Task success criteria definition

engineering

refinement is a Claude Code skill that locks the definition of success into human-signed, verifiable assertions before any implementation begins. It acts as Gate 1 in the driving-work-to-done framework: once a task is deemed worth tracking, refinement freezes an assertion block inside `{issue}/index.md` via a git commit — after that, implementation details can change freely, and only a wrong success definition warrants returning here. Each issue lives under `issues/{namespace}/{issue}/` with `index.md`, `loop-state.json`, and `measurement-ledger.json`; the `open-seed-issue.sh` script lets you capture an unplanned finding as a seed issue without signing assertions yet. Before sealing, `check-plan-answers.sh` validates that all plan fields (`what`, `when`, `why`, `how`, `environments`) are filled with sourced answers — `human`, `environment`, or `inferred_confirmed`. External outreach is tracked with `record-outreach.sh` to ensure every message sent has a recorded draft and a human sign-off. Assertions must include both positive and negative lists with explicit input conditions («when X, Y happens» rather than «should have X»), preventing the scenario where all assertions pass green