From quorum
Use when the user types /turbo-review or asks to convene multiple external models at once for a task, research, or review — a "quorum" pass. Fans the same prompt out to codex, opencode, and a fresh claude in parallel (read-only, in the background), then synthesizes their verdicts — flagging consensus (high confidence) and disagreement (needs scrutiny). Use proactively before committing to a risky design, plan, or diff.
How this skill is triggered — by the user, by Claude, or both
Slash command
/quorum:turbo-reviewThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Run **all available external models at once** and synthesize them into one answer. Instead of asking a single outside model, fan the same task out to `codex` (GPT-5.5), `opencode` (DeepSeek/Kimi/Qwen), and a fresh `claude` in parallel, then reconcile their verdicts. Where they agree, you have high confidence; where they disagree, you have a precise list of things to scrutinize.
Run all available external models at once and synthesize them into one answer. Instead of asking a single outside model, fan the same task out to codex (GPT-5.5), opencode (DeepSeek/Kimi/Qwen), and a fresh claude in parallel, then reconcile their verdicts. Where they agree, you have high confidence; where they disagree, you have a precise list of things to scrutinize.
Use for: pressure-testing a design/plan, reviewing a diff before merge, or a research question worth more than one model's view. Triggered by /turbo-review "...", "convene the quorum", "ask everyone", "запусти всех", or proactively before a costly/irreversible decision.
This skill composes the sibling skills codex, opencode, and claude — see each for per-CLI flag details.
The external CLIs run read-only — they review, research, and propose, but never touch files. You (the host agent) synthesize their input and make any edits yourself. This keeps the diff coherent and conflict-free, and it's portable across host agents. Grant an external model write access only if the user explicitly asks (-s workspace-write for codex, --permission-mode acceptEdits for claude) — and then to only one of them.
Write a single, self-contained prompt every model will receive. Include:
Launch every available CLI at once, each writing to its own file. Skip any CLI that isn't installed — degrade gracefully (two views still beat one). Use a shared <tag> for the run:
# codex — GPT-5.5, read-only
codex exec -s read-only -o /tmp/quorum-<tag>-codex.md "<prompt>" > /tmp/quorum-<tag>-codex.log 2>&1 &
# opencode — free opencode Zen model by default (swap to a premium one you've configured)
opencode run "<prompt>" --model "opencode/deepseek-v4-flash-free" > /tmp/quorum-<tag>-opencode.log 2>&1 &
# claude — fresh headless context, read-only (plan mode)
claude -p "<prompt>" --permission-mode plan --output-format text > /tmp/quorum-<tag>-claude.md 2>&1 &
wait # block until all of them finish
run_in_background: true so you're re-invoked as each finishes; then read each output file. The & … wait form above is the portable plain-shell equivalent for any host.<tag> per run. Each model can take minutes — running them concurrently makes the wall-clock the slowest single model, not the sum.Read every output file:
/tmp/quorum-<tag>-codex.md (verdict) / .log (if it failed)/tmp/quorum-<tag>-opencode.log (verdict is at the end)/tmp/quorum-<tag>-claude.mdProduce one consolidated answer:
You (the host) implement the agreed changes. The external models stay read-only.
claude pass is a fresh-context check rather than a different model; still useful, but codex + opencode add the most diversity.Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub alesha-pro/quorum --plugin quorum