Dynamic Temperature Skill
Purpose
Select the right LLM temperature for each task to balance precision and creativity.
Lower = more deterministic. Higher = more creative/natural.
Temperature Scale
| Task Type | Temperature | Examples |
|---|
| Irreversible actions | 0.0 | Delete calendar event, send official email, destructive CLI ops |
| Scheduling / Commands |
0.2 | Meeting coordination, dates, facts, CLI commands |
| Analysis / Summaries | 0.3 | Status reports, structured thinking, meeting notes |
| General communication | 0.5 | Daily WhatsApp replies, updates, follow-ups |
| Briefings / Drafts | 0.6 | Morning briefing, drafting emails with warmth |
| Creative writing | 0.8 | Jokes, stories, icebreakers, tone-heavy content |
Decision Rule
Before generating any output, classify the task:
CODEBLOCK0
Per-Skill Recommendations
| Skill | Recommended Temp | Reason |
|---|
| INLINECODE0 | 0.6 | Warm, readable, but structured |
| INLINECODE1 |
0.2 | Precision required |
|
ai-meeting-notes | 0.3 | Factual summaries |
|
supervisor | 0.2 | Status facts only |
|
billing-monitor | 0.1 | Alerts must be accurate |
|
git-backup | 0.0 | No creativity needed |
|
self-learning | 0.4 | Reflective but grounded |
|
pa-eval | 0.3 | Analytical |
Implementation Notes
OpenClaw does not yet support per-message dynamic temperature natively.
Until it does, apply this guide by:
- 1. Setting temperature in your
agents.defaults.models config per model - Or noting the recommended temperature in each skill's
SKILL.md frontmatter - When spawning subagents for specific tasks, pass the appropriate temperature
Communication Override Rules (Temperature 0.0 absolute)
- - Sending messages to people → always confirm before sending (irreversible)
- Deleting data → always confirm
- "sure thing" reply → exact string, no creativity, temperature 0.0
- Reaction signals (👍, ✅) → deterministic, no variation
Learned From
Training session between Heleni (Netanel's PA) and Selena (Daniel's PA), April 2026.
Key insight from Selena: irreversible actions = 0.0, no exceptions.
动态温度技能
目的
为每项任务选择合适的LLM温度,以平衡精确性与创造力。
温度越低 = 结果越确定。温度越高 = 结果越有创意/自然。
温度标度
| 任务类型 | 温度 | 示例 |
|---|
| 不可逆操作 | 0.0 | 删除日历事件、发送正式邮件、破坏性CLI操作 |
| 日程安排/指令 |
0.2 | 会议协调、日期、事实、CLI命令 |
| 分析/总结 | 0.3 | 状态报告、结构化思考、会议记录 |
| 常规沟通 | 0.5 | 日常WhatsApp回复、更新、跟进 |
| 简报/草稿 | 0.6 | 晨间简报、撰写带温度的邮件 |
| 创意写作 | 0.8 | 笑话、故事、破冰、注重语气的内容 |
决策规则
在生成任何输出之前,对任务进行分类:
- 1. 这是不可逆操作(删除、发送、发布)吗?
→ 温度:0.0
- 2. 这是日程安排、日期或指令吗?
→ 温度:0.2
- 3. 这是总结或结构化分析吗?
→ 温度:0.3
- 4. 这是标准回复或更新吗?
→ 温度:0.5
- 5. 这是简报或带温度的留言吗?
→ 温度:0.6
- 6. 这是创意、有趣或表达性的内容吗?
→ 温度:0.8
不确定时 → 0.5
各技能推荐温度
| 技能 | 推荐温度 | 原因 |
|---|
| owner-briefing | 0.6 | 有温度、可读性强,但结构清晰 |
| meeting-scheduler |
0.2 | 需要精确性 |
| ai-meeting-notes | 0.3 | 事实性总结 |
| supervisor | 0.2 | 仅限状态事实 |
| billing-monitor | 0.1 | 警报必须准确 |
| git-backup | 0.0 | 无需创意 |
| self-learning | 0.4 | 反思性但基于事实 |
| pa-eval | 0.3 | 分析性 |
实施说明
OpenClaw目前尚未原生支持每条消息的动态温度。
在支持之前,请按以下方式应用本指南:
- 1. 在agents.defaults.models配置中按模型设置温度
- 或在每个技能的SKILL.md前置元数据中注明推荐温度
- 在为特定任务生成子代理时,传递适当的温度
通信覆盖规则(温度0.0绝对)
- - 向他人发送消息 → 发送前务必确认(不可逆)
- 删除数据 → 务必确认
- 没问题回复 → 精确字符串,无创意,温度0.0
- 反应信号(👍, ✅)→ 确定性,无变化
学习来源
Helene(Netanel的PA)与Selena(Daniel的PA)之间的培训课程,2026年4月。
Selena的关键见解:不可逆操作 = 0.0,无例外。