auditlisted

Analyses code from several independent perspectives, then makes those perspectives argue with each other until they converge on a shared verdict. Each lens reviews blind, every finding is cross-examined by a lens that wants to kill it, conflicts are resolved as explicit trade-offs rather than averaged away, and the result is a ranked findings table with recorded dissent plus a full analysis. Use when the user says "/audit", "audit this code", "red-team this", "review this code from every angle", "review this code", "critique this", "analyse this code", "what's wrong with this", "look at this from every angle", "is this any good". For rewriting use rewrite, for improving in place use refactor, for UI use reskin. This skill judges code — for measuring data and writing the result up as a report, use report instead.
olcayseygan/crumora · ★ 0 · AI & Automation · score 72

Install: claude install-skill olcayseygan/crumora

# audit — put it on trial Read the code through **several independent pairs of eyes**, then put those eyes **in a fight**, and report only what survives. Fourth sibling of **rewrite**, **refactor** and **reskin**. Those three build; this one judges. Same house style: rounds, evidence over vibes, an explicit table at the end, and no pretending to be more certain than the evidence allows. The two failure modes it exists to prevent: - **The single-lens review** — one pass, one mindset, and everything outside that mindset is invisible. A performance reader never notices the lifecycle bug. - **The confident hallucination** — a finding that reads beautifully and is simply wrong. Nothing ships to the user until another lens has actively tried to kill it. --- ## 0. Pick the target and the panel If the user passed an argument, that is the target (`/audit src/parser.ts`, `/audit the working diff`). If not, ask **one question**: what to review — and default to the uncommitted diff if the repo has one. **Read the target completely before opinions start.** Not excerpts. If it is a diff, read the surrounding code too — a diff that looks fine in isolation and wrong in context is exactly what a review is for. Then **pick the panel**: 4-7 lenses, chosen for this target, declared up front and frozen. A default panel: | Lens | Hunts for | | --- | --- | | **Correctness** | Wrong results, off-by-one, bad state transitions, unhandled inputs, races, edge cases | | **Lifecycle & robust