jujugrowth-mcp 1.0.6 → 1.0.7

package/README.md

@@ -55,4 +55,17 @@ A token can be scoped to one business; the server then announces itself as
 `jujugrowth (yoursite.com)` and only ever touches that site. Generate scoped
 tokens in jujugrowth → Settings → Developer (one per business/repo).
 
+## Prompts to get started
+
+Open your AI in your site's project and try:
+
+- **See the work:** "List my jujugrowth recommendations and summarize each in one line."
+- **Implement one:** "Get the full brief for the most impactful recommendation, verify it against the code, then implement it here — show me a diff before committing. Mark it handled when it's in."
+- **Fresh analysis (Pro):** "Ask jujugrowth to run a fresh website analysis for this site, then list the new recommendations."
+- **Hands-off:** "Pull my open jujugrowth recommendations, implement the website/SEO ones in this repo (skip GA4-console or business-judgment items — flag those for me), open a PR, and mark each handled with a note."
+
+The server also tells your AI the workflow on connect (verify-before-changing,
+scope guardrails, always mark handled), so it follows the loop without you
+spelling it out each time.
+
 Requires Node ≥ 22. Optional env `JUJUGROWTH_API` (defaults to https://jujugrowth.com).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jujugrowth-mcp",
-  "version": "1.0.6",
+  "version": "1.0.7",
   "description": "MCP server connecting your AI coding assistant (Claude, Cursor, Codex) to jujugrowth — pull recommendations and apply them in your repo.",
   "type": "module",
   "bin": {

package/server.mjs

@@ -68,7 +68,7 @@ const TOOLS = [
   {
     name: "get_recommendation",
     description:
-      "Get the full implementable brief for one recommendation (title, body, and a ready-to-implement claude_prompt when present). Implement it in the user's repo, scoped to wording/structure/SEO — never touch checkout, auth, or analytics events.",
+      "Get the full implementable brief for one recommendation (title, body, and a ready-to-implement claude_prompt when present). Verify the brief's current-state claims against the actual code first (they can be stale), then implement in the user's repo scoped to wording/structure/SEO — never touch checkout, auth, or analytics events. After it's shipped, ALWAYS call mark_recommendation_handled with a note of what you changed.",
     inputSchema: {
       type: "object",
       properties: { tenantId: { type: "string" }, recId: { type: "string" } },
@@ -138,6 +138,16 @@ async function handle(msg) {
           name: SITE_LABEL ? `jujugrowth (${SITE_LABEL})` : "jujugrowth",
           version: "1.0.0",
         },
+        // Workflow the AI should follow unprompted (so the human never has to
+        // say "now mark it handled"). Surfaced to the model by MCP clients.
+        instructions:
+          `This server provides growth recommendations for ${SITE_LABEL ? `the site "${SITE_LABEL}"` : "the user's site(s)"} and connects them to your code work. Standard loop:\n` +
+          "1) list_recommendations.\n" +
+          "2) get_recommendation for one you'll act on.\n" +
+          "3) Before changing anything, VERIFY the brief's current-state claims against the actual code — they can be stale; skip what's already done.\n" +
+          "4) Implement in the user's repo, scoped to wording/structure/SEO. Never change checkout, auth, or analytics events. Show a diff / open a PR for the user to review.\n" +
+          "5) ALWAYS call mark_recommendation_handled with a short note of what you changed once it's shipped — don't wait to be told.\n" +
+          "Skip GA4-console toggles and business-judgment items — flag those for the user instead of editing code. This server is bound to ONE site; never act on another.",
       },
     });
   }