A creed is a structured set of principles that governs how an AI system behaves. Written in natural language. Parsed by machines. Portable across models and providers.
You already do something like this when you write a system prompt. You tell the model to be helpful, avoid harmful content, stay on topic. This works until it doesn't. System prompts are brittle: they can be overridden by clever inputs. They're invisible to users. They don't compose: running two AI agents together means their system prompts collide with no arbiter.
A creed is what a system prompt wants to be when it grows up.
What a creed looks like
Concretely, a creed is a structured document containing principles organized by category. Each principle has a name, a description, and a weight. Safety-critical principles can be marked mandatory. Stylistic preferences can be marked advisory. The creed travels with the AI: when it talks to another AI, when it gets exported to a different platform, when it gets audited.
This matters because AI is leaving the chatbox. Agents make purchases, send emails, negotiate contracts. When an agent acts on your behalf, you need to know what governs its behavior. When a business deploys AI that talks to customers, regulators need to inspect its operating principles. When two agents transact, they need a shared understanding of what's acceptable.
A system prompt is a whisper in the model's ear. A creed is a published commitment.
How creeds work at Creed Space
Author. Build a creed using the visual editor, or fork one from the community hub. Each principle can carry a candidate rule for the Policy Decision Point (PDP); principles without one are guidance.
Enforce. AI actions routed through Creed Space are checked against active creeds. Latency targets are under 10ms for known patterns; novel cases get a fuller review and take longer. Targets are local benchmark goals, not guaranteed timings.
Audit. Decisions are logged with the principle that applied and the reasoning given.
Export. Export your creed for ChatGPT, Claude, or any platform that implements VCP.
Why this is different from a blocklist
A blocklist says "don't say these words." A creed says "here's what you stand for." The difference is architectural. A blocklist catches specific outputs. A creed shapes reasoning. A blocklist is brittle against adversarial inputs. A creed gives the AI an understanding of why certain actions are appropriate or inappropriate, which generalizes to novel situations the blocklist never anticipated.
You probably already have values you want your AI to follow. A creed is how you make those values enforceable, portable, and auditable.