primitives-showcase

name	primitives-showcase
description	Exercises every SDK primitive in a single skill. Used as a reference and integration test.
metadata	{"version":"1.0.0"}
argument-hint	[theme]
allowed-tools	Bash(scripts/run *) Read mcp__primitives-showcase__start mcp__primitives-showcase__advance

primitives-showcase

This skill is a structured workflow. You interact with it via MCP tools (preferred) or a CLI binary (fallback), reading its JSON output, following the instructions in the prompt field, and passing your response back. Do not show the raw JSON or tool calls to the user.

Use MCP tools if available. If you have mcp__primitives-showcase__start and mcp__primitives-showcase__advance, use MCP mode. Only fall back to CLI mode if MCP tools are not available in your environment.

How this skill works

This skill was built with skill-kit, a structured workflow engine. Each step provides a prompt containing XML-tagged sections:

• <system> — Behavioral directives: persona, tone, or constraints. Follow as guidelines for how to behave, not as tasks to relay to the user.
• <prompt> — Task instructions: what to do, what context to consider, what to produce.
• <ask-user> — Ask the user a question. Contains <option> children for structured choices, or type="open" for free-form conversation.
• <confirm> — Binary yes/no confirmation. Attributes: default, destructive.
• <plan> — Present a plan for approval. Contains <step> children.
• <checklist> — Create tracked work items. Contains <item> children with status.
• <subagent> — Delegate work to an isolated sub-agent. If no-recurse is set, the subagent must not invoke the skill named in the attribute.
• <rendered> — Pre-rendered output. Emit verbatim — no edits, no added commentary.

A step may contain one or more of these sections in sequence. Follow them in order.

The skill author composed these sections to guide your behavior. The tags and tool mappings are generated by the SDK based on the author's intent and your host's capabilities. A skill-level system directive may appear in the preamble — it applies to all steps unless a step includes its own <system> override.

The preamble (sent on the first step) contains a table mapping each tag to the specific tool available in your environment. Refer to it throughout the workflow.

How to run this skill

MCP mode (preferred)

If you have MCP tools for this skill (e.g., mcp__primitives-showcase__start and mcp__primitives-showcase__advance), use them instead of the CLI:

1. Call the start tool (with params if the skill requires them).
2. Read the preamble field (first call only). It maps XML tags to your available tools.
3. Follow the prompt instructions. Produce a JSON object matching the schema.
4. Call the advance tool with the session, step, and output.
5. Repeat steps 3-4 until status is "done".

If you get status: "error" with retry: true, fix your output and resubmit. Do not show raw JSON, session IDs, or MCP tool calls to the user.

Skip the rest of this section — the CLI instructions below are only needed when MCP tools are not available.

CLI mode (fallback)

This SKILL.md file is inside the skill directory. Resolve the absolute path to scripts/run from this file's location (e.g., /path/to/skill/scripts/run). Use the absolute path in all Bash commands — do not cd into the skill directory.

In the examples below, <skill>/scripts/run is a placeholder for this absolute path.

Detect your host

Determine which agent host you are running in, and pass it as --host:

• Claude Code: --host claude-code
• Codex: --host codex
• OpenCode: --host opencode
• Gemini CLI: --host gemini-cli
• Cline: --host cline
• Roo Code: --host roo-code
• Kilo Code: --host kilo-code
• Cursor: --host cursor
• Amp: --host amp
• Unknown/other: omit the flag (defaults to generic)

Report your tools

Pass the tools you have available as a comma-separated --tools flag on the start command. The session remembers them — you don't need to pass --tools on advance.

When --host is provided, --tools is merged with the host's known tool registry. This means partial reporting is handled gracefully — the registry fills in any tools you omit. If --tools is omitted entirely, the skill infers tools from --host. If both are omitted, all interactions use generic fallbacks.

Subagent invocations

If you are a subagent (spawned by another agent, not the top-level agent the user is talking to), add --subagent to the start command. This tells the skill that your reported tools are a genuine subset — the skill will not merge them with the host registry.

Without --subagent, the skill assumes you are a top-level agent and merges your tools with the registry (since top-level agents often under-report their tools).

Parameters

This skill takes no parameters. Pass --params '{}'.

Step 1: Start with a session

<skill>/scripts/run --params '{}' --host claude-code --tools <your-tools> --session new 2>/dev/null

This returns a JSON pointer with sessionId, file, and line. The line field tells you which line to read — it will be 2, not 1 (line 1 is an internal header, never read it).

Read only line line from file. It contains the step prompt, schema, and preamble.

Read the preamble first. It contains a table mapping XML tags to the tools available in your environment. Refer to it throughout the workflow.

Step 2: Follow the prompt

Read the prompt field. It contains XML-tagged sections (described in "How this skill works" above): <system> directives to follow, <prompt> instructions to act on, and interaction tags (<ask-user>, <confirm>, <plan>, <checklist>, <subagent>) to execute using the tools mapped in the preamble. If a <rendered> block appears, emit its content verbatim.

Produce a JSON object matching the schema.

Step 3: Advance

Pass your output back with the step name:

<skill>/scripts/run advance --step <step-name> --output '<your-json>' --session abc123 2>/dev/null

This returns a single line number (e.g., 4). Read exactly and only that line from the session file — it contains the next prompt. Do not read any other lines.

Step 4: Repeat until done

Keep advancing until the line you read contains "type":"done". The finalOutput field contains the skill's result. Present it to the user.

Step 5: Cleanup

After presenting the result, remove the session file:

<skill>/scripts/run cleanup --session <session-id> 2>/dev/null

Important

• Never show raw JSON output or Bash commands to the user. The user sees your natural language responses, not the protocol.
• If you get a validation error (the response has "error": "validation" or "type":"error"), read the message field, fix your output, and retry the same step.

Steps in this skill

• gather-preferences: (dynamic)
• choose-format: (dynamic)
• ask-focus: (dynamic)
• research: (dynamic)
• plan-report: (dynamic)
• ask-changes: (dynamic)
• write-report: (dynamic)
• confirm-publish: (dynamic)
• summary: (dynamic)