sitefire-prompts

Installation
SKILL.md

Manage Sitefire prompts

Use the server's shared instructions and current tool contracts. The prompt-management guide owns the product explanation and selection guidance; do not recreate its parameter schemas here.

Start by inspecting the relevant setup with list_topics and list_prompts. Follow pagination when the user's scope exceeds one page. Keep lifecycle and generation state separate: an active prompt can still be generating. Discover labels and optional personas through list_filter_values.

For “what should we track?”, check existing runs with get_prompt_research. Reuse suitable completed research; otherwise start the authorized business area and market through start_prompt_research. Keep its request identity and run ID, follow the suggested polling interval, and review recommendations before adopting selected IDs through add_prompts. Starting or completing research is not activation. Research run discovery works in a later conversation.

Choose a coherent set of useful questions, accounting for ready and pending coverage and capacity. Apply the server's 3–7 recommendation as guidance, not a quota to fill with paraphrases. Personas are experimental; do not turn missing assignments into cleanup work.

When the user supplies exact customer questions, use manual add_prompts. Keep their wording and distinguish supplied evidence from proposed questions. Do not imply that research accessed CRM records or call notes unless those sources were actually provided or read. add_topics is the quick-generation alternative for known subjects; its existing-topic default skips. Request an explicit increment only when more questions are intended.

For topic consolidation, duplicate review, wording improvements, or retiring irrelevant monitoring, read the cleanup workflows. Use create_tags, update_tags, and retire_tags for label definitions. Topic renaming through update_topics changes the search-volume basis; prefer a tag for an internal label. Stop topic monitoring by archiving its prompts. Use the experimental persona tools (create_personas, update_personas, delete_personas) only when requested, and follow the delete preview and confirmation contract of delete_personas.

A preview does not consume its request identity: apply the reviewed business changes with the same request_id and dry_run: false. Keep that applied input and identity unchanged when retrying an uncertain apply. A new identity means new intent. Do not turn a conflict, failed generation, or unknown tracking overlay into an automatic new batch. Inspect the returned state and take the tool's recovery path. Re-read saved prompt IDs to verify texts, topics, states, and assignments; report pending work separately from completed changes.

Installs
1
First Seen
6 days ago