Skill Forging Master

Problem

Repetitive workflows often stay in tribal knowledge or one-off prompts. The model may know how to use tools, but not the team’s acceptance rules, edge cases, or reusable assets. A SKILL.md that is too long consumes context, while one that is too shallow fails to trigger or enforce behavior. It can also overengineer one-off tasks.

How it works

The skill turns skill creation into decide, design, build, evaluate, package.
- Start with Step 0: check whether the work is repeated, needs consistency, and is stable enough to solidify; reject lighter options such as global config, direct tool use, or a prompt.
- Choose Lv1–Lv5 complexity so simple workflows are not forced into references/ or scripts/.
- Generate a scaffold, split work into model judgment and script execution, and write an aggressive description with trigger phrases and counter-examples.
- Validate with evals, baseline comparison, and iterative convergence; package the local version, and if publishing to SkillHub, run compliance adaptation for LICENSE, hidden files, null bytes, MIT-0, and proprietary fields.

Boundaries

Use it for stable, reusable workflows, not always-on rules or one-time tasks. Missing domain details are surfaced as questions, not invented. Before editing existing skills, create a backup; after editing, re-check frontmatter, version, and structure.