Heidelberg AICurriculum

Claude's Workflows Actually Work Now | MCP Live Build

n8n ·2026-05-28 ·2 min read

Summary written by us from the video's transcript. The video, and everything in it, is n8n's work.

Learn how to build a Slack‑based workflow reviewer in n8n using Claude Desktop, the MCP, and public skills, with step‑by‑step setup, planning, and debugging guidance.

Takeaways

  • Define the exact user interaction up front; it guides prompt engineering and prevents scope creep.
  • Install the n8n‑skills and Superpowers plugins in Claude Desktop and restart to ensure they load correctly.
  • Always include a full list of skill file paths in the agent’s initial prompt so tools like `getSkill` can be invoked.
  • Read the entire plan generated by Claude, not just the summary, to catch hidden assumptions before building.
  • Use Slack app manifests for quick configuration and store tokens/secrets as n8n credentials for secure access.

Defining the Desired Workflow Review

The speaker starts by outlining the end‑user experience: DM or tag a Slack bot with an n8n workflow URL and receive a detailed review in the thread. The review must cover top‑level and sub‑workflows, provide an overall grade, overview, and findings categorized as must fix, should fix, and nice to have, formatted with Slack Block Kit.

Preparing Claude Desktop and Required Plugins

Five preparatory steps are required: create a working folder; connect the n8n MCP using the OAuth URL from n8n settings; install the n8n‑skills plugin via the marketplace path `n8n-io/skills`; add the Superpowers plugin for planning; write a comprehensive `Claude.md` to prioritize loading the n8n skill and set the mode to bypass permissions. The speaker emphasizes restarting Claude Desktop if plugins do not appear immediately.

Planning with Claude’s Brainstorming Skill

Claude is fed the transcript of the speaker’s earlier explanation and prompted to ask enough questions for uninterrupted execution. The brainstorming skill generates a plan document; the speaker stresses reading the full plan—not just the summary—to catch potential mistakes before proceeding.

Creating the Slack App Manifest

Using Claude, a valid Slack app manifest is generated, then manually uploaded at api.slack.com/apps via “Create New App from Manifest.” After creation, the bot token and signing secret are copied into new n8n Slack API credentials. The speaker also chooses a name for the bot (“workflow whisperer”).

Configuring n8n Credentials and Connecting Tools

An n8n API key with read‑workflow scope is created and stored as an n8n credential (base URL `<n8n‑url>/api/v1`). The speaker explains the role of each component: Claude Code runs locally, the MCP bridges Claude to n8n, and n8n skills teach the model specific tasks.

Iterative Debugging and Fixes

During execution, the workflow shows sticky notes stacked and missing skill calls. The speaker discovers that the agent never received a list of available skill files, causing `getSkill` failures. Claude suggests two fixes: load skill paths into the agent prompt and enforce strict use of skills for the review. After applying these changes, the workflow successfully calls six skill files and handles bad URLs gracefully.

Final Testing and Publication

The completed workflow is published, Slack event subscriptions are pointed to the production URL, and a test DM returns a full review on the first try. The speaker notes that n8n’s execution view makes tracing errors easy, while logical bugs require careful planning and agent‑prompt design.