kjanat Building Skills
Skills extend agent capabilities with specialized knowledge, workflows, and tools.
Quick Start
Minimal viable skill in 30 seconds:
mkdir my-skill && cat > my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: Does X when Y happens. Use for Z tasks.
---
# My Skill
Instructions go here.
EOF
Place in .opencode/skills/ (project) or ~/.config/opencode/skills/ (global).
Skill Type Decision Tree
What are you building?
├─ Instructions only → Simple skill (SKILL.md only)
│ Example: code-review guidelines, commit message format
│
├─ Domain knowledge → Reference-heavy skill (+ references/)
│ Example: API docs, database schemas, company policies
│
├─ Repeatable automation → Script-heavy skill (+ scripts/)
│ Example: PDF processing, data validation, file conversion
│
├─ Complex multi-step workflow → Multi-file skill (all directories)
│ Example: release process, deployment pipeline
│
└─ Large platform → Progressive skill
Example: AWS, GCP, Cloudflare (60+ products)
When to Create a Skill
Create a skill when:
- Same instructions repeated across conversations
- Domain knowledge model lacks (schemas, internal APIs, company policies)
- Workflow requires 3+ steps with specific order
- Code rewritten repeatedly for same task
- Team needs shared procedural knowledge
When NOT to Create a Skill
| Scenario | Do Instead |
|---|---|
| Single-use instructions | AGENTS.md or inline in conversation |
| Model already knows domain | Don't add redundant context |
| < 3 steps, no reuse | Inline instructions |
| Highly variable workflow | Higher-freedom guidelines |
| Just want to store files | Use regular directories |
Reading Order
| Task | Files to Read |
|---|---|
| New skill from scratch | anatomy.md → frontmatter.md |
| Optimize existing skill | progressive-disclosure.md |
| Add scripts/resources | bundled-resources.md |
| Find skill pattern | patterns.md |
| Debug/fix skill | gotchas.md |
| Review/validate skill | review-checklist.md |
In This Reference
| File | Purpose |
|---|---|
| anatomy.md | Skill directory structures |
| frontmatter.md | YAML spec, naming, validation |
| progressive-disclosure.md | Token-efficient design |
| bundled-resources.md | scripts/, references/, assets/ |
| patterns.md | Real-world skill patterns |
| gotchas.md | Common mistakes + fixes |
| review-checklist.md | Post-creation review + dispatch |
Scripts
| Script | Purpose |
|---|---|
scripts/init_skill.sh |
Scaffold new skill |
scripts/validate_skill.sh |
Validate skill structure |
scripts/package_skill.sh |
Create distributable zip |
Post-Creation Review
After creating or significantly modifying a skill, dispatch a reviewer agent (oracle or code-reviewer) to cross-validate against the best practices checklist. See review-checklist.md for the full checklist and prompt template.
Pre-Flight Checklist
Before using a skill:
SKILL.md starts with
---(line 1, no blank lines)name:field present, matches directory namedescription:includes what + when to useClosing
---after frontmatterSKILL.md under ~200 lines (use references/ for more)
All internal links resolve
Validate the skill using skills-ref:
uvx 'git+https://github.com/agentskills/agentskills.git#subdirectory=skills-ref' \ validate <SKILL_DIR> # Validate a skill skills-ref validate path/to/skill # Read skill properties (outputs JSON) skills-ref read-properties path/to/skill # Generate <available_skills> XML for agent prompts skills-ref to-prompt path/to/skill-a path/to/skill-b
Skill Locations
| Priority | Location |
|---|---|
| 1 | .opencode/skills/<name>/ (project) |
| 2 | ~/.config/opencode/skills/<name>/ (global) |
| 3 | .claude/skills/<name>/ (Claude-compat) |
| 4 | $(git rev-parse --show-toplevel)/skills (shared) |
Discovery walks up from CWD to git root. First-wins for duplicate names.
See Also
- Cloudflare Skill - Reference implementation