Guardrail Studio

Guardrail Studio is where you design a guardrail for an AI application, test it on real prompts, measure how well it performs, and keep improving it. It runs inside the WitnessAI console.

Opening the Studio

Open the console's app switcher and choose Guardrail Studio (Design and test guardrails) under Applications. The Studio opens at /studio.

You need to be signed in to the console. If you don't see the entry, Guardrail Studio hasn't been enabled for your organization.

To return to the rest of the console, use Back to Console in the Studio's left rail.

The screen at a glance

Area	What it does
Left rail	Studio and Back to Console. At the bottom: a status indicator (click it for system health), a dark/light toggle, and your initials.
Guardrails sidebar	Lists your guardrails as cards showing the name and {n} rules · {m} intentions. Templates opens the template library, and + creates a new guardrail. Drag the sidebar's right edge to resize it.
Main pane	The selected guardrail, with three tabs: Editor, Evaluate and Diagram. The Studio remembers the last tab you used.
Ask AI	A floating assistant, available everywhere in the Studio, that can create, change and test guardrails for you. See Guardrail Studio: Ask AI.

The main pane's header shows the guardrail's name. Click the name or the pencil icon to rename it. On the right of the header are View Payload (</>), Download config JSON, Save as Template, and the Enable switch.

Your first visit

With no guardrails yet, the sidebar shows No Guardrails Yet and the main pane shows Welcome to Guardrails with five ways to start:

Option	What it does
Guided Setup	A step-by-step wizard, or describe what you want in plain English
From Template	Start from a pre-configured guardrail for a common use case
Blank Guardrail	Start from scratch
Upload JSON	Import a guardrail from an exported JSON file
Paste JSON	Paste a guardrail JSON payload to create one

Each option is covered in

Guardrail Studio: Create a Guardrail.

Once you have guardrails, select one in the sidebar to open it. The Studio doesn't open one for you automatically.

Save what you want to keep. A guardrail is kept between visits only once it's saved. Guided Setup saves for you; everything else needs Save as Template. See

Guardrail Studio: Save, Share & Export.

How a guardrail works

Each guardrail performs classification, risk analysis and jailbreak detection in a single pass. The part you design is the classification: what your application is for, which kinds of request it should recognize, and what to do with each.

The parts of a guardrail

Part	What it is
Name	How the guardrail appears in the sidebar and in test results
Context	A description of the AI application and its domain, up to 1,000 characters. It helps the classifier understand domain-specific terms. For example: "Agricultural chatbot where poison and kill are benign pest control terms."
Rules	Each rule reads IF Intention is any of the following … THEN an action
Intentions	The kinds of request a rule recognizes. Each has a name and a definition that says what prompts should match it.
Enable	Turns the guardrail on or off. A new guardrail is enabled.

Rule actions

Action	On the rule	What happens
Allow	THEN Allow the user's prompt	The prompt passes.
Warn	THEN Warn the user with message: …	The prompt passes and is flagged for review. A warning message is required.
Block	THEN Block with message: …	The prompt is stopped. A custom block message is optional.
Route	THEN Route the user's prompt	The prompt is marked for a different handler. Route destinations aren't configured in the Studio, and in evaluations a Route match counts as allowed.

The default rule

Below your rules, For all other intentions is fixed to Allow. A prompt that matches none of your intentions is allowed. To stop something, give it an intention in a Block rule.

Built-in protections

Jailbreak detection and risk analysis run alongside your rules, and a guardrail checks both the user's prompt (input) and the model's reply (output). These are on for every new guardrail.

There's no switch for them in the editor. A guardrail created from a template, by Ask AI, or from an imported JSON file can carry different settings. See

Guardrail Studio: Save, Share & Export.

Guardrail Studio: Overview Opening the Studio The screen at a glance Your first visit How a guardrail works The parts of a guardrail Rule actions The default rule Built-in protections