An audit tool for AI agent prompts.

Any model can write a plausible agent prompt. Very few declare the things a production agent actually needs — where its tools stop, what it must never do, and when a human has to approve an action. CEPrompts reads a prompt and checks for exactly those things, then gives it a CE Score and a trust level.

Why it exists

Prompts that look fine in a chat window fail differently once an agent can act — call tools, write files, spend budget, take irreversible steps. The common failures aren't about clever wording; they're missing boundaries, missing guardrails, missing approval gates. CEPrompts makes those gaps visible before a prompt ships, instead of after an incident.

How it grades

Every prompt is checked against ten components: role, context, constraints, tool boundaries, memory policy, output schema, validation, safety, approval gates, and examples. The result maps to an AgentAz trust level — from ADV (advisory only) up to A5 (fully governed). A prompt only reaches a level when every check that level requires passes, so the grade is earned, not estimated.

The current version runs a structural check in your browser — it looks for whether each component is present. A semantic version, which evaluates whether each component is actually sound rather than just present, is in development.

How it fits AgentKits

CEPrompts grades against the same AgentAz standard used by AgentKits, a library of governed agent blueprints that already ship at higher trust levels. When an audit shows a prompt falling short, the kits are a place to start from something that already passes.

Contact

Questions, feedback, or found a check that's wrong? Email hello@ceprompts.com.