create-agent

name	create-agent
description	Distill a user's idea, viewpoint, or method into a reusable social-agent skill package that acts as an expression agent on their behalf.
argument-hint	[seed-or-slug]
version	2.0.0
user-invocable	true
allowed-tools	Read, Write, Edit, Bash

Language policy: Detect the user's language from the first message and keep that language for the whole run. 本技能根据用户首条消息自动选择中文或英文，并全程保持同一种语言。

赛博嘴替.skills / create-agent（V2）

产品定义

• 本技能用于生成 social-agent skill package，不是角色扮演或人物复刻工具。
• 品牌入口是 赛博嘴替.skills，命令入口是 /create-agent。
• 产物是一个可复用的 expression agent：代表用户解释其想法、观点或方法。
• 输入模式固定为 interview-first, attachments-assisted：
  • 访谈是主输入。
  • 附件用于证据、例子、语气校准，不是启动前置条件。
• 主场景是 解释想法：讲清楚、换受众表达、纠正误解、在对方没理解时继续解释。

触发条件（中文）

当用户说以下任一内容时启动：

• /create-agent
• 帮我做一个表达代理
• 我想蒸馏一个想法代理
• 帮我做一个能替我解释想法的 skill

当用户在已创建代理上要求增量更新时：

• 我有新材料 / 追加
• 这不对 / 应该改成
• /update-agent {slug}

当用户说 /list-agents 时，列出所有代理。 当用户说 /delete-agent {slug} 时，执行删除流程（需确认）。

操作契约（中文）

1) 创建流程（必须遵守）

1. 将 prompts/intake.md 作为唯一权威 intake 脚本执行（按脚本中的三轮 batched 提问与处理规则）。
2. 每轮都执行“整轮提问 -> 用户长回答 -> 本轮摘要 -> 轻确认后再进入下一轮”。
3. 三轮结束后，使用 prompts/preview.md 作为唯一权威 preview 脚本输出 4 类预览并请求修正。
4. 附件可在 Round 1/2 提前接收；但默认在三轮完成前不做深度消费，仅在“需要立即澄清 seed”时提前使用。
5. 预览确认后，继续补充使用附件中的例子、反例、术语与语气证据；若无附件仍可继续生成。
6. 生成代理包并写入目录。
7. 返回调用方式与可管理命令。

2) 三轮访谈摘要（仅作参考，操作以 prompts/intake.md 为准）

Round 1: 表达目标

• 你希望代理替你表达什么 seed？
• 这件事最希望对方带走的 1-3 个结论是什么？
• 常见使用场景是什么？

Round 2: 误解与纠偏

• 对方通常在哪些点误解你？
• 你通常如何重讲？
• 哪个例子/反例最有效？

Round 3: 互动风格

• 代理默认应当多直接、多耐心、多结构化？
• 面对不同理解水平如何切换表达方式？
• 哪些表达方式必须避免？

3) 目标产物形态

生成完成后，默认写入 ./agents/{slug}/，产物包含：

• SKILL.md：可直接调用的代理入口
• intent.md：表达目标、边界、非目标
• core_view.md：核心观点与推理主线
• explanation_routes.md：多条稳定解释路径
• audience_translation.md：面向不同受众的改写策略
• interaction_contract.md：互动行为与提问策略
• style_overrides.md：语气/深度/结构等可调参数
• misunderstandings.md：高频误解与纠偏动作
• expression_assets.md：可复用比喻、例子、反例、短句模板
• meta.json：版本、时间戳、来源摘要

3.1) V2 Builder 构建顺序（操作性约束）

在 intake 三轮完成、preview 修正完成（confirmed preview）、并完成附件补充使用后，必须按以下顺序构建 V2 包：

1. intent_builder
2. core_view_builder
3. explanation_routes_builder
4. audience_translation_builder
5. interaction_contract_builder
6. style_overrides_builder
7. misunderstandings_builder
8. expression_assets_builder

一致性锚点（强约束）：

• 所有 builders 必须对齐已确认的 intake summaries + confirmed preview。
• Builders do not invent new conclusions beyond confirmed material.
• 若证据不足，使用各 builder 的 empty-state fallback，而不是补造结论。

4) 进化流程

对 /update-agent {slug}：

1. 读取现有代理包。
2. 对“新增材料/补充信息”，将 prompts/merger.md 作为权威 merge 路由脚本执行。
3. 对“这不对/应该改成/这个说法不像我”类显式纠正，将 prompts/correction_handler.md 作为权威 correction 路由脚本执行。
4. 只做增量更新，不丢失已确认结论。
5. 更新 meta.json 的 version 与 updated_at。
6. 返回变更摘要（新增了什么、替换了什么、为什么）。

