Duolingo AI Product Strategy — Map of Content

Most AI roadmaps are lists of features. Duolingo's public record reads differently — the model gets pointed at whatever makes the incumbent expensive, at whichever internal constraint gates everything else, and at the stages of a pipeline where errors are local and cheap. This skill packages that judgment for products that have nothing to do with language learning.

This is a graph. Scan the descriptions, follow only the [[wikilinks]] you need.

Where this skill stops: it answers "build it, here, for this reason." Making the model behave once built — prompting, decomposition, memory, evals, verification, agent permissions — belongs to the sibling skill duo-llm-feature-engineering.

Where to aim

• [[references/attack-the-incumbents-physical-bottleneck]] — when a competitor's guarantee rests on scarcity, generation dissolves their whole cost structure, and you must re-supply the guarantee you deleted.
• [[references/point-ai-at-unit-economics-not-features]] — steps, authoring and evaluation are the three cost lines worth attacking; a feature is copied in a quarter, a cost structure is not.
• [[references/spend-the-automation-budget-on-the-gate]] — find the one constraint everything queues behind, and justify the spend by the work it unlocks rather than by throughput.

Getting the problem right before you build

• [[references/name-the-real-ml-problem-first]] — your objective is often not the one the off-the-shelf model was trained for; state it as a scoring function first.

Dividing the labor