← ClaudeAtlas

cross-modal-reviewlisted

Run a second-model quality gate on staged changes, a skill draft, or any generated output. A cheaper, independent model reviews the primary output for correctness, safety, and completeness before the user sees the final result.
manastalukdar/ai-devstudio · ★ 1 · Code & Development · score 77
Install: claude install-skill manastalukdar/ai-devstudio
# Cross-Modal Review I act as a second, independent reviewer — catching errors, blind spots, and assumptions the primary pass missed. Inspired by gbrain's cross-modal quality gate pattern. ## Token Optimization **Expected range**: 400–1,500 tokens (second-pass review), 50 tokens (trivial change, skip) **Patterns used**: Git diff scope default (staged only), early exit (no reviewable content), progressive disclosure (verdict first → issues → details) **Early exit**: If the target is a single-line change or a comment-only edit, report "Trivial change — cross-modal review skipped" and stop. ## When to Use Invoke after: - `/review` — to get a second opinion on the primary review output - Generating a new skill draft — to validate it passes conformance before writing - Writing complex business logic — to catch logic errors the primary pass may rationalize - Security-sensitive changes — to add a second set of eyes before commit Can also be invoked directly: `/cross-modal-review` with staged changes as the default target. ## Step 1 — Identify Target ```bash # Default: staged changes git diff --cached --stat git diff --cached --name-only # If no staged changes, check for an argument (file path or prior output) [[ -z "$ARGUMENTS" ]] && git diff --cached --quiet && echo "Nothing to review" && exit 0 ``` ## Step 2 — Frame the Second-Pass Review Review the target with deliberate adversarial framing — assume the primary pass was optimistic. Check specifically: **Correctness*