Guardrails

Evaluate prompts against guardrail policies before batch generation.

Note

Guardrails is the first layer of Creytix Prompt Guard — the pass/fail check that runs before a prompt is allowed into batch generation.

Overview

Before Creytix runs a batch of AI generation — a set of image prompts, a content pack, a rewrite job — Guardrails scans every prompt against a fixed set of policies. It's built to catch the mistakes that actually happen: a credential typed into a tracked prompt, a job quietly defaulting to the wrong model, or one brand's reference material leaking into another brand's generation run.

Catching these before generation runs means a bad prompt costs you a failed scan, not a wasted render or a customer-facing asset you have to pull back.

How it works

Every prompt scanned by Guardrails is checked across a set of categories, and the result lands in one of two buckets:

Fail — a hard stop. Plaintext secrets, banned model defaults, and missing required guardrails block the batch from running at all.
Warn — a flag for human review. Unstructured prompts, missing brand-isolation footers, or other soft issues surface for a person to check before the batch ships.
Re-scan — once flagged items are fixed, the same prompt or pack is re-run through the scan before it's cleared.

Steps

•
Point Guardrails at a prompt pack
Select the prompt or prompt pack you're about to use for a batch generation job.
•
Run the scan
Guardrails checks it against the current policy set — secrets, model defaults, cross-brand isolation, and structure.
•
Review fails vs. warns
Fails block the batch outright; warns are flagged for someone on your team to look at before proceeding.
•
Fix and re-run
Update the prompt and re-scan before the batch goes out, rather than shipping around a flagged issue.

Capabilities

Secret detection — catches API keys, tokens, and connection strings that shouldn't be in a tracked prompt.
Model-default checks — flags prompts calling for a model outside your approved defaults.
Cross-brand isolation — stops one brand's reference assets or copy from leaking into another brand's generation job.
Structure checks — flags unstructured, hard-to-review "wall of text" prompts.
Fail vs. warn — hard fails block a batch; warns surface for human review instead of stopping everything.

Limits & honest scope

Guardrails is a policy scanner, not a content critic — it checks whether the prompt going in is safe and well-formed, not whether the output that comes out is any good.