harden-skilllisted

# Harden a skill Use this skill to make a skill's *wording* bind. Authoring a skill is only half done when the prose reads well; the other half is whether an agent under load actually does what the prose intends. This skill finds where the guidance will fail under pressure or be rationalized away, and prescribes the instruction form that holds. Governing intuition: **you cannot harden a behavior you have not named, and you should not add guidance for a failure you have not seen.** State the behavior each passage is trying to produce, the failure an agent falls into without it, and match the form to that failure — do not reach for a prohibition by reflex. It produces a hardening plan: per-behavior verdicts, recommended rewrites, and a verification protocol the agent (or user) then applies as a separate act. **Input**: the skill instructions under authoring or revision — a `SKILL.md` (draft or existing), a specific rule or section, or a proposed wording change; if none is given, ask which skill or passage to harden. **Boundary**: diagnoses and prescribes only — produces a hardening plan; does not edit the target skill or run the tests. ## Workflow ### Phase 1: Inventory the intended behaviors Walk the target prose and, for each passage of guidance, write one line: *what agent behavior is this trying to produce?* Guidance that maps to no concrete behavior is decoration — flag it. A passage may carry several behaviors; split them, because different behaviors may need diffe