context

Route based on the first token in $ARGUMENTS:

• list — qmd context list. Show all configured contexts grouped by collection.
• add [path] "<text>" — qmd context add $ARGUMENTS.
  • One-arg form (add "text") uses cwd and auto-detects the collection.
  • Two-arg form takes an explicit path. Path formats:
    • qmd://collection/path — virtual path inside a collection (recommended)
    • / — global context, applies to every collection
    • . or relative path — auto-resolved from cwd
    • Absolute filesystem path — collection detected by longest-prefix match
  • Collection names must match [a-zA-Z0-9_-]+.
• rm <path> (alias remove) — qmd context rm $ARGUMENTS. Accepts both filesystem paths and qmd:// URIs.

Context is hierarchical: when a search hits a path, qmd concatenates ALL matching prefixes (global → root → specific subdir), joined with \n\n. Adding context at multiple levels is additive, not overriding.

What context actually does: the qmd MCP server returns the document's full hierarchical context block in every search result's context field. Claude (and other LLM callers) sees the context alongside the snippet so it knows what kind of document it's reading. Context is purely descriptive metadata. It does NOT influence chunk selection, reranking, or scoring — those are driven by the query terms and the optional intent parameter, not by the YAML context: field.

So context's job is to TELL the model "this snippet comes from a Convex schema example, not a tutorial blog post." A short, specific context (one sentence) is better than a long one because the full string gets prepended to every result's snippet.

The /qmd:add SKILL writes the initial root context once at clone/register time. Use this command to update the root context or add deeper paths later (e.g. qmd://next.js/docs/api → "Stable App Router API reference").

If $ARGUMENTS is empty or the subcommand is unrecognized, print: usage: /qmd:context <list|add|rm> [args].