AI Workflow Patterns

Key jargon

Term	Plain-language meaning
Prompt chain	A sequence where one model output becomes a later step's input.
Routing	Classifying a request and sending it to a specialized path.
Parallelization	Running independent subtasks concurrently.
Evaluator-optimizer	A loop in which one component critiques and another revises against a stopping rule.

Key concepts

Patterns can be combined, but every added branch increases testing and recovery complexity.
Use explicit intermediate artifacts and stop conditions to keep orchestration inspectable.

Concept map

flowchart LR
    A["Decompose job"] --> B["Select chain route or parallel pattern"]
    B --> C["Evaluate intermediate artifacts"]
    C --> D["Join and deliver"]

Pattern	Use	Main risk
Prompt chain	Decompose staged transformations	Early errors propagate
Router	Select specialist/model/tool	Misrouting and hidden fallback
Parallel fan-out	Independent research or candidate generation	Duplicate work and costly synthesis
Evaluator–optimizer	Draft and revise against a rubric	Evaluator shares the same blind spots
Orchestrator–workers	Dynamic task decomposition	Unbounded work and conflicting outputs
Map–reduce	Process many documents then aggregate	Lost cross-document relationships
Human gate	Approve risky transition	Approval fatigue or vague preview

Design method

For each node define input schema, output schema, owner, model/tool, timeout, retry rule, evidence, and transition condition.

Dumpster lesson

The prompt-chaining research intake found a useful principle: structured intermediate artifacts make failures localizable. Treat the paper’s reported scores as author claims, not universal proof.

Exercise

Design a three-stage research digest: cluster evidence, synthesize each cluster, then reconcile claims. Specify how a failed cluster is represented.