25 terms that name specific AI agent failure modes and the architectural patterns that prevent them. Each term is a concept you can act on, not jargon for its own sake. See how they interlock →
The diagnostic discipline of taking a real prompt from your own work, running a structured audit aga…
A binary policy specifying that certain operations require explicit unlock phrases before execution.…
The end-state framing of all the above: the surface of structured artifacts: system prompts, AGENTS.…
The structured set of bounds, rules, and tiered policies that shape a model's output to match intent…
The framing of a prompt as a problem in which the model searches for an output that satisfies every …
The discipline of allocating limited context tokens between different categories of information: sys…
The totality of inputs a model receives at the moment of generation: the system prompt, the conversa…
A structured extraction of a specification's intent, requirements, out-of-scope items, and open ques…
The direction and magnitude by which an AI system's outputs deviate across demographic groups when a…
An explicit map of upstream and downstream relationships for a unit of code, constructed before any …
A tiered policy specifying what the agent does at each level of scope ambiguity: refuse, narrow, con…
A constraint pinned to a structurally privileged position the model re-reads every turn: typically t…
An explicit statement of what "correct" means for a unit of work, written before any code or test, a…
An append-only artifact capturing decisions or context that must persist across sessions. Existing e…
The interface where a human's goal is translated into a model-readable specification. The point at w…
The high-dimensional mathematical manifold in which a language model represents concepts. Each model…
A model configured to evaluate other model output against a structured rubric. Used for semantic cri…
Explicit specification of what the model must not do. The "Guardrails" slot of the Three-Constraint …
The property that each step in a multi-step prompt sequence preserves the contracts established in e…
A five-level framework for assessing and advancing a practitioner's or organisation's prompt archite…
Structured adversarial testing against an enumerated category list: boundary inputs, malformed input…
The direction and magnitude by which a model's interpretation moves away from the user's intent over…
The principle that every well-formed prompt encodes three constraints: Intent (what you want the mod…
The ratio of constraint-bearing tokens to total tokens in a prompt. A high-density prompt encodes ma…
The full layered validation construct that converts agent output into a trusted artifact: Ground Tru…
