OpenClaw Startup
First-run setup wizard for new OpenClaw agents. Interviews the user, generates
tailored workspace files, scaffolds memory, installs operational crons, and
recommends companion skills.
Use when
- - setting up a new OpenClaw agent from scratch
- onboarding a new user who just installed OpenClaw
- resetting an agent workspace to a clean baseline
- someone says "set up my agent" or "help me get started"
Do not use when
- - the agent is already configured and running (use
openclaw-guide instead) - the user wants to change one specific config setting
- the task is skill authoring (use
skill-builder) - the task is security hardening of an existing agent (use
agent-hardening)
Phases
Run these in order. Each phase confirms with the user before proceeding.
Phase 1 — Identity Interview
Resuming a partial setup: Before starting the interview, check if any of
SOUL.md, USER.md, or IDENTITY.md already exist. If they do, show the user
what's already configured and ask: "Want to pick up where we left off, start fresh,
or keep these and skip to memory setup?" Respect their choice.
Ask the user 8 questions to understand who they are and what the agent should do.
Use the interview guide at references/interview-guide.md.
From the answers, generate:
- -
SOUL.md — agent identity, personality, tone, boundaries - INLINECODE8 — human's key details
- INLINECODE9 — agent name, emoji, presentation
- INLINECODE10 — customized from
assets/agents-template.md with principles
tailored to the user's main job (Q4) and communication style (Q5)
Show the user the generated files and ask for approval before writing.
Phase 2 — Memory Scaffolding
Run the install script to create the memory directory structure:
CODEBLOCK0
This creates:
- -
memory/ directory with daily log structure - INLINECODE13 — durable fact index
- INLINECODE14 — current priorities tracker
- INLINECODE15 — commitments and deferred actions
- INLINECODE16 — error and learning capture (agent-memory-loop)
Phase 3 — Operational Crons
Review recommended crons with the user. See references/cron-recipes.md.
If openclaw CLI is not available: Skip cron installation. Instead, show the user
the recommended cron configs from references/cron-recipes.md and explain they can
install them later once OpenClaw is set up. Proceed to Phase 4.
If openclaw CLI is available:
Always recommend:
- - Heartbeat (configurable frequency)
- Workspace backup (daily)
Recommend if applicable:
- - Weekly memory consolidation
- Session health watchdog
- Security monitoring
Install only what the user approves. Use the OpenClaw cron tool directly.
Phase 4 — Security Baseline
Walk through the security checklist at references/security-baseline.md.
Key items:
- - Channel access controls (allowlist vs open)
- Injection defense rules in SOUL.md
- Escalation framework (green/yellow/red)
- Tool restrictions for sub-agents
Phase 5 — Companion Skills
Recommend installing these skills based on the user's needs:
Always — extract text/transcripts from URLs, videos, PDFs |
clawhub.com/skills/summarize |
|
agent-hardening | If the agent handles sensitive data or external channels |
clawhub.com/skills/agent-hardening |
|
openclaw-backup | If the user wants disaster recovery |
clawhub.com/skills/openclaw-backup |
Install via ClawHub if available:
CODEBLOCK1
Phase 6 — Verification
After setup, run a quick health check:
- - Verify all generated files exist and are non-empty
