Skip to main content
Instructions policies answer “What should happen?” You describe the required action or behavior in plain language. Limits uses that description to evaluate each request and return allow, block, or escalate.

​
What instructions are for

  • Natural language — No need to write structured rules. Describe the policy in prose or bullets.
  • Flexible interpretation — The system interprets your intent (e.g. “Block requests over $10,000” or “Escalate when the user is high risk”).
  • Same three outcomes — Every request still gets a clear allow, block, or escalate decision.
Use instructions when you prefer to specify behavior in words rather than in field/operator form, or when the logic is easier to describe than to express as conditions.

​
How instructions work

  1. In the policy editor, you write instructions in the Instructions tab (e.g. bullets or short paragraphs).
  2. For each request, Limits evaluates the payload and your instructions together.
  3. The result is still allow, block, or escalate with a reason.
Instructions are evaluated per request; the system does not expose a separate “condition list” — the natural language is the policy.

​
Writing good instructions

  • Be explicit about outcomes: “Block any payment over $5,000.” “Escalate when risk level is high.”
  • Mention key fields if they matter: amount, user risk, resource type, etc.
  • Default behavior: State what happens when nothing special applies (e.g. “Otherwise allow.”).
Example set of instructions:
  • Block any request where the amount exceeds $10,000.
  • Escalate to review when the user’s risk level is high.
  • Allow all other requests and log them.

​
Example

Policy in Instructions mode:
  • “Block any payment over $5,000.”
  • “Escalate to review when risk level is high.”
  • “Otherwise allow and log the request.”
Request { amount: 100 } → Allow.
Request { amount: 6000 } → Block.
Request { amount: 100, user: { risk_level: 'high' } } → Escalate.

​
Evaluating instructions (SDK and API)

  • SDK: Use limits.evaluate(policyKeyOrTag, prompt, response) where prompt is the user prompt and response is the LLM output (or input to validate). See SDK Policies.
  • API: POST /api/policies/{policyKey}/evaluate/instructions with body { "request": { "prompt": "...", "input": "..." } }. See API Reference.
You can use a policy key or a tag. With a tag, the strictest result across all matching policies wins: Block → Escalate → Allow.

Conditions

Define policies with structured rules instead.

Guardrails

Content safety, PII, and moderation policies.

Platform: Policies

Create and edit policies in the UI.

SDK: evaluate()

Evaluate instructions from your app.