
| **Type**          | Framing                             |
| **Blocked**       | `"In today's fast-paced world"`     |
| **Replace with**  | Cut entirely — start with the point |
| **Default scope** | Suggestions only                    |
| **Default state** | Off                                 |

This phrase is an archetypal AI opener. The rule removes it and leaves the rest of the sentence intact. It's off by default because the scope is narrow; turn it on if you find it appearing in your drafts.

## Scope options

Each rule (built-in or custom) has a scope that controls where it applies.

| Scope           | Where the rule runs                      |
| --------------- | ---------------------------------------- |
| **Everywhere**  | Both inline Tab suggestions and rewrites |
| **Suggestions** | Inline Tab suggestions only              |
| **Rewrites**    | Rewrites only                            |

You can set a different scope for each rule independently. For example, you might block em dashes everywhere but limit the framing rules to suggestions only.

## Adding custom guardrails

You can define your own rules in the **Add a guardrail** form inside AI preferences.

<Steps>
  <Step title="Open AI preferences">
    Click your user menu and select **AI preferences**, then scroll to the **Blacklist** section.
  </Step>

  <Step title="Choose a type">
    Select **Syntax** for literal text replacements (a specific word, phrase, or punctuation character). Select **Framing** for rhetorical patterns or sentence-level constructions.
  </Step>

  <Step title="Enter the blocked text">
    Type the exact text you never want to appear in the **Never say…** field. This is matched case-insensitively against AI output.
  </Step>

  <Step title="Optionally add a replacement">
    Type a preferred alternative in the **Say this instead** field. If you leave it empty, Dear removes the blocked text and cleans up surrounding whitespace.
  </Step>

  <Step title="Set the scope">
    Choose **Everywhere**, **Suggestions**, or **Rewrites** from the **Apply to** dropdown.
  </Step>

  <Step title="Save the rule">
    Click **Add guardrail**. The rule appears in the table immediately and is on by default.
  </Step>
</Steps>

<Note>
  Guardrails run after generation, not before. They rewrite the model's output rather than steering the prompt. This means very long or complex patterns may not be caught in every case — keep rules concise and specific for best results.
</Note>

## Managing rules

Each rule in the blacklist table has:

* **On/off toggle** — disable a rule without deleting it.
* **Scope selector** — change the scope at any time.
* **Delete button (×)** — remove the rule permanently.

The built-in rules (dash, framing A, framing B) can be toggled and re-scoped but not deleted.

## Enabling and disabling all guardrails

The master **Blacklist** toggle at the top of the section turns all guardrails on or off at once. When disabled, no rules run — neither built-in nor custom. Individual rule states are preserved so re-enabling restores your previous configuration.

## Related

* [Inline completions](/ai/inline-completions) — where guardrails run for Tab suggestions
* [Voice blending](/ai/voice-blending) — control the style of AI output alongside guardrails