Legal War Game

Legal War-Game (法律战沙盘)

Red-teams a legal/terms page by running open-ended mock litigation against it. The goal is to find which clauses actually hold up under adversarial oral argument — and which get torn apart — so you can harden the text.

The deliverable is a per-scenario report: full case file (起诉状 → 答辩状 → 庭审记录 → 判决书) plus a loophole review (复盘).

The driver is war-game.js

.claude/skills/legal-war-game/war-game.js is a Workflow script. Run it with the Workflow tool, passing the inputs via args. You do NOT edit the script to change the scenario — you pass different args.

How to run (the agent path)

Call Workflow with scriptPath pointing at the script and args carrying the four inputs:

Workflow({
  scriptPath: ".claude/skills/legal-war-game/war-game.js",
  args: {
    owner:    "<你方身份画像 — 只给你方/法官/复盘官, 不给对方>",
    legal:    "<你的法律/条款全文摘要 — 同上, 保密>",
    scenarios: [
      {
        key: "vcg-infringement",
        title: "视觉中国图片侵权索赔之诉",
        claimantBrief: "<对方律师视角: 他是谁、要什么、手里有什么牌. 不给他看你的条款>",
        claimantGoal:  "尽可能拿到赔偿, 用惯用的批量维权与举证施压打法.",
        respondentGoal: "驳回或最大限度压低索赔. 手持你方条款, 可援引防御."
      }
    ],
    law: "中华人民共和国法律",      // optional, default
    issueCount: 3,                   // optional, default 3 焦点
    hardCapPerIssue: 40              // optional, 仅防脚本失控
  }
})

It runs in the background; you get a task-notification with the JSON result when done.

The four inputs (what to fill)

• owner — who YOU are (the defendant / clause-holder), with the facts that matter legally (e.g. minor + legal guardians, non-profit, student). Given to your side, the judge, and the reviewer — never to the claimant.
• legal — the full text/summary of your clauses, translated into the bracketed summary form. Same visibility as owner. This is the "shield" being tested.
• scenarios — one entry per adversarial situation. For each:
  • claimantBrief — the opposing lawyer's point of view ONLY: who they are, what they want, what evidence they hold. Do not leak your clauses or identity here — blind opposition is the whole point (a real plaintiff doesn't get your terms up front).
  • claimantGoal / respondentGoal — what each side is trying to achieve.
• law — governing legal system for the judge.

The process it runs (why it's built this way)

1. Pleadings (书面诉答) — claimant drafts 起诉状 blind; you draft 答辩状 holding the clauses.
2. Hearing (开放庭审) — judge frames N issues, then loops each issue: claimant argues → you defend → judge rules one of {继续追问 / 本焦点查明 / 本焦点查明(制止重复)}. The issue only closes when the judge declares it resolved — no fixed round cap. This is what separates real oral debate from one-shot document ping-pong.
3. Verdict + Review — judge rules on each clause's validity and the case; a reviewer then separates "fixable by editing clauses" from "only fixable by ops/evidence", and refuses to over-engineer for situations you won't enter.

Lessons baked in (from 6 live iterations — don't relearn these)

• Blind the claimant. Never put your clauses/identity in claimantBrief. Leaking them makes the plaintiff superhuman and the test useless.
• Let the judge control the loop. A fixed round count produces repetitive "each side restates" debate. The judge-resolves-it loop is what surfaces real weak points (it once chased ownership for 7–8 rounds into RAW-fingerprint forensics).
• Read the review for the split that matters: clauses only make you "not lose"; facts/evidence make you "win". Most high-risk leftovers are ops problems (source attribution logs, response-time discipline, traffic data), not text.
• Don't over-engineer. Reviewer is instructed to skip clauses for situations you won't enter and to reject offensive/aggressive clauses you can't back in court.

Output handling

The result is an array of { key, title, caseFile, review }. Archive it to _drafts/legal-war-game-<n>.md (or wherever the user keeps drafts) and summarize the verdicts + the fixable-vs-ops loophole split. Commit only when the user says to.