Prompts

The DTPR MCP server exposes each of the seven DTPR Agent Skills — plus the two shared reference documents the schema-tier skills cite — as registered MCP prompts. Any MCP-aware client (Cursor, Continue, Cline, Claude Desktop, custom clients, the Anthropic Agent SDK) can list and load them via the standard prompts/list and prompts/get methods.

Listed prompts

The skill bodies are returned verbatim from the published SKILL.md files (frontmatter included), with one rewrite: plugin/dtpr/references/<slug>.md citations are translated to dtpr-references-<slug> prompt names so the skill body stays self-coherent when loaded outside the plugin checkout.

prompts/list

Returns the registered prompts, each with a short description suitable for a client picker. The full skill description (often 1–2 KB of trigger phrases and sibling routing) is preserved inside the body — the listing description is truncated to the first sentence so picker UIs stay readable.

// → POST /mcp
{ "jsonrpc": "2.0", "id": 1, "method": "prompts/list" }

// ← response
{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "prompts": [
      {
        "name": "dtpr-describe-system",
        "description": "Describe an AI system, algorithm, or automated decision process as a DTPR (Digital Trust for Places and Routines) AI datachain."
      },
      …
    ]
  }
}

prompts/get

Returns one prompt body as a single user message. The host injects the message into its conversation, exactly the way Claude Code's skill subsystem injects skills when the plugin is installed.

// → POST /mcp
{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "prompts/get",
  "params": { "name": "dtpr-element-design" }
}

// ← response
{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "description": "Brainstorm one proposed DTPR element …",
    "messages": [
      {
        "role": "user",
        "content": {
          "type": "text",
          "text": "---\nname: dtpr-element-design\n…"
        }
      }
    ]
  }
}

Capabilities

The MCP server advertises the prompts capability on initialize:

{
  "capabilities": {
    "tools": {},
    "resources": {},
    "prompts": {}
  }
}

Per the 2025-06-18 MCP spec, only user and assistant roles are valid in prompts/get messages. DTPR prompts always return a single user message.

v1 limits

• Zero arguments. Every prompt accepts no arguments. The skill body is self-contained; the user prompt comes after the skill body is injected. Typed arguments are reserved for v2 if a real consumer needs them.
• No localization. Skill bodies are English-only.
• No notifications/prompts/list_changed. Clients should re-fetch on session start. The bundled prompt set ships with each API deploy and changes only on PR-driven release.
• Reference rewrite is path-only. Citations like Read \plugin/dtpr/references/comprehension-rubric.md`becomeRead `dtpr-references-comprehension-rubric` (load via MCP prompts/get)`. Surrounding prose is preserved verbatim — slightly awkward but unambiguous.
• Research corpus is unavailable in MCP-prompt mode. Skills that mention plugin/dtpr/research/INDEX.md degrade gracefully (every skill already documents this fallback path) — the agent simply skips the corpus lookup phase.

Related

• Connecting — endpoint, headers, wire format.
• Envelope — ok / err envelopes shared by tools (prompts use the standard MCP prompts/get shape directly, not the DTPR envelope).
• Skills — per-skill reference, the source of every prompt body.
• Other harnesses — using prompts from outside the Claude plugin.