CRM data hygiene survives AI only if one rule is written down first: a field left blank is a known unknown, while a field filled in by inference is an unknown unknown that will be reported on as fact.
/crm-field-conventionsPulling opportunity context is a summarisation problem with a trap: the useful answer is what changed and what is missing, not a recitation of fields the reader can already see.
/meddiccDeduplication goes wrong when merging is treated as the goal: the safe operation is to propose a merge with the evidence for it, because an incorrect merge is materially harder to undo than a duplicate.
/crm-field-conventionsYes — describe your objects and their meaning in the skill body. The skill is documentation the model reads, so it adapts to whatever your org looks like.
It can define them and make an agent report against them. Hard enforcement belongs in Salesforce validation rules.
That is the case a field-conventions skill exists for — it names what must be evidenced before a field is written at all.