The 10 skills real AI-engineering intern roles (like Clinikally YC S22's) actually ask for. Drag to orbit · tap a star to inspect · mark what you know.
Each skill carries a weight reflecting how often it appears in real AI-engineering intern postings (agents 15%, LLM APIs 13%, Python 12%, RAG 12%, evals 11%, multi-agent 9%, prompting 8%, deployment 8%, vector DBs 7%, Git/SWE 5%). Your score is simply the sum of the weights you have marked:
readiness = Σ weight(skill) for every skill you know
The "learn next" suggestion is a greedy pick: the highest-weight skill you don't have yet whose prerequisite edges in the constellation are all satisfied. That mirrors how learning actually compounds — RAG before you know embeddings is memorizing recipes; agents before tool-calling is cargo-culting frameworks.
When a YC startup like Clinikally (S22) writes "build AI agents and multi-agent systems, LLM applications," the day-to-day work is rarely training models. It is: wiring frontier-model APIs into product features, designing tool schemas and prompts, grounding answers in company data with RAG, and — the part most candidates skip — writing evals so quality is measurable. Hiring managers consistently rank a shipped public project (an agent with traces and an eval suite) above coursework.
1) Depth beats breadth: 70% on this tree with one polished project outperforms 100% checkbox knowledge. 2) Read the model provider docs, not just framework docs — frameworks churn, APIs endure. 3) Cost awareness is a skill: interns who can say "this agent run costs $0.04 and here's why" stand out immediately.
The constellation models ten core AI engineering skills as a dependency graph where foundational techniques unlock downstream capabilities. Marking skills you possess turns their stars green and increments your total readiness percentage by each node's empirical posting weight. The explorer dynamically inspects unlearned nodes, greedily surfacing the highest-impact skill whose prerequisite path is fully unlocked.