Scaffold

A scaffold in AI prompting is a structured template that guides models through complex reasoning by decomposing problems into explicit steps. Rather than asking a model to solve a problem directly, scaffolding provides the skeleton of a solution process: first identify the key variables, then establish relationships between them, then apply the appropriate formula, then verify the result. The model fills in each section, constrained by the structure to follow a logical progression. Scaffolding works because language models are better at completing patterns than inventing them from scratch. When given a clear framework, the model's outputs at each step are more focused and accurate. The cumulative effect of high-quality intermediate steps produces better final answers than unstructured generation. Common scaffolding patterns include question decomposition, where complex questions are broken into simpler subquestions answered sequentially; role-based scaffolds, where different perspectives analyze a problem; and verification scaffolds, where the model checks its own work against criteria. Scaffolding differs from chain-of-thought prompting in its prescriptiveness. Chain-of-thought asks the model to think step by step, while scaffolding specifies what those steps should be. Advanced AI agents use dynamic scaffolding, selecting and adapting structures based on problem type.