Intent Engineering

Exposes three tools that audit agent intent specs against a 25-item rubric derived from production agent research. Feed it a spec via `audit_intent_spec` and you get a score out of 25, detected anti-patterns like outcome-as-activity or missing stop rules, and concrete fix recommendations. `generate_intent_spec_scaffold` gives you paste-ready YAML templates for blank, level-1-mvr, or full 9-section specs. `assess_retrofit_level` triages existing prompts or SKILL.md files into L1/L2/L3 retrofit buckets based on blast radius and autonomy. Useful when you're about to hand a spec to an agent and want to catch underspecification before it ships something confidently wrong. The server itself scores 23/25 when audited by its own tool.