Prompt Guard overview

AI safety — guardrails, evals, and prompt packs before customer-facing AI ships.

Note

Creytix Prompt Guard is part of the same discipline that keeps Creytix Inspector honest — a gate that has to pass before AI-generated work reaches a customer-facing surface.

Overview

Creytix Prompt Guard is the check that runs on a prompt before it's used to generate anything a customer sees — copy, images, video, or bulk content. Instead of trusting whoever wrote the prompt to remember every rule, Prompt Guard scans it against a fixed set of guardrail policies, tracks whether prompt quality holds up over time with an eval loop, and logs what every generation actually cost.

Because it sits in front of batch generation, a bad prompt gets caught before it turns into a wasted render, a leaked credential, or a customer-facing asset that doesn't match brand — not after.

How it works

Prompt Guard runs in three layers:

Guardrail scan — a deterministic, effectively free check runs on every prompt before it's used for batch generation, catching things like leaked secrets, disallowed model defaults, or content bleeding across brands.
Eval loop — a golden set of known-good prompts is re-run whenever a prompt or rule changes, and a smaller sample gets deeper (model or human) review on a set cadence.
Cost log — every generation records the surface, model, and token spend that produced it, so batch AI work never runs blind on cost.

Steps

•
Open Prompt Guard
From the Creytix Control Panel, open Prompt Guard to see the current status of your prompt packs.
•
Run a guardrail scan
Point a prompt or prompt pack at the scan before running a batch generation job.
•
Review fails and warns
Hard fails block the batch outright; warns surface for a human to review before you proceed.
•
Check eval history and cost
Review how a prompt has performed over time and what its generations have cost by surface and model.

Capabilities

Guardrail scan — deterministic checks catch secrets, unsafe or banned patterns, and structural problems before a prompt reaches generation.
Eval loop — golden test prompts are re-run to catch regressions whenever a prompt or rule changes.
Cost logging — every generation is tied back to the surface, model, and token spend behind it.
Prompt packs — approved, reusable prompt templates for teams shipping AI content on a regular cadence.

Limits & honest scope

Prompt Guard checks the prompt and the pipeline around it — it does not judge whether the generated output itself is good; that's still a human call, or a job for Creytix Inspector on the site-quality side.

Judge-based (LLM) evals are sampled, not run on every single generation — full review still happens on a schedule, not in real time.
It's built for content moving through Creytix's own generation pipelines, not a general-purpose prompt-injection firewall for arbitrary third-party AI apps.
New guardrail categories get added as real mistakes happen; treat the ruleset as growing, not exhaustive on day one.