Amazon AI-assisted OA: how to prepare for the repository task

Candidates describe it as the second question of the coding challenge. You get a small existing application with failing unit or integration tests, and you fix bugs or implement missing pieces until the tests pass. An AI assistant is built into the assessment environment. Several candidates were not given this format and received two conventional questions instead.

Choose the stack whose request flow you can already trace, because you cannot switch once the task starts. If you know none of them, candidates suggest Django as the quickest to learn from zero. Learn how a URL reaches a view or controller, how the model saves to the database and how a test calls an endpoint. One candidate reported that the C++ option looked different in the practice assessment, closer to building something without provided tests, so try each option in the practice assessment before deciding.

• Django: URL patterns, views, models and the ORM, serializers or forms, and the test client.
• Node.js: Express routes, middleware, a data layer such as a MongoDB model, and the test runner.
• Spring Boot: controllers, services, repositories, dependency injection and JUnit tests.

Usually not, according to candidate reports. Many describe an assistant that explains code, narrows down where a bug is and helps with syntax, but refuses to write the fix. That matches the guarded mode HackerRank documents for its assistant, which helps with syntax and concepts without providing complete solutions. Amazon has not published which mode or models its assessment uses.

Plan to write the change yourself. Use the assistant to summarize an unfamiliar file, explain an error or confirm what a function is supposed to do, and ask narrow questions rather than pasting the whole task. Do not use outside AI tools unless your instructions allow them; Amazon's student assessment page says your code must be your own and that browser use is logged.

Candidates who ran out of time say the biggest time sink was reading the whole codebase. Let the tests tell you where to look instead.

• Read the README and the names of the failing tests first. They usually point to the endpoint or model involved.
• Run the tests and read the first failure: the expected and actual status code, response body or database state.
• Trace that request from route to handler to model, and ask the assistant to explain any part you don't recognize.
• Make the smallest change that fixes the failing behavior, then rerun all tests, not only the one you targeted.
• Move to the next failing test. Leave a few minutes at the end to rerun everything and submit.

Practice on a small web project you did not write. Break one handler at a time, such as removing a save call, returning the wrong status code or skipping a permission check, and fix it from the failing test within 15 minutes. Then practice on a real open-source bug where you also need to find the relevant files. Set your assistant to explain rather than edit for some sessions, because that is closer to the reported OA assistant.

Keep the algorithm half of the challenge in your plan. Candidates report medium problems on graphs, binary search, dynamic programming and array partitioning. Rehearse those without AI unless your instructions say otherwise.

PraxisAI's challenges are real bugs from open-source Python projects, many of them in Django, with failing tests that decide whether your fix is correct. They are a way to practice reading an unfamiliar repository and fixing a bug from its tests under a timer. PraxisAI's agent can write code, unlike the assistant candidates describe, so ask it to explain and locate rather than edit if you want to rehearse the OA conditions. PraxisAI is not affiliated with Amazon or HackerRank and does not reproduce Amazon's assessment.