Skip to main content

Agent authoring

FlipCard is designed to be authored by humans or LLM agents. The @microsoft/flipcard-mcp server gives any MCP‑capable host a small, opinionated tool surface for producing valid manifests.

Configure the host​

Add the server to your MCP client (Copilot Studio, Claude Desktop, VS Code MCP, etc.):

{
  "mcpServers": {
    "flipcard": {
      "command": "npx",
      "args": ["-y", "@microsoft/flipcard-mcp"]
    }
  }
}

A reliable agent loop for producing a card looks like this:

  1. Discover — call list_categories to learn the seven presets and their intent. Do this once per session.
  2. Plan — pick the category whose intent matches the user's request.
  3. Generate — call generate_manifest with the chosen category and any provided titles.
  4. Refine — edit design.front, schema.back.fields, and workflow.actions in place.
  5. Validate — call validate_manifest on the final draft.
  6. Persist — write the JSON wherever your application stores cards.

Prompt scaffolding​

A short system prompt that works well:

You are a FlipCard authoring assistant. When the user asks for a card:

  1. Call list_categories once to ground yourself.
  2. Pick a category whose intent matches — never the closest color.
  3. Call generate_manifest with category, frontTitle, and backTitle.
  4. Edit the result so design.front.summary is one sentence and schema.back.fields reflects the data the user described.
  5. Call validate_manifest and report the result.
  6. Output the final JSON in a fenced json block.

Quality bar​

Before returning a manifest:

  • id is stable, kebab‑case, and unique within the user's collection.
  • title is human‑readable, not a duplicate of id.
  • design.front.title is short enough to fit a card header.
  • design.front.summary is one sentence and adds information the title doesn't.
  • The category matches intent, not color preference.
  • validate_manifest returns { valid: true }.

See also​