add-mcp-prompt

name	add-mcp-prompt
description	Use when adding a new MCP prompt (slash command) to the gdevelop-mcp server. Prompts inject a guided workflow message — they don't run code directly.

Adding an MCP prompt

Prompts in gdevelop-mcp are workflow templates surfaced as slash commands (e.g. /gdevelop:add-hero). They inject a structured user message that tells the agent which tools to call, in what order, and where to pause for user confirmation.

Checklist

• Decide if a prompt is the right primitive. A prompt is appropriate when the workflow has 3+ steps, involves multiple tools, and benefits from a fixed "rail". For single-step actions, a tool is enough.
• Add the prompt in src/prompts.ts inside registerPrompts(server).
• Use a kebab-case name (e.g. import-pack, setup-platformer).
• Define args with Zod — keep them few and meaningful.
• Write the prompt body as a numbered workflow: 1. inspect 2. discover 3. plan (dryRun if a write is involved) 4. wait for user approval 5. apply 6. verify
• When the workflow involves writing, always include dryRun: true before the apply step.
• Run npm test (smoke-tests prompts/list).
• Update README.md's prompts table.

Template

server.prompt(
  "my-prompt-name",
  "One-sentence description of what this workflow does.",
  {
    projectPath: z.string().describe("Absolute path to the project"),
    something: z.string().describe("..."),
  },
  ({ projectPath, something }) => ({
    messages: [
      {
        role: "user",
        content: {
          type: "text",
          text: [
            `<one-line goal statement>`,
            "",
            "Steps (do NOT modify until I approve):",
            `1. inspect_project({path: "${projectPath}"})`,
            `2. ...discovery tool calls...`,
            "3. Build the edit_project payload, dryRun:true.",
            "4. Show me the summary, wait for approval.",
            "5. Apply with dryRun:false. backup defaults to true.",
            "6. inspect_project to verify.",
          ].join("\n"),
        },
      },
    ],
  }),
);

Style guide

• The prompt body is read by the model — be explicit about which tools to call (give them by name, with their argument structure).
• Make confirmation points explicit: say "WAIT for my confirmation" before any write.
• Keep the prompt under ~30 lines — long prompts get ignored or paraphrased.