An AI skill is a wrapped prompt. The trap: you duplicate the prompt in ten places and none is tested. SkillForge treats each skill like business code.
What a skill contains
- A targeted prompt (role, constraints).
- A RAG retriever (pgvector) for document context.
- Guardrails: refusal of forbidden categories, automatic PII redaction on output.
Evaluation before prod
Each skill has snapshots: expected input → expected output. On change, we replay. If output drifts beyond a threshold, promotion is blocked.
Why version
Like code: you roll back, compare two versions, know who changed what. An « improved » skill that worsens a real case shows up immediately.