What the work actually is

You are not answering coding questions — you are authoring the problems. Each task starts from a real engineering situation in a large codebase: a regression that only surfaces under a specific config, a feature that touches five modules and a migration, a dependency upgrade that quietly changes behaviour. You write the prompt so that a coding agent has to investigate rather than pattern-match, then you build the verifier: a deterministic test suite that passes on a genuinely correct fix, fails on the shortcut, and doesn't flake. The whole thing ships as a Docker image that builds and runs the same way on someone else's machine.

The hard part is usually the verifier, not the prompt. Agents are good at finding the test and writing to it — deleting the failing assertion, special-casing the input, stubbing the broken path. A task that can be gamed is a task that gets rejected, so you're expected to think adversarially about your own grading logic and document what a passing solution must actually demonstrate.

What the platform screens for

  • A real codebase. The listing is explicit: you need a large, complex, private codebase you own and have the right to author tasks against. This is the most common disqualifier. Toy projects, tutorial repos, and code owned by a current employer don't clear it.
  • Docker fluency in practice. Not "I've used containers" — pinned base images, reproducible builds, deterministic test runs, sensible handling of network and time dependencies.
  • Architecture-level experience. You've maintained something substantial long enough to reason about how a change ripples across modules.
  • Hands-on agent use. You've run coding agents on long tasks and can describe where they fail, not just that they're impressive.
  • Technical writing. Prompts and review criteria are read by other engineers and by graders. Vague specs create unscoreable tasks.

Logistics

Fully remote and asynchronous. The qualifying assessment is unpaid and typically involves authoring a sample task end to end — prompt, container, tests — which is a meaningful time investment before any compensation. Passing enters an ongoing pool; project volume varies, and rates within the $70–100/hr band are set per project, not guaranteed. Most contributors work in self-directed blocks rather than fixed hours, with review cycles and revision requests handled over the platform.