5) 管理命令

默认代理目录：

• ./agents/{slug}/
• list / update 在默认 ./agents/ 视图下会自动识别同级 ./colleagues/ 旧产物，并按兼容对象处理。

/list-agents：

python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py --action list --base-dir ./agents

/update-agent {slug}：

• 走“进化流程”，对指定代理做增量更新。

/delete-agent {slug}：

rm -rf agents/{slug}

执行删除前必须口头确认。

Tool Rules（共享）

• 文本/Markdown/JSON/图片/PDF 读取：Read
• 附件解析或采集脚本：Bash 调用 tools/ 下脚本
• 文档写入：Write / Edit
• 版本变更：更新 meta.json

若采集脚本失败：

• 先报告错误原因和可行替代方案
• 可退化为“仅访谈 + 手工附件读取”模式
• 不得因为缺少附件而拒绝创建

---

English Version

Product Contract

• This skill creates a social-agent skill package, not a roleplay clone.
• Front-door brand: 赛博嘴替.skills; command surface: /create-agent.
• The generated artifact is an expression agent.
• Input model is interview-first, attachments-assisted.
• Primary scenario is explain ideas with audience adaptation and misunderstanding correction.

Trigger Conditions

Start when the user says any of:

• /create-agent
• Help me create an expression agent
• I want to distill an idea into an agent
• Build a skill that explains my idea for me
• Create an agent for this concept

Evolution mode:

• I have new material / append
• This is wrong / change this
• /update-agent {slug}

List mode:

• /list-agents

Delete mode:

• /delete-agent {slug} (confirmation required)

Operational Flow

1. Treat prompts/intake.md as the authoritative intake script and execute it round-by-round.
2. For each round, follow: batched prompts -> long-form user answer -> round summary -> light confirmation before proceeding.
3. After Round 3, treat prompts/preview.md as the authoritative preview script and request corrections across 4 categories.
4. Attachments may be offered early (Round 1/2), but defer deep attachment use until after the three rounds unless needed to clarify the seed immediately.
5. After preview confirmation, use attachments as supporting evidence/examples (or continue without attachments).
6. Generate the reusable agent package.
7. Return invocation and management commands.

Interview Rounds (Reference Only; prompts/intake.md is operative)

Round 1: What should this agent express?

• What seed should be expressed?
• What are the top 1-3 takeaways?
• In which scenarios is this explanation needed?

Round 2: Why do people misunderstand it?

• Where do misunderstandings happen?
• How do you usually re-explain?
• Which examples/counterexamples work best?

Round 3: How should it interact?

• Preferred directness, patience, and structure?
• How to adapt for weaker/stronger listeners?
• What expression styles should be avoided?

Target Package Shape

Generated agents are written to ./agents/{slug}/ by default.

• SKILL.md
• intent.md
• core_view.md
• explanation_routes.md
• audience_translation.md
• interaction_contract.md
• style_overrides.md
• misunderstandings.md
• expression_assets.md
• meta.json

Operative V2 Builder Sequence

After intake rounds are complete, preview corrections are confirmed (confirmed preview), and attachment evidence is incorporated, build the V2 package in this exact order:

1. intent_builder
2. core_view_builder
3. explanation_routes_builder
4. audience_translation_builder
5. interaction_contract_builder
6. style_overrides_builder
7. misunderstandings_builder
8. expression_assets_builder

Consistency anchor (hard rule):

• Keep every builder aligned to confirmed intake summaries and confirmed preview.
• Builders do not invent conclusions that are not supported by confirmed material.
• Use empty-state fallback text when evidence is missing.

Management Commands

/list-agents:

python3 ${CLAUDE_SKILL_DIR}/tools/skill_writer.py --action list --base-dir ./agents

/update-agent {slug}:

• Run incremental evolution update for the target agent.

/delete-agent {slug}:

rm -rf agents/{slug}

Require explicit confirmation before delete.

Compatibility Notes

• Compatibility posture for V1: soft migration.
• Primary package path and command surface use ./agents/{slug}/ and /create-agent.
• Legacy ./colleagues/{slug}/ packages are compatibility inputs only, not the active output target.
• Expression-agent build orchestration now uses:
  • prompts/intent_builder.md
  • prompts/core_view_builder.md
  • prompts/explanation_routes_builder.md
  • prompts/audience_translation_builder.md
  • prompts/interaction_contract_builder.md
  • prompts/style_overrides_builder.md
  • prompts/misunderstandings_builder.md
  • prompts/expression_assets_builder.md
• Legacy persona/work prompt files are kept in place only as deprecated migration stubs.
• User-facing invocation target is /{slug}.
• Default package path is ./agents/{slug}/.
• Default list and update flows also auto-detect same-level ./colleagues/{slug}/ packages and mark them as compatibility targets.