skill-forge
Drive one of YOUR skills (a .claude/skills/<name>/SKILL.md plus bundled scripts) toward a measurable bar: loop generate → exercise → judge until a judge-scored eval gate passes (fail-closed), then stage a PR for a human to merge. The eval-driven cousin of one-shot scaffold-sky-workflow. Runs against your repo; all mutation in a throwaway worktree. v1 does not execute model-generated scripts (static bash -n / node --check only). Bounded by max_budget_usd + loop.max; idempotent. Fire via: skyway run skill-forge --var skill_name=<name>