configure
Installation
SKILL.md
Global Rules
- Execute silently between tool calls. Do not output planning, progress, transition, waiting, or tool-result narration between calls. Execute tool calls silently and proceed directly to the next call. Parallelize independent tasks by batching tool calls into one turn whenever possible. Before the final output, speak only when the skill explicitly requires user input, approval, an exact notice, or material error/blocked reporting. Do not invent checkpoints.
- Keep the final output concise. Return only the requested result or deliverable. Omit process recaps, tool-call details, redundant preambles or conclusions, and data already shown in a widget.
- Ground dynamic or custom relationship and field names before relying on them. Fixed standard fields that this skill explicitly marks as requiring no grounding need no extra grounding call. On a name error, use the skill's documented grounding path when present; otherwise report the error instead of guessing or re-firing the same shape.
- Cite every value exactly as queried; never fabricate; distinguish a blank value from a value that was not queried. Link each Salesforce record inline:
https://[instanceUrl]/lightning/r/[SObjectType]/[Id]/view. - Show human labels, never API/field literals. In anything the user sees, print each field's grounded
label(for example, "Deal Risk", notDeal_Risk__c) and record Names, never raw Ids or__cAPI names. - NEVER use
discoverordescribe, and never call an API or endpoint not written in this skill. Every Salesforce URL you need is in the skill. Don't guess REST paths: on a 404 or unknown-path error, fall back to a documented query in the skill, not to discovery. If you need a capability such as email, docs, Slack, calendar, or web research, use the other connector/MCP tools already available to you. Endpoint guessing and discovery add needless round-trips. Use only the skill-authorizeddispatch_readonlyanddispatchcalls, directly with the queries given.
Rules:
- Empty
MINEscope is not a failure. A brand-new user legitimately has no activity — record "none yet", never invent records. - Artifacts can only call MCP tools you declare in their
mcp_tools, and the host blocks any tool you did not actually call this session. So (1) resolve the EXACT fully-qualified tool name in-session before creating an artifact — do not guess or list candidate variants; (2) declare only tools you verified this session inmcp_tools; (3) never make an artifact depend on a server that may not be connected. Your Sales Intelligence view calls exactly one tool (the read-only dispatch tool).