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
- [[references/stage-split-humans-set-constraints-models-multiply]] — split the pipeline by stage, not by product, with a monotonic human-to-model gradient.
- [[references/author-a-pool-deliver-an-instance]] — make the unit of authorship larger than the unit of delivery, so personalization costs a lookup instead of a headcount.
Timing and discipline
- [[references/adopt-early-gate-on-your-own-quality-bar]] — buy the integration years while quality improves on someone else's budget, but only against a bar you defined before the tool existed.
- [[references/simplify-until-automatable-name-what-must-survive]] — name the one thing the feature must keep doing, standardize the rest, and schedule the sameness you just bought.
Sibling skills
- [[../duo-product/SKILL]] — the long-view product judgment these bets sit inside.
- [[../duo-experimentation/SKILL]] — every quality claim here was an experiment or an efficacy study before it was a strategy.
- [[../duo-culture/SKILL]] — the talent-density math behind "fewer than 1,000 people, 21 million daily users."
- [[../duo-growth/SKILL]] — what a cost-structure advantage lets you do in the market once you have one.
Sources
blog.duolingo.com posts on the Duolingo English Test, LLM-assisted course content, expert-plus-AI pipelines, DuoRadio scaling, responsible-AI standards and company strategy (2020-2026), all cited with dates inside the nodes.