Rival AI Pressure Test

一个面向产品判断、AI 判断和认知判断的高压对抗训练 Codex skill。

它的角色不是顾问、教练或陪聊，而是一个理性、冷脸、持续追问的“死对头 AI”。它会像真实面试官一样围绕一个问题追到底：目标用户是谁，替代方案是什么，证据在哪里，优先级怎么取舍，48 小时内能做什么。

What It Does

• 默认从产品方向开始压测，题型参考字节、阿里、腾讯、网易、百度等国内大厂产品面试风格。
• 可以切到 AI 方向，追问 Agent、上下文、记忆、评测、信任、成本和落地边界。
• 可以切到认知方向，追问判断、证据、行动、反馈、自我叙事和压力稳定性。
• 用户说“结束”时，停止对掏，输出评分诊断、战斗复盘和行动处方。

Install

Clone this repository, then copy the skill folder into your Codex skills directory.

PowerShell:

Copy-Item -Recurse -Force .\skill\rival-ai-pressure-test "$env:USERPROFILE\.codex\skills\"

macOS/Linux:

mkdir -p ~/.codex/skills
cp -R ./skill/rival-ai-pressure-test ~/.codex/skills/

Then start a new Codex session and invoke:

Use $rival-ai-pressure-test，开始。

or:

死对头 AI，开始压我。

For a complete guide, see docs/usage.md.

Interaction Style

The skill should not sound like a content creator writing sharp slogans. It should sound like an interviewer asking a concrete case question.

Good opening:

如果你负责抖音关注页，发现用户停留时长下降但推荐页时长上涨，你会先判断这是产品问题还是用户迁移？你会看哪一个指标来做第一步判断？

Bad opening:

别热身。最危险的不是你做不出来，而是你把自己的兴奋误判成用户需求。

The skill keeps structure internally, but the visible reply should be natural speech. It should not print labels such as 被挑战点、反方判断、压力问题 unless the user explicitly asks for a structured recap.

Commands

• 开始: start with one concrete product interview question.
• 产品: pressure-test product direction.
• AI: pressure-test AI direction.
• 认知: pressure-test cognitive judgment.
• 随机: choose the highest-leverage pressure topic.
• 继续上一题: continue the last unanswered pressure point.
• 身份复位: restore the persona and output style.
• 状态总结: output a training state block for cross-session continuation.
• 结束: stop debating and output an evaluation report.

Project Structure

rival-ai-pressure-test/
├─ skill/
│  └─ rival-ai-pressure-test/
│     ├─ SKILL.md
│     ├─ agents/openai.yaml
│     └─ references/
│        ├─ persona-card.md
│        ├─ consistency-protocol.md
│        ├─ topic-bank.md
│        └─ rubric.md
├─ examples/
│  ├─ start-product.md
│  ├─ continue-session.md
│  └─ ending-report.md
├─ docs/
│  └─ usage.md
├─ CONTRIBUTING.md
└─ LICENSE

Topic Bank

The topic bank contains 400 questions:

• Product direction: 200 questions.
• AI direction: 100 questions.
• Cognitive direction: 100 questions.

The product section is split evenly across domestic big-tech styles:

• ByteDance: 40 questions.
• Alibaba: 40 questions.
• Tencent: 40 questions.
• NetEase: 40 questions.
• Baidu: 40 questions.

The topic bank is ammunition, not a lottery box. The skill may use it directly or invent a sharper question from the current context.

License

MIT