← ClaudeAtlas

code-reviewlisted

Reviews and audits existing Roblox/Luau code instead of writing it. Use when the user asks to review a file, diff, or pull request, to audit a place or system, to score architectural health or maturity, to find leaks, exploits, or data-loss risks in code that already exists, or asks how good or how safe their current implementation is. Reports one severity per finding (Blocker / Correctness / Advisory) behind a confidence gate, and scores a whole project 1-5 across security, lifecycle, performance, and replication. Not for writing or refactoring code — hand that to the best-practices skill; not for chasing a reported symptom down to its cause when no file has been named yet — that is the diagnose skill; not for Studio tooling, sync, or playtest questions - that is the studio-ops skill.
andrian-syh/roblox-optimum · ★ 1 · Code & Development · score 77
Install: claude install-skill andrian-syh/roblox-optimum
# Roblox Code Review Judging code that already exists. Nothing here writes code — the moment the task turns into producing or changing Luau, the **best-practices** skill owns it and this one steps aside. **Goal, in priority order:** true → severe-in-the-right-order → actionable → brief. A wrong finding costs more than a missed one. Trust lost to a false positive does not come back on the next review. ## Session Invariants (must survive compaction) ```text ROBLOX CODE REVIEW - INVARIANT CARD 1 One severity per finding, chosen after the confidence gate: Blocker - security, data loss, a guaranteed leak. Correctness - a real bug with a concrete failure scenario. Advisory - style, layout, micro-optimization. 2 CONFIDENCE GATE - all four before a finding is reported: - Trace both sides of paired logic (the create and the destroy, the fire and the receive). Half a trace is not a finding. - Assume the odd shape may be intentional until the trace says otherwise. - Demand a concrete failure scenario: inputs or state, then the wrong outcome. "Could be a problem" is not a finding. - Verify the API against the target environment, not against memory. 3 Read the false-positive catalog BEFORE reporting. Every rule has shapes that only look like violations. 4 Advisory items are PROPOSED, never applied. Unrelated code is never reformatted. Nothing is refactored on this skill's initiative. 5 Engine facts are cited or marked unverified. Ne