skill-forge

Skill Forge is a meta-skill that guides the creation of high-quality, production-grade skills for Claude Code. It treats a skill as an "onboarding guide" that transforms a general-purpose agent into a specialized one with procedural knowledge, domain expertise, and bundled tools. Its governing principle, stated as an "Iron Law," is that every line in a skill must justify its token cost: if a line does not make the model's output better, more consistent, or more reliable, it should be cut. The skill assumes Claude is already capable and adds only what the model does not already know.

The workflow is organized as a trackable checklist spanning nine steps: understanding the skill (collecting at least three concrete usage examples and trigger keywords), planning architecture, initializing via a bundled init_skill.py script, writing the description, writing the SKILL.md body, building resources, reviewing, packaging via package_skill.py, and iterating on real usage. Certain steps are marked REQUIRED or BLOCKING so they are not skipped. The skill emphasizes that the description field determines whether a skill triggers automatically and is discoverable by search, and warns against placing "When to Use" information in the body, which loads only after triggering.

Use it when you want to create, build, design, write, update, improve, refactor, debug, or package a skill. It explains the canonical directory layout (SKILL.md plus optional scripts/, references/, and assets/), keeping SKILL.md under 500 lines while moving detail into on-demand reference files. It teaches writing techniques such as question-style instructions, anti-pattern documentation, and Iron Law plus red-flag signals, along with confirmation gates for destructive or costly operations and an optional parameter system using flags and arguments. Much of the deeper guidance lives in bundled reference files (description-guide, architecture-guide, writing-techniques, workflow-patterns, parameter-system, output-patterns) that are loaded only when a given sub-step needs them, demonstrating the progressive-loading strategy it advocates.