Skip to content

Routing

This page is about how the agent should interpret plain-language Zotero requests.

It is not a command catalog.

Route by user intent

What the user saysWhat the agent should inferWhat the reply should focus on
“Find papers in my Zotero library about reward hacking”surface candidate items firstwhich items match and why
“Look up Smith2024”direct citation-key lookupmetadata, citation, fallback state
“Show me the last 10 items added to the library”recent-N enumerationwhat was added recently, not keyword matches
“Pull the PDF annotations and notes for this paper”evidence extraction for one itemmetadata, children, annotations, missing capability
“Create an llm-safety workspace”build a long-lived topic workspacenaming, import scope, index prerequisites
“Save this as a Zotero saved search”store reusable search conditionswhat conditions were saved
“Download attachment ATCH005”local attachment downloadwhich attachment, where it was saved
“Add a note to this item”controlled mutationwhat will change and whether write access exists
“Preview a merge for KEEP001 and DUPE001 first”manual merge previewwhat will be filled or moved, and when the write happens
“Show me the current config and default profile”configuration inspectioneffective config, default profile, missing pieces

Quick rule of thumb

  • one item or a few items: treat it as lookup or evidence extraction
  • a long-lived topic set: treat it as a workspace request
  • explicit “last 10” / “recent N” phrasing: treat it as library recent --count
  • “save this filter” or “reuse this later”: treat it as a saved-search request
  • “download the attachment”: treat it as an attachment task, not upload
  • “preview this merge first”: treat it as item merge, not immediate duplicate merge
  • any mutation: go through the safety gate first
  • any “why is this broken?” question: start with doctor or config

When to run doctor first

Run doctor first when:

  • this is a new environment
  • the task will mutate the library
  • the task depends on PDF / outline / annotation / attachment behavior
  • the task needs semantic indexing or workspace query
  • the task depends on citation-key lookup
  • the user is troubleshooting configuration or profile state

Division of labor between skills and CLI docs

  • Skills pages answer: “How should I ask, and how will the agent interpret it?”
  • CLI pages answer: “What command and flags exist underneath?”

Read the skills pages first. Use the CLI pages as reference.