Claude Code Skills for MIOSA

Skills extend Claude Code with MIOSA-specific knowledge. Install them once and Claude automatically loads the right one when you ask for cloud infrastructure - no manual prompting.

What you get

Each skill ships with an examples.md (or reference.md) bundled file Claude loads on demand. SKILL.md bodies stay under 100 lines - minimum context cost, maximum signal.

Install

git clone https://github.com/Miosa-osa/miosa-skills.git
cp -r miosa-skills/* ~/.claude/skills/

Or per-project:

cp -r miosa-skills/* .claude/skills/

Verify

Open Claude Code in any project and ask:

What skills are available?

You should see all 5 miosa-* skills listed.

How they activate

You don’t need to invoke them manually. Each skill declares trigger phrases in its description and when_to_use frontmatter. Say things like:

• “create a sandbox and run this” → miosa-sandbox loads
• “take a screenshot and click the login button” → miosa-computer loads
• “deploy this repo to production” → miosa-deploy loads
• “I need a Postgres database” → miosa-data loads

You can also invoke directly with /miosa-sandbox, /miosa-computer, etc.

Structure (official Agent Skills spec)

~/.claude/skills/miosa-sandbox/
├── SKILL.md         # YAML frontmatter + tight overview (<100 lines)
└── examples.md      # Workflow patterns, loaded on demand

Each SKILL.md includes:

• description - what the skill does, when to use it
• when_to_use - trigger phrases for automatic activation
• allowed-tools - pre-approved MCP tools (no permission prompts)

Body covers: lifecycle, sizes, conventions, the core loop, errors. Detailed examples live in the sibling file so they don’t bloat context until you need them.

What’s in each skill

miosa-mcp

• Resource map (sandbox vs computer vs deployment vs database vs storage vs volume)
• Universal conventions (IDs, status polling, sizes, paths, timeouts)
• Decision tree pointing to the right resource skill
• reference.md with the full 145-tool catalog by category

miosa-sandbox

• Core loop: create_sandbox → exec → destroy_sandbox
• Size guidance (small=4GB, medium for builds, etc.)
• Heredoc trick for multi-file projects
• examples.md: build-and-serve, data processing, snapshot reuse, PR testing, benchmarks

miosa-computer

• Screenshot → identify → click/type → screenshot → verify loop
• Coordinate system (1024×768, absolute pixels)
• Mouse/keyboard tool table
• examples.md: browser automation, form filling, checkpoints, services, drag-drop, hotkeys

miosa-deploy

• GitHub → deployment flow
• Sandbox promotion via sandbox_deploy
• Framework auto-detection table
• examples.md: Next.js+DB, preview branches, rollback, env updates

miosa-data

• Database creation + credential wiring
• Storage buckets + presigned URLs (PUT and GET)
• Volume vs storage decision table
• Cron schedule reference
• examples.md: full-app patterns, backup pipelines, client-direct uploads

Customizing

The skills are intentionally tight. To extend:

1. Add an examples.md entry under the skill that needs the pattern
2. Add paths: to a skill’s frontmatter to restrict it to certain file types
3. Override allowed-tools if you want manual approval on certain tools

To override a skill’s listing (e.g. hide one):

// .claude/settings.local.json
{
  "skillOverrides": {
    "miosa-computer": "off"
  }
}

Why skills instead of CLAUDE.md

CLAUDE.md content loads into context on every session. Skills load only when relevant. For 145 tools across 6 resource types, that’s the difference between burning ~3,000 tokens per session vs ~200 tokens (just the descriptions).

Updating

cd ~/.claude/skills
git -C miosa-skills pull
cp -r miosa-skills/* .

Claude Code watches the skills directory and picks up changes mid-session - no restart needed.

Related

• MCP guide - the 145-tool surface these skills wrap
• CLI - miosa mcp install to configure the MCP server
• Quickstart - get a sandbox running in under 60s