← ClaudeAtlas

rjv-gated-buildlisted

Use for long, high-stakes multi-step builds (financial/money/prod) where an INDEPENDENT reviewer grills and approves each step before it ships. Gated loop, role casting across models, compressed anchor doc, crash/compaction durability, model economy (cost-routing hard rule). Triggers: 'gated build', 'grill each step', 'adversarial review loop', 'overnight run', 'high-stakes build'.
rjvim/ai-skills · ★ 0 · AI & Automation · score 63
Install: claude install-skill rjvim/ai-skills
# Gated build — adversarial multi-agent construction For work where being *plausibly wrong* is expensive. One agent can't reliably catch its own errors — the blind spot that produced the bug reviews the bug. Quality comes from **independence**: a second agent whose job is NO-AGREE, grilling every step to an explicit APPROVED. Economics: compress everything agents read (specs, notes, grill prompts) so you can afford a second agent to disagree. Efficiency funds redundancy. ## 1. Casting the roles Three roles: **Orchestrator** (holds anchor doc, casts, integrates), **Author** (drafts), **Reviewer** (grills to APPROVED). Cast by task, not habit. **Orchestrator — the human casts it at launch. ASK, don't assume.** You start the run, so you pick the driver; it records once in `## Cast` and every resume obeys it — no re-negotiation. Choose by run-length × stakes: | Orchestrator | Pick when | Reviewer it then requires | |---|---|---| | **Opus** (flagship) | short / interactive, or design-heavy where the driver's own judgment carries the loop | may self-review MECHANICAL steps only; high-stakes still gets an independent cross-vendor reviewer | | **Sonnet** (mid) | long autonomous / overnight — the all-night loop at ~1/5 Opus per-turn cost | never self-reviews non-trivial work → cast an **Opus** reviewer; live-money adds a **Codex** cross-grill | | **Codex** (external CLI) | repo-heavy mechanical drives; or when you want the **zero-cost cheap-lane only Codex has** (local-Ollama ag