← ClaudeAtlas

jurylisted

Run another skill N times in parallel and merge the outputs into a frequency ledger - surfaces non-determinism and unions findings across runs. For high-stakes reviews and experiments.
sgtlaugh/skills · ★ 2 · Code & Development · score 76
Install: claude install-skill sgtlaugh/skills
# /jury - Repeat a skill N times, merge into a ledger Runs a target skill N times concurrently in **independent** contexts, then merges the per-run findings into one comparison table: rows = deduped findings, cells = each run's severity, plus an **`X/M` frequency** column (of N launched, S succeeded, M full-fidelity). Surfaces recall variance and unions what any single run would miss. Best for skills that emit discrete, file-anchored findings (e.g. a code-review skill). ## Usage ``` /jury <skill-name> <its-args> [--n N] # default N=3, min 2, cap 6 /jury <skill-name> <its-args> --headless # jurors as root processes (for targets that self-fan-out) /jury --merge <runs-dir> # re-merge saved raws, no re-run ``` `<skill-name>` is **whatever skill you want repeated** - read it from the input, never assume a particular one. It's a skill name + its args, not a nested slash-command (a leading slash is just stripped). The harness only ran `/jury`; everything after is text for this skill to act on by fanning out (Step 3), never by invoking the target in this session. ## Step 1 - Parse - `--merge <dir>` → skip Steps 2-4; go straight to Step 5 reading the raws in `<dir>`. No target is parsed, so the fidelity gate uses its majority fallback and the overview says so. - Otherwise: **the target name is the first non-flag token after `/jury`, copied verbatim from the input.** Do NOT replace it with any skill named elsewhere in this document - copy exactly what was typed (