Template Recategorization: MCP as Hermeneutic Bridge

╔══════════════════════════════════════════════════════════════════╗
║  TEMPLATE RECATEGORIZATION                        ✓ VALIDATED    ║
║  ────────────────────────────────────────────────────────────    ║
║  8 templates │ 5/5 criteria │ 0 rollbacks │ ~45 min total        ║
║                                                                  ║
║  ┌──────────┐     ┌──────────┐     ┌──────────┐     ┌─────────┐  ║
║  │  SLACK   │     │  CLAUDE  │     │   MCP    │     │AIRTABLE │  ║
║  │  THREAD  │ ──► │   CODE   │ ──► │  BRIDGE  │ ──► │   DB    │  ║
║  └──────────┘     └──────────┘     └──────────┘     └─────────┘  ║
║                                                                  ║
║    Human           Agent          Protocol         System        ║
║   (intent)     (interpretation)   (access)        (state)        ║
╚══════════════════════════════════════════════════════════════════╝

Hypothesis

AI agents equipped with MCP (Model Context Protocol) tools can participate in
thehermeneutic circleof data curation—where understanding the whole
(category taxonomy) and the parts (individual templates) mutually inform each other.

Context: The Problem

Webflow templates in our CMS were incorrectly categorized under "Public Services"—a
category intended for government entities (police departments, fire stations,
municipal utilities, political offices).

Instead, templates for dentists, design agencies, law firms, and digital nomad
lifestyles had accumulated there. This misalignment degrades user experience:
someone browsing for government templates finds lifestyle coaches.

Method: MCP Integration

The Model Context Protocol (MCP) allows Claude Code to interact directly with
external systems. By configuring the Airtable MCP server, the agent gains:

{`// ~/.claude.json
{
  "mcpServers": {
    "airtable": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "airtable-mcp-server"],
      "env": {
        "AIRTABLE_API_KEY": "pat..."
      }
    }
  }
}`}

Grounding: The Hermeneutic Circle

Data curation is inherently hermeneutic. To categorize a template correctly, you must:

1. Understand the category taxonomy— What does "Public Services" mean?
2. Examine the individual template— What does this template represent?
3. Return to the whole— Does this template belong in this category?
4. Refine understanding— Should the category definition change?

"The part can only be understood from the whole and the whole only from the parts."
— — Wilhelm Dilthey, on the hermeneutic circle

Traditional database queries break this circle: you either query by category
(ignoring template content) or by template (ignoring taxonomy structure).
MCP-enabled agents can traverse both simultaneously through conversation.

Workflow Demonstration

Success Criteria

Tool Complementarity in Practice

This experiment demonstrates the complementarity principle from the CREATE SOMETHING
philosophy: Claude Code excels at interpretation and creation, while humans provide
context and verify results.

Subtractive Triad Application

Technical Learnings

Results

The experiment successfully recategorized 8 templates from the "Public Services"
category. Each template was reassigned to appropriate categories that accurately
reflect its content and intended audience.

Limitations

What This Proves / Doesn't Prove

Reproducibility

• Claude Code CLI installed
• Airtable account with API access (Personal Access Token)
• airtable-mcp-server npm package
• Knowledge of target category definitions

1. Configure MCP server in~/.claude.json(see Configuration above)
2. Restart Claude Code to load MCP tools
3. Identify miscategorized records via human review or query
4. Query category table to obtain linked record IDs
5. Update records with corrected category IDs
6. Verify changes in Airtable and downstream systems

• MCP server may not load from project-level config—use global config
• Linked record fields require ID lookup, not human-readable names
• Category definitions may be tribal knowledge—document them first

Conclusion

The hypothesis isvalidated. Claude Code with MCP tools
can participate meaningfully in the hermeneutic circle of data curation.
In this experiment, 8 templates were successfully recategorized through
conversational interaction—demonstrating that AI agents can:

• Query and interpret data (understanding parts: individual templates)
• Apply category definitions (understanding whole: "Public services = government")
• Propose and execute corrections (synthesis: remove miscategorizations)
• Learn from human feedback (iteration: preserve appropriate secondary categories)

"The bridge is a thing that gathers."
— — Heidegger, Building Dwelling Thinking

MCP serves as a bridge—gathering human intent, agent capability, and system
state into a unified workflow. The protocol recedes; the curation emerges.