AI Coding — Maze Solver

Requirements

Starter codebase: a maze with S start, E end, # walls, . open cells. The interviewer hands you a small Python project with solver.py, a print helper, and a tests/ folder. You progress through 4-5 sub-tasks.

• Q1 — Bug fix, AI typically disallowed. Either the path symbol overwrites the S/E markers in the printed output, or the diff-style output is malformed. Fix by giving the start/end glyphs display priority in the print logic.
• Q2 — BFS / DFS infinite-loop bug. Search loops forever because a visited set is missing. Add dedupe before enqueue.
• Q3 — Directional gates. Cells marked > or < force the next move direction. Patch get_neighbors / move to consult the current cell glyph.
• Q4 — Keys and doors. Lowercase letters are keys; uppercase letters are doors. Expand state from (x, y) to (x, y, collected_keys_bitmask). Candidates may need to revisit cells once they hold new keys.
• Q5 (since April 2026) — Bombs. A bomb glyph destroys walls within a fixed radius (commonly 2-cell Chebyshev or a 1-cell + cross). Generate a get_affected_area(x, y) helper, fold the destroyed-walls set into search state, decide whether destruction is one-shot or persistent.
• Q5 alt — Energy / shortest weighted path. Cells carry energy costs; minimize total energy to reach the exit. Dijkstra is the expected answer. Some loops use this instead of bombs.

Trap to watch

One candidate reports the Q1 starter contains an intentionally invalid test case that must be commented out before the bug fix can be validated — the interviewer expects you to recognize the bad test, not patch the codebase to satisfy it. If 30+ min into Q1 the failing test still won't pass after seemingly correct fixes, ask whether any test cases themselves are wrong.

Examples

One candidate's run: S.....#...E style 8×10 grid; Q1 fixes the * path overwriting S; Q2 adds visited to BFS; Q3 forbids leftward moves on > cells; Q4 introduces a/A, b/B pairs; Q5 places one bomb that, when stepped on, clears a 5×5 wall block. Multiple threads report 7 hidden test cases with the bar set at passing the first 4.

Notes

• Since mid-March the AI assistant writes directly into the CoderPad editor (Cursor-style). Model picker exposes Claude Opus 4.6, GPT-5.x, and Sonnet — Opus 4.6 is the consensus pick for this prompt.
• The interviewer will ask you to explain whatever the AI wrote. Generating a working solution but failing the explanation is a common rejection path.
• Reaching Q4 is the typical "strong" bar; Q5 is reach-territory and was added in April 2026.
• Underspecified glyph rules: confirm with the interviewer whether bombs are triggered on entry, exit, or by adjacency before prompting the AI.

Preparation

• Drill the canonical Q1-Q4 ladder by hand first so you can dictate the algorithm to the AI rather than read what it produces blind.
• Practice on the public Hello Interview maze harness (hellointerview.com/practice/ai-coding) — same class layout, with a runnable AI panel.
• Rehearse the prompt loop: "decide algorithm yourself → break into small functions → prompt the AI per function → read + explain." Letting the AI drive end-to-end fails the explanation follow-up.
• For Q5, pre-write a mental template for radius-N blast geometry and bitmask state extension so you don't burn time deriving it live.