fusion-advisorlisted
Install: claude install-skill muslewski/llm-armory
# Fusion Advisor (two-model peer fusion at the reasoning layer)
## Overview
Turns your advisor session (Claude Code / Opus) into a **two-model advisor**. At a high-leverage
decision point, the advisor consults **Grok** (different vendor → different blind spots) as a
context-isolated peer, reconciles the two positions by **verifying claims — not voting**, and
emits the advisor artifact: a plan, a `/loop` decision, a review. Execution then flows down to
grok/sonnet executor children via the `armory` launcher (`bin/llm`). **Fusion decides; armory executes.**
Naive "two models beat one" is false (see Guardrails). Fusion pays off only on hard judgment,
and only with structure. This skill IS that structure.
Advisor half stays here (frame, reconcile, verify, decide). Peer half is one isolated Grok
consult per invocation. Never a debate club.
**Requires:** the `grok` CLI (Grok Build) and this repo's `armory` / `bin/llm` launcher. The
advisor is Claude Code (Opus); either peer can be swapped for another frontier CLI.
## When to use / when NOT (cost gate)
**Use at high-leverage points only:** design forks, plan approval, a genuine "am I sure?"
moment, `/loop` course-corrections, risky diffs (auth, billing, migrations, concurrency,
cross-zone). This is the same bar you would apply to an adversarial review.
**Do NOT use for:** routine or mechanical steps; execution dispatch (that is the `armory`
launcher); or as a second code executor. Fusion costs 2–10× tokens and over-fusing
in