Full-Stack AI-Pairing Assessment

A five-round AI-pairing assessment delivered as the "Instacart Full-Stack Engineer Assessment". Two rounds make you elicit a feature spec by chatting with an AI agent; you then implement each spec with an in-editor Claude Code assistant in a provided full-stack repo (FastAPI backend) under full unit + integration test coverage. Feature families include a library/book search flow, a notification/hold system, and a quick backend filtering bug fix hidden behind a frontend symptom; the final implementation may also require administrative controls and observability. Backend-only applicants may still be routed here; the real bottleneck is AI-tool fluency and speed, not algorithms. Requirements Five timed rounds, delivered by email as the "Instacart Full-Stack Engineer Assessment". The invite title varies between candidates, and backend-only applicants may still receive the full-stack version. One completion window was two weeks from receipt. Round 1 — Requirement elicitation. Chat with an AI agent to extract the spec for a feature. The agent withholds detail, so clarifying thoroughly is part of the signal. One starting prompt: a user finds the frontend table too long; the elicited spec was two dropdowns plus a text box, though finer details (dropdown contents, labels) were hard to pull out. Another variant asks for a book/library search feature. Round 2 — Implementation. In a provided project directory, implement the round-1 feature using an in-editor Claude Code assistant. Copy-pasting the round-1 AI conversation is blocked. Full unit + integration test coverage is required; submitting runs the entire suite. Round 3 — Bug fix. A frontend page renders a wrong value or unfiltered result set; one fix was a small reordering, while another was in the FastAPI backend path before results were returned. Expect to trace the symptom across frontend JavaScript and Python backend code. Round 4 — Second elicitation. A harder spec, again drawn out by chatting with the AI agent: either two features spanning frontend and backend, with idempotency guarantees and full test coverage, or a notification feature whose exact stack and behavior must be obtained from the AI product manager. Round 5 — Implementation of the round-4 features under tight time (one candidate had only ~30 minutes left). Round-2 code does not carry into round 5. Notes Example round-4/5 task A library-style notification + hold system: a patron borrowing an item receives two notifications, one 24 hours before the due time and one at the due time. A patron can place a hold on an item; multiple patrons can hold the same item simultaneously. When a held item is returned and becomes available, notify the holding patrons in FIFO order, serving the first N holders. The logic is intentionally tangled, and the exact rules must be elicited from the AI agent rather than read from a written spec. Some versions expand the final implementation with administrative controls and observability requirements. Logistics & gotchas The whole assessment rewards AI-pairing fluency and typing speed more than algorithmic depth. Round-2 work is not preserved into round 5; budget time accordingly. Commit after each finished feature; CodeSignal can lose code during the assessment. Terminal error output cannot be copy-pasted back into the AI assistant, which slows debugging. The Claude Code assistant runs slowly, and running the full test suite on submit can freeze the page. Running two AI agents in parallel against the same project directory caused them to edit the same files and break the project; keep agents scoped to separate areas. Preparation Practice eliciting a precise spec from an AI assistant: ask for inputs, outputs, edge cases, and idempotency expectations explicitly rather than accepting the first answer. Build a small full-stack feature (FastAPI backend + simple frontend) end to end with an AI editor, writing unit and integration tests as you go, on a timer. Rehearse a FIFO hold/notification or scheduled-reminder feature so the round-4/5 family is muscle memory.