菜单
Open a gated, multi-round critique episode with a different-model subagent that attacks your assumptions before you commit to a consequential, uncertain decision. Use on "pair", "spar", "rubber duck", "get a second opinion", "discuss amongst yourselves", or "consult another model before committing".
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/lossyrob/skills --skill spar命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Drive a backlog of GitHub issues to PRs autonomously and sequentially. The loaded session becomes an orchestrator that triages issues into S/M/L tiers, spawns PAW implementer (and optional PAW Review) worker terminals that coordinate over telex instead of GitHub-comment polling, gates each PR through a preference/human-floor merge review, and auto-merges or routes to human review. Use when asked to work through a backlog/queue of issues autonomously, run an autonomous issue-fixing pipeline, or orchestrate PAW sessions across many issues.
Launch a new terminal running Copilot CLI with a requested title, color label, and working directory. Supports macOS iTerm2/Terminal.app and Windows Terminal, prompt-driven sessions, existing-session resume, explicit terminal selection, and new/current window targeting.
Convene a bounded multi-model, multi-perspective deliberation to produce a recommendation with confidence, preserved dissent, and a transcript. Use on "council", "panel", "deliberate", "multi-agent review", "advisory council", "get diverse perspectives", or "discuss this with specialists".
正在显示 SKILL.md
基于 SOC 职业分类
| name | spar |
| description | Open a gated, multi-round critique episode with a different-model subagent that attacks your assumptions before you commit to a consequential, uncertain decision. Use on "pair", "spar", "rubber duck", "get a second opinion", "discuss amongst yourselves", or "consult another model before committing". |
Get a sharp second opinion from a different model before you commit to a consequential decision. This is a pairing skill with one divergence: between agents there is no ego to protect, so the best pairing is sparring — the partner is adversarial in service of the work. Keep the structure of pairing — a shared goal and the driver holding the pen — and drop the ego-protecting hedging that turns review into a rubber stamp.
Open one when you are about to make, revise, or abandon a consequential commitment under uncertainty. That is the whole rule; the cases below just instantiate it:
Do not spar on routine coding, obvious fixes, mechanical refactors, ordinary test failures, or "please review this diff." Opening episodes for trivial or poorly scoped questions is the failure mode — it produces latency theater and trains everyone to rubber-stamp. Open few episodes; make each one deep enough to change your thinking.
Run the duck with the task tool as a rubber-duck agent, overriding model to
the pairing model. For a real back-and-forth, run it in background mode and use
write_agent to send follow-up rounds, so the duck retains the episode's context
across turns. If your Copilot CLI build does not offer a rubber-duck agent type,
use a general-purpose subagent instead and put the sparring role (attack
assumptions, find failure modes, do not validate) directly in its prompt — the
mechanics are identical, only the agent type changes.
If the caller does not specify a mode, infer it: a decision you own → integrate; a decision the operator owns, or an explicit request for a brief → brief-back.
Callers (prompts or operators) may layer context-specific instructions on top of this skill: add domain triggers, pin a pairing model, force a mode, or point the episode at a specific artifact. The mechanics above do not change.
