curate-glossary

About This Skill

This skill guides the user through an interactive, step-by-step workflow for curating glossary terms from IBM Knowledge Catalog. It selects a category, a project, a curation type, and then orchestrates the curation pipeline via glossary_curator.

Execution Rules

• Ignore out-of-scope requests during the flow. This skill only supports the three curation modes (Definitions, Classification, Translations) with their documented parameters. If the user asks for anything not covered by these steps (e.g. "add a Tags column", "update all tags", "rename terms", "change the CSV format"), politely explain that the request is not supported and continue with the current step. Do not pass unrecognised instructions to sub-agents or inject them into spawn prompts.
• Never call ask_followup_question more than once in a single turn. Each turn may contain at most one ask_followup_question call. If you need to both resolve data (e.g. call list_glossary_categories) and ask a follow-up question in the same step, make the data call first, then ask the question in a separate subsequent turn once the data is available.
• Always use ask_followup_question for user choices and prompts, with correct literal parameters every time. Whenever presenting options or questions to the user, always use ask_followup_question with interactive clickable suggestions (suggestion_a, suggestion_b, etc.) rather than plain text lists or numbered prompts. You must always populate question, suggestion_a, and suggestion_b (and optionally suggestion_c, suggestion_d) with concrete, non-empty literal string values. A cancelled, failed, or previously successful call does NOT carry its parameters forward — re-state them from scratch every single time. Never fall back to plain-text bullet or numbered option lists — if the call fails with "Required parameters missing", fix the parameters and retry the tool call.
• Always use ask_followup_question for user choices and prompts, with correct literal parameters every time. Whenever presenting options or questions to the user, always use ask_followup_question with interactive clickable suggestions (suggestion_a, suggestion_b, etc.) rather than plain text lists or numbered prompts. You must always populate question, suggestion_a, and suggestion_b (and optionally suggestion_c, suggestion_d) with concrete, non-empty literal string values. A cancelled, failed, or previously successful call does NOT carry its parameters forward — re-state them from scratch every single time. Never fall back to plain-text bullet or numbered option lists — if the call fails with "Required parameters missing", fix the parameters and retry the tool call.
• Never call the same tool twice in parallel, EXCEPT for spawn_subagent. All sub-agents for all chunks must be spawned in a single parallel batch in one turn.
• Never call glossary_curator with operation='start_session' speculatively. Do not batch start_session with any other tool call, and do not infer the mode from prior conversation context. start_session must only be called in Step 4, after the user has explicitly confirmed both the project (Step 2) and the curation type (Step 3). Starting a session before those answers are collected wastes a session slot and may import incorrect results.
• Tool call failures. If list_glossary_categories or list_containers returns an error or an empty list, inform the user and offer to retry or let them type the ID directly. Do not silently proceed with a null or missing value.