What Is an Instruction?
An instruction is a rule you give an agent: in this scenario, do this. Categories cover the common cases out of the box. Instructions let you handle the ones that are specific to your organisation, such as always blocking a project codename, or allowing a scenario your team has already reviewed. Instructions are forward-looking. Once you add one, the agent applies it to every trace from then on.What an Instruction Is Made Of
There is no separate “Escalate” action. Notification is the high-criticality setting, not a third action.
Triggers
The trigger decides when an instruction applies. Pick the approach that fits the rule.- Semantic
- Regex
- String match
Matches a scenario by meaning, even when the wording differs. Describe the situation in plain language.Example: “Customer is discussing their account balance or transaction history.”
Create an Instruction
Open Instructions
Go to Agents → Instructions and start a new instruction.
Describe the scenario
Name it, pick a category, and choose the trigger that matches the rule.
Set action and criticality
Choose Allow or Block, then Low or High criticality.
Assign it to an agent
Assign the instruction to the agent that should use it. Instructions appear grouped by category on the agent.
Save an Instruction From an Assessment
You do not have to write every instruction from scratch. When you review an assessment and decide how a scenario should be treated in future, you can save it as an instruction directly from that decision. The new instruction keeps a link back to the assessment it came from, so you always know where it originated.Test Before You Save
Open the testing environment from the flask icon in the top navigation. You can test a single instruction in isolation, or dry-run a Data Protector configuration (role, categories, and instructions) against a sample prompt. Nothing you test touches live traffic. See Configuring agents.Assessments
Review verdicts and save instructions from them.
Data Protector
See how instructions sit alongside categories.