常设指令授予你的智能体对已定义程序的永久操作权限。你无需为每个任务提示智能体,而是定义具有明确范围、触发器和升级规则的程序,智能体会在这些边界内自主执行:“你负责每周报告。每周五编制并发送,只有在看起来有问题时才升级。”
为什么需要常设指令
没有常设指令: 你需要为每个任务提示智能体,例行工作会被遗忘或延迟,而你会成为瓶颈。
有了常设指令: 智能体会在定义好的边界内自主执行,例行工作会按计划发生,而你只需参与异常和审批。
它们如何工作
常设指令定义在你的 Agent 工作区文件中。推荐做法是将它们直接写入 AGENTS.md(每个会话都会自动注入),这样智能体始终能在上下文中获得它们。对于更大的配置,你也可以将它们放在专用文件中,例如 standing-orders.md,并从 AGENTS.md 引用它。
每个程序指定:
- 范围 - 智能体被授权执行的内容
- 触发器 - 何时执行(计划、事件或条件)
- 审批关卡 - 行动前需要人工签署确认的内容
- 升级规则 - 何时停止并请求帮助
智能体会通过工作区引导文件在每个会话中加载这些指令(请参阅 Agent 工作区了解自动注入文件的完整列表),并结合用于基于时间强制执行的 cron 作业来执行它们。
将常设指令放入 AGENTS.md,以保证每个会话都会加载它们。工作区引导会自动注入 AGENTS.md、SOUL.md、TOOLS.md、IDENTITY.md、USER.md、HEARTBEAT.md、BOOTSTRAP.md 和 MEMORY.md,但不会注入子目录中的任意文件。
常设指令的结构
## Program: Weekly Status Report
**Authority:** Compile data, generate report, deliver to stakeholders
**Trigger:** Every Friday at 4 PM (enforced via cron job)
**Approval gate:** None for standard reports. Flag anomalies for human review.
**Escalation:** If data source is unavailable or metrics look unusual (>2σ from norm)
### Execution steps
1. Pull metrics from configured sources
2. Compare to prior week and targets
3. Generate report in Reports/weekly/YYYY-MM-DD.md
4. Deliver summary via configured channel
5. Log completion to Agent/Logs/
### What NOT to do
- Do not send reports to external parties
- Do not modify source data
- Do not skip delivery if metrics look bad - report accurately
常设指令加 cron 作业
常设指令定义智能体被授权执行的内容。Cron 作业定义它发生的时间。它们协同工作:
Standing Order: "You own the daily inbox triage"
↓
Cron Job (8 AM daily): "Execute inbox triage per standing orders"
↓
Agent: Reads standing orders → executes steps → reports results
cron 作业提示应引用常设指令,而不是重复它:
openclaw cron add \
--name daily-inbox-triage \
--cron "0 8 * * 1-5" \
--tz America/New_York \
--timeout-seconds 300 \
--announce \
--channel imessage \
--to "+1XXXXXXXXXX" \
--message "Execute daily inbox triage per standing orders. Check mail for new alerts. Parse, categorize, and persist each item. Report summary to owner. Escalate unknowns."
示例 1:内容和社交媒体(每周周期)
## Program: Content & Social Media
**Authority:** Draft content, schedule posts, compile engagement reports
**Approval gate:** All posts require owner review for first 30 days, then standing approval
**Trigger:** Weekly cycle (Monday review → mid-week drafts → Friday brief)
### Weekly cycle
- **Monday:** Review platform metrics and audience engagement
- **Tuesday-Thursday:** Draft social posts, create blog content
- **Friday:** Compile weekly marketing brief → deliver to owner
### Content rules
- Voice must match the brand (see SOUL.md or brand voice guide)
- Never identify as AI in public-facing content
- Include metrics when available
- Focus on value to audience, not self-promotion
示例 2:财务运营(事件触发)
## Program: Financial Processing
**Authority:** Process transaction data, generate reports, send summaries
**Approval gate:** None for analysis. Recommendations require owner approval.
**Trigger:** New data file detected OR scheduled monthly cycle
### When new data arrives
1. Detect new file in designated input directory
2. Parse and categorize all transactions
3. Compare against budget targets
4. Flag: unusual items, threshold breaches, new recurring charges
5. Generate report in designated output directory
6. Deliver summary to owner via configured channel
### Escalation rules
- Single item > $500: immediate alert
- Category > budget by 20%: flag in report
- Unrecognizable transaction: ask owner for categorization
- Failed processing after 2 retries: report failure, do not guess
示例 3:监控和告警(持续)
## Program: System Monitoring
**Authority:** Check system health, restart services, send alerts
**Approval gate:** Restart services automatically. Escalate if restart fails twice.
**Trigger:** Every heartbeat cycle
### Checks
- Service health endpoints responding
- Disk space above threshold
- Pending tasks not stale (>24 hours)
- Delivery channels operational
### Response matrix
| Condition | Action | Escalate? |
| ---------------- | ------------------------ | ------------------------ |
| Service down | Restart automatically | Only if restart fails 2x |
| Disk space < 10% | Alert owner | Yes |
| Stale task > 24h | Remind owner | No |
| Channel offline | Log and retry next cycle | If offline > 2 hours |
执行-验证-报告模式
常设指令与严格的执行纪律结合时效果最好。常设指令中的每个任务都应遵循此循环:
- 执行 - 完成实际工作(不要只是确认收到指令)
- 验证 - 确认结果正确(文件存在、消息已送达、数据已解析)
- 报告 - 告诉所有者完成了什么以及验证了什么
### Execution rules
- Every task follows Execute-Verify-Report. No exceptions.
- "I'll do that" is not execution. Do it, then report.
- "Done" without verification is not acceptable. Prove it.
- If execution fails: retry once with adjusted approach.
- If still fails: report failure with diagnosis. Never silently fail.
- Never retry indefinitely - 3 attempts max, then escalate.
此模式可以防止最常见的智能体失败模式:确认任务但没有完成。
多程序架构
对于管理多个关注点的智能体,请将常设指令组织为边界清晰的独立程序:
## Program 1: [Domain A] (Weekly)
...
## Program 2: [Domain B] (Monthly + On-Demand)
...
## Program 3: [Domain C] (As-Needed)
...
## Escalation Rules (All Programs)
- [Common escalation criteria]
- [Approval gates that apply across programs]
每个程序都应有:
- 自己的触发节奏(每周、每月、事件驱动、持续)
- 自己的审批关卡(某些程序比其他程序需要更多监督)
- 清晰的边界(智能体应知道一个程序在哪里结束,另一个程序从哪里开始)
最佳实践
应该做
- 从较窄权限开始,随着信任建立再扩展
- 为高风险操作定义明确的审批关卡
- 包含“不要做什么”部分 - 边界和权限同样重要
- 与 cron 作业结合,实现可靠的基于时间执行
- 每周查看智能体日志,验证常设指令正在被遵循
- 随着需求演变更新常设指令 - 它们是动态文档
- 第一天就授予宽泛权限(“做你认为最好的事”)
- 跳过升级规则 - 每个程序都需要一个“何时停止并询问”的条款
- 假设智能体会记住口头指令 - 将所有内容写入文件
- 在单个程序中混合关注点 - 不同领域使用不同程序
- 忘记用 cron 作业强制执行 - 没有触发器的常设指令会变成建议