The vocabulary
The working vocabulary of this site. Some terms are names given here to failure patterns that kept showing up without one; others, like context engineering, are wider industry terms used here in a specific sense. Each links to where it was first evidenced, not just defined in the abstract.
Context design
Understanding the customers, business reality, technology and operational history surrounding a brief, before deciding what to build. The output can be an interface, a process change, an automated workflow or a content system; the discipline is the same regardless of what it produces.
First appeared: Practised across a decade of retail UX and digital transformation work, named explicitly once AI made the underlying skill visible as its own discipline.
Context debt
The accumulated backlog of decisions, exceptions and tribal knowledge a company never wrote down, because a person could always be asked instead. Invisible for as long as humans were doing the work, since a person facing a gap pauses and asks. A model facing the same gap doesn't pause; it fills the gap with something plausible and moves on.
First appeared: Named after repeatedly seeing the same undocumented-exception pattern surface across unrelated AI rollouts.
Context Debt · How Context Debt shows up: five patterns · The Context Debt Map · The debt came due
The Context Stack
The seven layers business context is actually made of: facts, history, rules, precedence, exceptions, boundaries, and verdict shape. Built by reading back through the site's own demos for what recurs, not assembled as a framework first and illustrated after.
First appeared: Derived from what was already load-bearing across the five original context-brief demos.
Batch blindness
A failure that only exists at the level of a set of outputs, not any single one of them. Every item in a batch can be generated correctly relative to its own brief and the batch can still read as wrong once viewed together, because nothing in a single-item generation loop is positioned to see the set.
First appeared: Surfaced from a real production batch of 51 category lifestyle images, generated one category at a time with a deliberately thin brief.
Fidelity drift
The opposite shape of fault to batch blindness: a problem that exists inside a single output in isolation, invisible unless someone checks that specific item's fine detail against what it's meant to represent. Synthesis isn't reproduction; broad shape survives regeneration reliably, fine printed or licensed detail doesn't.
First appeared: Surfaced from the same 51-image batch that surfaced batch blindness, at a different stage of the same pipeline.
Context engineering
The infrastructure layer: retrieval, memory, tool schemas, the pipes that get material in front of a model at all. Context design is the judgement layer sitting on top of it: deciding what that material should actually contain. The two are complementary, not competing, and most AI failures blamed on one are really a gap in the other.
First appeared: Distinguished explicitly once the two terms started being used interchangeably in the wider industry.