Lobster Config Guide

Provides a step-by-step guide to configure Lobster AI Agent behavior via AGENTS.md, group rules, and skill plugins for optimal interaction.

Security Scan

This is an instruction-only guide for configuring a Lobster/OpenClaw agent; it requests no credentials, installs nothing, and its instructions are consistent with the stated purpose.

Purpose & Capability

Name/description claim to provide configuration guidance; the SKILL.md contains only step-by-step recommendations for AGENTS.md, group rules, and plugin selection—no unrelated requirements (no env vars, binaries, or installs).

Instruction Scope

Instructions stay on-topic: they tell how to write AGENTS.md, when to use group rules, and how to choose plugins. They do not instruct the agent to read or exfiltrate files, access secrets, or call external endpoints. They do reference restarting OpenClaw and avoid changing the system prompt unless sure, which is an appropriate caution for a configuration guide.

Install Mechanism

No install spec and no code files — instruction-only skill. Nothing is written to disk or downloaded by the skill itself.

Credentials

The skill requests no environment variables, credentials, or config paths. The occasional mention of internal items (e.g., system prompt key) is advisory and not a request for access to secrets.

Persistence & Privilege

Defaults are used (not always:true). The skill does not request permanent presence or elevated privileges, nor does it modify other skills' configs; it's a passive guidance document.

Assessment

This skill is a text-only how-to for writing agent config files and choosing plugins; it does not install software or ask for credentials. Before applying its recommendations: (1) review any changes you make to system-level prompts (it rightly warns not to change them lightly), (2) back up existing AGENTS.md or system prompt values so you can revert, and (3) only enable or whitelist plugins you trust from known sources. If you need stronger guarantees, ask the publisher for provenance or a source URL before applying the guidance in production agents.

lobster-config

来源：爸爸（youyoude）整理自龙虾社区匿名贡献者

让 AI Agent（龙虾）变聪明的核心不是训练，而是写规矩。

---

核心理念

给龙虾配置行为规范，只需要写好三份文件，缺一不可。

---

文件一：AGENTS.md（最重要）

这是龙虾的"大脑"，决定它是谁、怎么说话、什么时候开口什么时候闭嘴。

基础结构（控制在 100 行以内）

# 身份定位
一句话说清它是谁

# 说话风格
- 口语化，像个真实的人
- 别有 AI 味（禁用"很高兴为您服务""请问有什么可以帮您"）
- 能用纯文本就别加 markdown
- 重点加粗就够了

# 行为边界
- 规定什么时候回、什么时候闭嘴
- 群聊里的发言规则

# 输出格式
纯文本优先，不需要列表/表格就不加

原则：越短越好。超过 100 行模型反而抓不住重点。

---

文件二：群规则（customPrompt）

每个群单独配，没要求就别硬写。

空着比乱写强。

---

文件三：Skill 插件

去 ClawHub 按需安装，别贪多。

装了记得开白名单权限，否则等于没装。

---

两个最容易踩的坑

坑 1：改完配置不重启

配置改完，必须重启 OpenClaw 才能生效。

坑 2：temperature 设太离谱

推荐值：0.7 左右

太高输出太随机，太低输出太死板。

---

配置顺序建议

1. 先写好 AGENTS.md — 这一份搞定效果立竿见影
2. 再配群规则 — 有需要再写，没有就空着
3. 最后装 Skill — 按需选择，不要堆砌

---

附加建议（来自实践）

• 别改系统 prompt（Compaction memoryFlush.systemPrompt）除非你很确定要改
• 每次改动后测试 — 看看回复有没有变自然
• 听龙虾的输出 — 如果有 AI 味就继续调

---

整理自龙虾社区贡献者经验，由 youyoude 发布。