skill-creator

name

skill-creator

description

Guide for creating effective opencode skills. Use for creating or updating skills that extend agent capabilities with specialized knowledge, workflows, or tool integrations. Examples: - user: "Create a skill for git workflows" → define SKILL.md with instructions and examples - user: "Add examples to my skill" → follow the user: "query" → action pattern - user: "Update skill description" → use literal block scalar and trigger contexts - user: "Structure a complex skill" → organize with scripts/ and references/ directories - user: "Validate my skill" → check structure, frontmatter, and discovery triggers

Skill Creator

Create opencode skills that extend agent capabilities with specialized knowledge and workflows.

What Skills Provide

• Specialized workflows - Multi-step procedures for specific domains
• Tool integrations - Instructions for file formats, APIs, libraries
• Domain expertise - Company-specific knowledge, schemas, business logic
• Bundled resources - Reusable scripts, references, and assets

Skill Locations

Scope	Path
Project	.opencode/skill/<name>/SKILL.md
Global	~/.config/opencode/skill/<name>/SKILL.md

• Project skills: Team-shared, repo-specific (e.g., our-api-patterns, project-deploy)
• Global skills: Personal tools for all projects (e.g., pdf-editor, commit-helper)

For project paths, OpenCode walks up from cwd to git worktree root.

Skill Structure

skill-name/
├── SKILL.md              # Required - frontmatter + instructions
├── scripts/              # Optional - executable code (Python/Bash)
├── references/           # Optional - docs loaded on-demand
└── assets/               # Optional - templates, images, fonts

SKILL.md Format

---
name: skill-name
description: [Self-contained workflow summary — see guidelines below]
---
# Instructions here (markdown body)

Name: Short, hyphen-case identifier. Should be descriptive but concise (max 64 chars).

Description: Agent sees this + the name before loading. Must be self-contained with:

• What workflow/capabilities it provides
• "Use proactively when" trigger contexts
• 3-5 concrete examples

CRITICAL: The description field is the primary trigger mechanism.

Skills are SOPs/workflows, NOT agents. DO NOT use role descriptions like "You are a..." or "[Role] expert."

Before loading, the agent sees only the name and description in <available_skills>:

<available_skills>
  <skill>
    <name>skill-name</name>
    <description>...</description>
  </skill>
</available_skills>

The description must be self-contained — agents won't load a skill just to "see what it does."

Name + description should work together:

• Name: Short, hyphen-case identifier (e.g., typescript-advanced)
• Description: Self-contained workflow summary with capabilities, triggers, and examples

Description pattern (LLM-optimized):

---
name: skill-name
description: |-
  [Action verb/capabilities]. Use for [specific cases]. Use proactively when [contexts].
  
  Examples:
  - user: "query" → action
  - user: "query" → action
---

Dense, machine-parseable, specific. Avoid prose.

CRITICAL YAML SYNTAX: Multi-line descriptions with examples MUST use literal block scalar (|-). The hyphen strips the trailing newline. Do NOT use plain YAML with unquoted colons or lists:

# ❌ WRONG - breaks YAML parsing
description: Handle plugins. Examples:
- user: "..." → action

# ✅ CORRECT - use |- for multi-line
description: |-
  Handle plugins.
  
  Examples:
  - user: "..." → action

Example:

---
name: typescript-advanced
description: |-
  Handle TypeScript 5.9 advanced typing, generics, strict configs, type errors, migrations,
  erasable syntax compliance, and test writing. Use proactively for complex generics,
  conditional types, utility types, TS compiler config, or test authoring.
  
  Examples:
  - user: "Create a type-safe event emitter" → implement with generics and mapped types
  - user: "Migrate to strict TypeScript" → add discriminated unions, exhaustive checks
  - user: "Build typed API client from OpenAPI" → generate request/response types with inference
  - user: "Write unit tests" → create strict, typed tests with realistic fixtures
---

Requirements:

• Start with action verb (NOT "You are" or "[Role] expert")
• List specific capabilities (vague "helps with X" = ignored)
• Include "Use proactively when" trigger contexts
• Provide 3-5 concrete user: "..." → ... examples
• Use |- literal block scalar for multi-line descriptions (plain YAML with lists/colons breaks parsing)
• Dense, LLM-parseable — description alone must justify loading

Reduce redundancy: Don't repeat description content in SKILL.md body.

Bundled Resources

Directory	Purpose	When to use
scripts/	Reusable Python/Bash code	Same code rewritten repeatedly
references/	Docs, schemas, API specs	Info agent needs while working
assets/	Templates, images, fonts	Files used in output (not loaded)

MUST NOT include: README.md, CHANGELOG.md, INSTALLATION_GUIDE.md, or other auxiliary docs. Skills contain only what the agent needs to do the job.

Core Principles

Be Concise

The context window is shared. Only add info the agent doesn't already have.

