You already know: every technical layer of an AI system (D1–D20); evaluating claims and running experiments (D20); cost models for training and serving (D12, D15, D17).
Why this matters for Atlas. Staff engineers are judged on outcomes across teams, not on their own commits. The core skills are: writing a technical strategy that turns goals into a few guiding choices; choosing problems with explicit value, confidence and cost, as a portfolio rather than one bet; running multi-team technical work (milestones, critical path, risk burn-down); and growing other engineers so the work scales beyond you. None of this is mystical. Each has a concrete practice, and this lesson gives them with small tools that make the reasoning explicit.
1. Technical strategy: diagnosis, guiding policy, coherent actions
The question. What is a technical strategy, concretely?
Start from nothing. A useful structure (Rumelt's kernel): a diagnosis of the situation (what's actually limiting us, with evidence), a guiding policy (the few choices that address it, including what we will not do), and coherent actions (concrete work that implements the policy and reinforces itself). A strategy that lists every team's wish is not a strategy; one that names trade-offs is. For Atlas: "Diagnosis: answer quality is limited by retrieval on lecture notes (60% of wrong answers trace to missed passages), and cost per active student is growing 8% a month. Guiding policy: fix retrieval before changing models; no new modalities this half; every change must pass the shared eval gate. Actions: hybrid retrieval with contextual chunks, a per-course eval set, routing easy questions to a smaller model."