skillopt-improve-skilllisted

# SkillOpt Improve Skill ## Goal Improve an existing skill toward a concrete objective using SkillOpt's discipline: treat the skill document as trainable state, run task rollouts, reflect on scored traces, propose bounded edits, apply only selected edits, and accept updates only when validation improves. Do the work end to end when possible. If the user gives only a skill and a goal, create an evaluation plan, run a lightweight local optimization loop, and leave a clear improved artifact plus optimization notes. Ask only when the goal cannot be scored or the target skill cannot be found. When running bundled scripts, use the directory containing this `SKILL.md`. In Claude Code, prefer `${CLAUDE_SKILL_DIR}`. In other environments, locate the installed or unpacked skill directory first. ## Resource Map - Read `references/skillopt-method.md` when deciding between the official SkillOpt package and a local Codex loop, or when configuring epochs, learning rate, slow update, meta skill, or validation gates. - Read `references/eval-design.md` when the user did not provide eval cases or the improvement goal needs a measurable scoring rubric. - Use `scripts/skillopt_workspace.py` to initialize an optimization workspace, apply SkillOpt-style JSON patches, summarize rollout results, and compare baseline vs candidate results. ## Workflow 1. Locate the target skill. - Prefer an explicit path from the user. - If the user names a skill, search `${CODEX_HOME:-$HOME/.codex}/skill