New skill: mcp-guide#34
Conversation
Interactive, module-based guide for getting value out of the Mixpanel MCP server. Teaches orient → analyze → investigate → build → govern → chain workflows one concept at a time, with tiered starter/advanced paths and built-in troubleshooting. Tool-agnostic — works across any MCP-connected AI client. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
@rpatel05 Please run https://github.com/mixpanel/ai-plugins/tree/main/.claude/skills/review-skill on this skill and put the score in your testing plan (should be above 70%). |
| **Reference (open anytime):** | ||
| - **Module 7:** Prompting principles and known limits | ||
|
|
||
| Ask where they'd like to start. New to MCP? Start at Module 0. |
There was a problem hiding this comment.
The can be more specific e.g. "If you're new to MCP, start with Module 0. Which module do you want to start with?"
There was a problem hiding this comment.
Or you could even instruct them more directly with something like: "Reply with a module number to begin".
| @@ -0,0 +1,343 @@ | |||
| --- | |||
| name: mcp-guide | |||
| description: Use this skill whenever the user asks about the Mixpanel MCP server, what it does, what it can answer, what tools are available, how to phrase questions, why to use it, or what value it provides. Trigger on phrases like "how do I use Mixpanel MCP", "what can I ask Mixpanel", "Mixpanel in Claude", "Mixpanel natural language", "connect Mixpanel", "Mixpanel through AI", "show me what Mixpanel MCP can do", or any open-ended question about Mixpanel's MCP capabilities. Also trigger when the user has just connected Mixpanel and is figuring out where to start, even if their question is vague (e.g. "what now?", "what should I try first?"). Do NOT skip this skill on the assumption a question is simple. The skill contains canonical tool names, prompt patterns, and known gotchas that are easy to get wrong from memory. | |||
There was a problem hiding this comment.
"Do NOT skip this skill on the assumption..." can be more agent directive with: "If the user is asking a basic question, always invoke this skill first."
| > "List the events that relate to checkout. For each, give the description and top three properties. Then suggest which combination best represents `purchase intent` based on how this project is actually instrumented." | ||
| That version surfaces Lexicon gaps instead of assuming. | ||
|
|
||
| #### → if it didn't work |
There was a problem hiding this comment.
These types of conditions (also lines 94) are vague. Agents will do better with more concrete routing rules like: "if user mentions 'error' , 'didn't work', 'empty' then ..."
| **Concept:** The highest-leverage thing MCP does is combine Mixpanel with your other connected tools (error monitoring, trackers, CRM, feedback) for synthesis that used to take a week of manual coordination. The discipline that makes it work: decompose into one tool per turn, quantitative first, qualitative last. | ||
|
|
||
| **Try this:** | ||
| > Pick one tool besides Mixpanel that you have connected. Run a 2-step chain: get a Mixpanel finding first, then a follow-up that pulls from the second tool to deepen it. |
There was a problem hiding this comment.
This could be more procedural in order it is meant to be executed: 1. get MP finding 2. pick 2nd tool (or prompt user) 3) find a follow up with 2nd tool
zoeabrams-eng
left a comment
zoeabrams-eng
left a comment
There was a problem hiding this comment.
Will merge once comments are addressed and pr passes /review-skill
2 participants
What
Adds a new
mcp-guideskill — an interactive, module-based guide that teaches users how to get real value out of the Mixpanel MCP server. Covers orient → analyze → investigate → build → govern → chain workflows, one concept at a time, with tiered starter/advanced paths and built-in troubleshooting.Why
New MCP users often either underuse the server (treating it as a faster report builder) or get stuck on schema discovery and prompt structure. This skill provides a structured onboarding path that teaches the four-part prompting rule (Behavior + Population + Timeframe + Shape), surfaces common gotchas (wrong event names, missing descriptions, permission errors), and progressively introduces advanced workflows like user-level investigation, Lexicon governance, and cross-tool chaining. It's tool-agnostic — works across any MCP-connected AI client (Claude, OpenAI, Gemini, Cursor, etc.).
Type
Testing
Starter path — orient and analyze
Result: Returns project list, top events by volume, flags undocumented events, and reads like a schema orientation — not a raw data dump.
Advanced path — investigate with Flows
Result: Moves from aggregate funnel to individual user investigation. Returns distinct IDs, post-drop-off paths, and a hypothesis for why users didn't convert.
Prompting principles (Module 0)
Result: Surfaces 2-3 relevant principles (specificity, iterate don't restart, one ask per turn) without dumping the entire reference list.