Skill Editor

Expert at creating and maintaining skills following Anthropic's prompt engineering best practices, targeting both Claude Code and OpenCode.

Skill Locations

Skills can be placed in project-local or global directories. Both tools search multiple paths:

Claude Code

• •Project: .claude/skills/<skill-name>/SKILL.md
• •Global: ~/.claude/skills/<skill-name>/SKILL.md

OpenCode

• •Project: .opencode/skills/<skill-name>/SKILL.md
• •Global: ~/.opencode/skills/<skill-name>/SKILL.md
• •Also reads Claude-compatible paths: .claude/skills/<skill-name>/SKILL.md
• •Also reads agent-compatible paths: .agents/skills/<skill-name>/SKILL.md

Claude Code SKILL.md Format

---
name: lowercase-hyphenated
description: Clear, specific description with trigger keywords.
argument-hint: "[optional-hint]"
disable-model-invocation: true/false    # true = user-only
user-invocable: true/false              # false = background only
allowed-tools: Tool1, Tool2
context: fork                           # Run in subagent
agent: Explore                          # Subagent type when forked
---

Instructions Claude follows when skill is active.

OpenCode SKILL.md Format

---
name: lowercase-hyphenated              # Must match directory name
description: Clear, specific description with trigger keywords.
license: MIT                            # Optional
compatibility: opencode                 # Optional
metadata:                               # Optional, string-to-string map
  audience: developers
---

Instructions the agent follows when skill is loaded.

OpenCode Frontmatter Fields

OpenCode Skill Invocation

Skills are NOT directly user-invoked in OpenCode. The agent discovers them via a built-in skill tool. The agent sees all skill names/descriptions and calls skill({ name: "skill-name" }) to load on demand.

Permissions can be controlled in opencode.json:

{
  "permission": {
    "skill": {
      "*": "allow",
      "internal-*": "deny"
    }
  }
}

Skill Directory Structure

<skill-name>/
  SKILL.md           # Main instructions (required)
  references/        # Detailed reference material (optional)

Key Principles

1. •Description is critical - Both tools use it for automatic loading/discovery
2. •Keep SKILL.md under 500 lines - Put detailed reference material in references/
3. •Use $ARGUMENTS for user input (Claude Code)
4. •Use !command`` for dynamic context (Claude Code)
5. •Set disable-model-invocation: true for side-effect workflows (Claude Code only)
6. •Name must match directory - name frontmatter must equal the directory name

Workflow

When creating a new skill:

1. •Determine the skill's purpose and which tools it targets (Claude Code, OpenCode, or both)
2. •Choose a descriptive lowercase-hyphenated name
3. •Write a description rich in trigger keywords so the agent loads it automatically
4. •Draft the SKILL.md body with clear, concise instructions
5. •If instructions exceed ~300 lines, split reference material into references/ files
6. •Create the SKILL.md in the appropriate directories:
   • •For Claude Code: .claude/skills/<name>/SKILL.md
   • •For OpenCode: .opencode/skills/<name>/SKILL.md
   • •For both: create in both locations (or just .claude/skills/ since OpenCode reads it too)

Quality Checklist

• • Description includes trigger keywords for automatic discovery
• • SKILL.md is focused and under 500 lines
• • Reference material in separate files if needed
• • Appropriate invocation controls set
• • Name matches directory name
• • Frontmatter matches target tool's format