codex-prompting
解决 Codex 特有的提示陷阱，让模型按预期工作
仅适用于 Codex，提示过拟合可能导致行为偏差
为 Codex 编写或微调提示，适配其 agent 化执行细节与习性
帮你写或微调给 OpenAI Codex（GPT-5.x-codex）用的提示词，让 Codex 按预期做事。常在需要它执行代码任务、启动子代理（subagents）或恢复/分叉会话时用到。特别之处是会顾及 Codex 的 agent 化习性，能教它使用默认不认识的能力，减少提示词和模型“打架”的情况。
▸ 展开 SKILL.md 英文原文
Author or tune a prompt/instruction aimed at OpenAI Codex (the CLI/agent powered by GPT-5.x-codex), accounting for Codex's specific harness quirks. Assembles a Codex-directed prompt from per-model reference notes plus fill-in templates, and — critically — teaches Codex to use capabilities it does NOT know about by default (spawning subagents, resuming/forking other sessions). Use when the user wants to "write a prompt for codex", "codex への指示を作る/組み立てる", "codex 向けプロンプト", "codex に渡すプロンプト", tune phrasing for codex, pick a codex model + reasoning effort, or get codex to use subagents/sessions. NOT for designing a full autonomous multi-wave handoff package from repo analysis (that is codex-orchestrator-brief), NOT for Claude driving codex sessions live as orchestrator (that is codex-tdd-orchestration), and NOT for optimizing a generic non-codex prompt (that is prompt-optimizer).
帮我安装这个 skill：https://raw.githubusercontent.com/Eotel/skills/main/codex-prompting/SKILL.mdcurl -fsSL "https://raw.githubusercontent.com/Eotel/skills/main/codex-prompting/SKILL.md"# codex-prompting Codex is not a generic chat model with a code mode; it is a GPT-5.x model that has been **post-trained for a specific agentic harness** (the Codex CLI / Codex cloud). It has strong baked-in priors about how to plan, edit files, run tools, write preambles, and finish a task. A prompt that ignores those priors fights the model; a prompt that aligns with them gets a senior-engineer-grade result. This skill is the knowledge base for writing **prompts aimed at Codex** plus templates to assemble them. ## What this produces A finished Codex-directed instruction — one of: - A **task prompt** you paste into a Codex session (or `codex exec "..."`). - A **developer/system prompt**