- Verify crons are registered and have valid next-run times
- Verify the agent can respond to a basic prompt
- Show the user a summary of everything that was set up
Interview principles
- - Ask one question at a time — don't dump all 8 at once
- Accept short answers — infer reasonable defaults from context
- Show what you're generating before writing it
- Never write to core files without user confirmation
- Keep the tone conversational, not bureaucratic
References
- -
references/interview-guide.md — the 8 identity questions with guidance - INLINECODE27 — recommended cron configurations
- INLINECODE28 — security checklist for new agents
- INLINECODE29 — full mapping of interview answers → template placeholders + model recommendations
Assets (templates)
- -
assets/soul-template.md — SOUL.md template with placeholders - INLINECODE31 — USER.md template
- INLINECODE32 — IDENTITY.md template
- INLINECODE33 — AGENTS.md starter
- INLINECODE34 — HEARTBEAT.md starter
Scripts
- -
scripts/install.sh — creates memory scaffolding and .learnings directory - INLINECODE36 — post-setup health check
OpenClaw 启动向导
面向新 OpenClaw 代理的首次运行设置向导。通过访谈用户,生成定制化工作区文件,搭建记忆框架,安装运维定时任务,并推荐配套技能。
使用场景
- - 从零开始设置新的 OpenClaw 代理
- 引导刚安装 OpenClaw 的新用户上手
- 将代理工作区重置为干净基线
- 当有人说帮我设置代理或帮我开始使用
禁止使用场景
- - 代理已配置并运行中(请使用 openclaw-guide)
- 用户只想修改某个具体配置项
- 任务是技能编写(请使用 skill-builder)
- 任务是对现有代理进行安全加固(请使用 agent-hardening)
执行阶段
按顺序执行以下阶段。每个阶段在继续前需获得用户确认。
阶段 1 — 身份访谈
恢复部分设置: 开始访谈前,检查 SOUL.md、USER.md 或 IDENTITY.md 是否已存在。若存在,向用户展示已配置内容并询问:您想从上次中断处继续、重新开始,还是保留这些文件并跳转到记忆设置?尊重用户的选择。
向用户提出 8 个问题,了解用户身份及代理应具备的功能。使用 references/interview-guide.md 中的访谈指南。
根据回答生成:
- - SOUL.md — 代理身份、个性、语气、边界
- USER.md — 用户关键信息
- IDENTITY.md — 代理名称、表情符号、呈现方式
- AGENTS.md — 基于 assets/agents-template.md 定制,根据用户主要工作(问题 4)和沟通风格(问题 5)调整原则
向用户展示生成的文件,获得批准后方可写入。
阶段 2 — 记忆框架搭建
运行安装脚本以创建记忆目录结构:
bash
bash {baseDir}/scripts/install.sh
该脚本将创建:
- - memory/ 目录,包含每日日志结构
- MEMORY.md — 持久化事实索引
- ACTIVE.md — 当前优先级追踪器
- memory/FUTURE_INTENTS.md — 承诺与延迟操作
- .learnings/ — 错误与学习记录(代理记忆循环)
阶段 3 — 运维定时任务
与用户一起审查推荐的定时任务。参见 references/cron-recipes.md。
如果 openclaw CLI 不可用: 跳过定时任务安装。改为向用户展示 references/cron-recipes.md 中的推荐定时任务配置,并说明待 OpenClaw 设置完成后可稍后安装。继续执行阶段 4。
如果 openclaw CLI 可用:
始终推荐:
酌情推荐:
仅安装用户批准的定时任务。直接使用 OpenClaw 定时任务工具。
阶段 4 — 安全基线
逐项检查 references/security-baseline.md 中的安全检查清单。
关键项目:
- - 频道访问控制(白名单模式 vs 开放模式)
- SOUL.md 中的注入防御规则
- 升级框架(绿色/黄色/红色)
- 子代理的工具限制
阶段 5 — 配套技能
根据用户需求推荐安装以下技能:
始终推荐 — 从 URL、视频、PDF 中提取文本/转录内容 |
clawhub.com/skills/summarize |
| agent-hardening | 如果代理处理敏感数据或外部频道 |
clawhub.com/skills/agent-hardening |
| openclaw-backup | 如果用户需要灾难恢复 |
clawhub.com/skills/openclaw-backup |
如果可用,通过 ClawHub 安装:
bash
clawhub install <技能名称>
阶段 6 — 验证
设置完成后,运行快速健康检查:
- - 验证所有生成的文件存在且非空
- 验证定时任务已注册且具有有效的下次运行时间
- 验证代理能够响应基本提示
- 向用户展示所有已设置内容的摘要
访谈原则
- - 每次只问一个问题 — 不要一次性抛出全部 8 个问题
- 接受简短回答 — 根据上下文推断合理的默认值
- 在写入前展示正在生成的内容
- 未经用户确认,绝不写入核心文件
- 保持对话语气,避免官僚风格
参考资料
- - references/interview-guide.md — 8 个身份问题及指导说明
- references/cron-recipes.md — 推荐的定时任务配置
- references/security-baseline.md — 新代理安全检查清单
- references/placeholder-map.md — 访谈回答与模板占位符及模型推荐的完整映射
资源文件(模板)
- - assets/soul-template.md — 含占位符的 SOUL.md 模板
- assets/user-template.md — USER.md 模板
- assets/identity-template.md — IDENTITY.md 模板
- assets/agents-template.md — AGENTS.md 起始模板
- assets/heartbeat-template.md — HEARTBEAT.md 起始模板
脚本
- - scripts/install.sh — 创建记忆框架和 .learnings 目录
- scripts/verify.sh — 设置后健康检查