• Challenge each paragraph: "Does this justify its token cost?"
• Prefer examples over explanations
• SHOULD keep SKILL.md under 500 lines

Match Freedom to Fragility

Freedom Level	Format	Use When
High	Text instructions	Multiple valid approaches
Medium	Pseudocode/parameterized scripts	Preferred pattern exists
Low	Specific scripts	Fragile ops, consistency critical

Progressive Disclosure

1. Metadata (name + description) - Always loaded (~100 words)

   • The description is the PRIMARY discovery mechanism
   • Agents see this and decide whether to load the skill
   • If description is vague, the skill will never be used

2. SKILL.md body - Loaded when skill triggers

   • Core workflow and detailed instructions
   • Keep focused on what the agent needs to do the work

3. Bundled resources - Loaded on-demand by agent

   • Move variant-specific details to references/

Keep core workflow in SKILL.md. Move variant-specific details to references/.

Example structure:

cloud-deploy/
├── SKILL.md (workflow + provider selection)
└── references/
    ├── aws.md
    ├── gcp.md
    └── azure.md

Agent loads only the relevant provider file.

Creation Process

1. Understand → Gather concrete usage examples
2. Plan → Identify reusable scripts/references/assets
3. Initialize → Create directory and SKILL.md manually
4. Edit → Write SKILL.md frontmatter + body, add resources
5. Validate → Verify the skill loads and triggers correctly
6. Iterate → Test, improve, repeat

Step 1: Understand

Gather concrete examples of how the skill will be used. Ask:

• "What should this skill do?"
• "What requests should trigger it?"
• "Can you give example user queries?"

Skip only if usage patterns are already clear.

Step 2: Plan

For each use case, identify reusable resources:

If you find yourself...	Add to...
Rewriting same code	scripts/
Re-discovering schemas/docs	references/
Copying same templates	assets/

Examples:

• pdf-editor: "Rotate this PDF" → scripts/rotate_pdf.py
• bigquery: "How many users today?" → references/schema.md
• frontend-builder: "Build me a todo app" → assets/react-template/

Step 3: Initialize

Create the skill directory and SKILL.md manually:

# Global skill (personal tools)
mkdir -p ~/.config/opencode/skill/my-skill

# Project skill (team-specific)
mkdir -p .opencode/skill/my-skill

Create SKILL.md with frontmatter:

---
name: my-skill
description: |-
  [Workflow/capabilities]. Use for [specific cases]. Use proactively when [contexts].
  
  Examples:
  - user: "query" → action
  - user: "query" → action
---
# [Skill Name]

[Instructions start here]

Description template example:

---
name: typescript-advanced
description: |-
  Handle TypeScript 5.9 advanced typing, generics, strict configs, type errors, migrations,
  and test writing. Use proactively for complex generics, conditional types, utility types, TS compiler
config, or test authoring.

Examples:
- user: "Create a type-safe event emitter" → implement with generics and mapped types
- user: "Migrate to strict TypeScript" → add discriminated unions, exhaustive checks
- user: "Build typed API client from OpenAPI" → generate request/response types with inference
- user: "Write unit tests" → create strict, typed tests with realistic fixtures
---

Add optional directories as needed:

cd my-skill
mkdir scripts references assets  # only create what you'll actually use

Step 4: Edit

Writing guidelines:

• MUST use imperative form ("Run the script", not "You should run")
• SHOULD use bullet points over prose
• SHOULD link to references for detailed info
• MUST test all scripts before including

Frontmatter requirements:

• name: lowercase-hyphen format, must match directory name exactly
• description: CRITICAL — follow the format specified in above. This is how agents discover your skill.

Step 5: Validate

Verify the skill is discoverable:

1. Check structure:

   • SKILL.md exists in skill directory
   • Directory name matches name: in frontmatter exactly
   • YAML frontmatter is valid

2. Test discovery:

   • The skill should appear in agent's <available_skills> section
   • Description should be specific enough to match relevant queries

3. Verify triggers:

   • Read your description — would you know when to use this skill based solely on it?
   • Do the examples cover the main use cases?

Note: Skills are used directly from their directories. No packaging or installation step required.

Step 6: Iterate

After real usage:

1. Notice where the skill fails to trigger or provides unclear guidance
2. Update the description to include missing trigger contexts
3. Add more examples to the description or SKILL.md body
4. Re-validate discovery and triggers

Agent Permissions

Control skill access per-agent in agent config:

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

Values: "allow", "deny", "ask". Use "*" as wildcard default.

<question_tool>

Batching: Use the question tool for 2+ related questions. Single questions → plain text.

Syntax: header ≤12 chars, label 1-5 words, add "(Recommended)" to default.

When to ask: Ambiguous request, multiple skill patterns apply, or scope unclear.

</question_tool>