← ClaudeAtlas

prismlisted

Run an independent read-only Claude Code review from Codex when the user requests Claude review or cross-model adversarial review. Return findings without automatic fixes or review loops.
SWHee/skills · ★ 0 · AI & Automation · score 56
Install: claude install-skill SWHee/skills
# Prism Use the bundled `scripts/review.py`, resolved relative to this SKILL.md's real path. The host is Codex; Claude is the reviewer. Do not substitute a Codex review if Claude fails. ## Run - Setup request: `python3 <skill>/scripts/review.py --check`. - Default: `python3 <skill>/scripts/review.py --cwd <repo> --mode review`. - Adversarial request: add `--mode adversarial` instead of `--mode review`. - Branch review: add `--base <user-specified-ref>`. Default scope is tracked staged/unstaged changes plus untracked files. A clean working tree does not implicitly select another branch. - Forward a requested focus with `--focus <text>` and the user's language with `--language Korean` (or the appropriate language). - Model is pinned to `claude-opus-5-5`. Never substitute Sonnet or another model, including on failure. Changing the model requires a new user request and implementation change. - Choose `--effort high|xhigh` in the current Codex turn, using the task context and a brief inspection of the scoped diff. No classifier, subagent, or extra model call is needed. Default to high for localized changes with clear contracts. Choose xhigh for nontrivial concurrency, authorization/trust boundaries, irreversible data/migration behavior, or cross-component design interactions that require tracing multiple failure paths. File count alone is not sufficient. Honor an explicit high/xhigh request. If uncertain, use high; never escalate after seeing findings automatically. - Before ex