What problem does it solve? Research software work often fails because changes are made without matching the process to the risk: trivial edits get buried in ceremony while risky behavioral changes ship untested, unverified, or entangled with unrelated work in a dirty checkout. This Skill matches the workflow to each change's risk, size, and isolation needs so evidence, reproducibility, and user intent are preserved. ## Core Features & Use Cases - Risk-proportional routing: Classifies tasks from read-only inspection to publication-facing deliverables and prescribes the appropriate level of testing, isolation, and review. - Git worktree cutoff rules: Decides when an isolated worktree is required versus when working in place is safe, with machine-checked decision cases and Worktrunk or Git fallback commands. - Evidence-based completion claims: Requires running the proving command after the final change and reporting observed output before claiming work is complete, fixed, or passing. - Use Case: You need to fix a failing test in an R analysis repository while another agent session is active. The Skill directs you to create an isolated worktree, write a regression test, fix the failure at its source, rerun the reproducer, and only then report the verified result. ## Quick Start Ask the AI to fix a bug in your research codebase and let it decide the worktree, testing, and verification steps appropriate to the change.