Declare what must never happen in your AI workflow, find the silent failures that would let it happen, harden the action boundary, and prove each fix under injected faults with faultkit, a local fault-injection proxy (needs faultkit installed: Homebrew, AUR, or go install; no telemetry).
Skills 6 in the Claude listing
harden Use when asked to make an AI or LLM workflow resilient, add a guard, policy check, or safety boundary around an irreversible action, prevent an agent from ever doing something, or fix a finding from a prove Generates or selects a faultkit scenario for a business invariant, writes the deterministic gate test if the project lacks one, uses the installed faultkit, runs it locally, and reports the proof stat Prove all Finds every business invariant in the project, keeps each as a faultkit scenario and gate in .faultkit/invariants/ (or as not_generated when no fault expresses it yet), runs them locally, and reports review Use when asked to assess, review, audit, or explain the resilience of an AI or LLM workflow, agent, or tool loop; when someone mentions silent failures, business invariants, what happens when the mode run Same as prove-all. values Declares the business value and the unacceptable outcomes of an AI workflow in .faultkit/values.md, from the user's own words or from a review's draft, and writes the file only after the user says yes