skill-creator

What I do

Create and validate OpenCode skills following the Anthropic Agent Skills spec. Skills that violate these rules fail silently: not discovered, not matched, or waste tokens.

Golden-template rule: copy, don't compose

NEVER write a new skill from a blank file. Find the existing skill closest in shape to what you need (a command-reference skill, a workflow skill, an audit skill), copy its SKILL.md as the starting point, and modify the minimum: name, description, and the body sections that differ. The frontmatter idioms, section order, and command style come along for free — which is exactly what freehand composition gets wrong. If you find yourself rewriting most of the template, you picked the wrong exemplar; pick another one.

Path resolution for references/

Skills load from different locations per tool (~/.claude/skills/, ~/.config/opencode/skills/, project .claude/skills/). Inside a SKILL.md, paths like references/foo.md are ALWAYS relative to the directory containing that SKILL.md — never to the project CWD. When authoring, write reference links relative; when following them as an agent, resolve them against the SKILL.md location before reading.

Directory structure

.claude/skills/{skill-name}/
├── SKILL.md              # lean entry point — always loaded when the skill triggers
└── references/           # OPTIONAL: depth loaded on demand
    └── {subtask}.md

One directory per skill. Directory name must match name in frontmatter. The entry file is always SKILL.md (uppercase). The Agent Skills spec caps SKILL.md at 500 lines; in this repo prefer ~200 (everything in SKILL.md hits the context whether the task needs it or not). When a section is only relevant to a subset of tasks or has grown into step-by-step depth, move it to references/{subtask}.md and add a routing table in SKILL.md (| task | read |) pointing to it — keep inline only orientation, path-selection, and gotchas the agent must see before choosing a path. Keep each reference under ~200 lines.