Senior Software Engineer — Author High-Difficulty AI Coding Tasks (Claude Code / Docker)

About the project We're building a benchmark of realistic, high-difficulty software engineering tasks used to evaluate frontier coding models. You'll act as the task author: pick a real open-source codebase, write the opening prompt, calibrate difficulty against a strong model, validate the task in an isolated environment, and submit it for review. This is not a "fix this bug for a client" job — the deliverable is the task itself, plus evidence of how a top model performed on it. Each task takes roughly 2 hours, of which no more than 1–1.5 hours should be spent inside Claude Code. What you'll do Claim a task slot from our portal (each slot pre-seeds task type, domain, language, codebase type, prompt style, and interaction mode). Choose a real, runnable application (not a library, tutorial, or SDK), pin it to a specific commit, and confirm it meets the slot's codebase requirements (enterprise-scale or personal-scale, with a working test suite). Write the opening prompt in the assigned style — Casual Prose (20–60 words), Explicit Spec (450–1,200 words), or Error/Artifact Brief (real reproducible error + one short line). Run a difficulty check in a fresh Docker lane using Claude Opus 5.5 at High effort. If the model solves it cleanly in one shot, discard and pick a harder problem. Write the private high-level goal and 2–3 repo-specific decision points, then validate the task in a second fresh lane. Collect the conversation log and final diff and attach them in the portal. Log discarded one-shot runs in the Discarded Run Log. Must-have Strong software engineering background (5+ years; senior-level judgment expected). Comfort reading unfamiliar codebases quickly and identifying realistic, bounded engineering problems. Experience with Docker and command-line tooling; able to set up and run projects from a repo or Compose file. Working knowledge of at least one mainstream stack (Python/JS/TS/Go/Java, etc.) and its testing conventions. Excellent written English — prompts must read like a real colleague's Slack message or a professional ticket. Discipline with process: pinning commits, fresh lanes for every run, honest reporting of results. Nice-to-have Prior experience writing evals, benchmark tasks, or AI training data. Familiarity with Claude Code (or similar agentic coding tools) and model-effort settings. Experience with CI, infrastructure, or operational debugging (for "Operate"-type tasks). What we provide Access to our task portal (task slots, filters, Trainer Card, submission and AutoQC). Official setup bundle, Docker lane image, checksum-verified downloads, and a prework.py helper for setup and output collection. Claude Max or Ultra subscription reimbursed (claimed via a linked form). Pro is not sufficient for active work. Clear written onboarding covering workflow, task types, codebase rules, prompt styles, calibration, and scoring. Compensation Hourly rate (please quote yours). Performance bonuses per approved task, tiered by approval date To apply Please answer briefly: Your strongest stack(s) and years of experience. One example of a real codebase you know well enough to design a hard, bounded task around (repo + why). Your experience with Docker and agentic coding tools (Claude Code, Codex, etc.). Your weekly availability and target hourly rate. Confirm you can commit to ~2 hours per task and are comfortable working in English.