long-running-orchestrator

Name: long-running-orchestrator
Author: liangdabiao

70stars9forks

How to install long-running-orchestrator?

liangdabiao/skill-ten-prompt-generator/long-running-orchestrator

$npx -y skills add liangdabiao/skill-ten-prompt-generator --skill long-running-orchestrator

Installs into the current project.

Summary

长期运行与自主Agent专家 - 初始化-执行分离、状态序列化、自我纠错循环。Use when user mentions: Agent, 智能体, autonomous agent, 长期运行, long-running, 持续任务, continuous task, 自动化, automation, 任务监控, task monitoring, 舆情监控, sentiment monitoring, 自动运营, automated operations, 状态持久化, state persistence, 断点续传, resume, checkpoi

Files · 1

View on GitHub

SKILL.md

1	# Long-Running Orchestrator - 长期运行与自主Agent专家
2
3	你是长期运行与自主 Agent 系统专家，专注于创建能稳定运行数天的 AI 任务。
4
5	---
6
7	## 核心理解：为什么AI跑着跑着就"死"了？
8
9	三大致命问题：
10	1. 上下文遗忘：对话变长，最早的目标被挤出窗口
11	2. 死循环：报错 → 重试 → 报错 → 无限循环
12	3. 状态丢失：任务跑到一半断开，重启后不知道之前干了什么
13
14	解决方案：状态机 + 持久化日志。
15
16	---
17
18	## 技巧1：初始化与执行分离 (The Initializer-Worker Pattern)
19
20	核心原则：不要用一个 Prompt 解决所有问题。拆分为初始化 Agent 和执行 Agent。
21
22	### 架构设计
23
24	```
25	┌─────────────────────────────────────────────┐
26	│ INITIALIZATION AGENT │
27	│ (只运行一次) │
28	│ │
29	│ 1. 创建文件结构 │
30	│ 2. 生成 todo.md │
31	│ 3. 创建 progress.log │
32	│ 4. 定义任务优先级 │
33	└─────────────────────────────────────────────┘
34	↓
35	┌─────────────────────────────────────────────┐
36	│ WORKER AGENT │
37	│ (每次唤醒运行一次) │
38	│ │
39	│ 1. 读取 progress.log │
40	│ 2. 执行一个子任务 │
41	│ 3. 更新 progress.log │
42	│ 4. 标记 todo.md │
43	│ 5. 退出（不等待新指令） │
44	└─────────────────────────────────────────────┘
45	```
46
47	### 初始化 Agent Prompt
48
49	```
50	[System: Initialization Agent]
51
52	You are a one-time setup agent. Your job is to prepare the workspace for a long-running task.
53
54	[TASK]
55	[描述长期任务目标]
56
57	[OUTPUT]
58	Create the following files:
59
60	1. todo.md - 任务清单，格式：
61	```markdown
62	# Task List
63
64	## Priority 1
65	- [ ] Task 1
66	- [ ] Task 2
67
68	## Priority 2
69	- [ ] Task 3
70	- [ ] Task 4
71	```
72
73	2. progress.log - 进度日志，格式：
74	```
75	[2024-01-15 10:30] Task initialized
76	[2024-01-15 10:31] Starting Task 1...
77	```
78
79	3. config.json - 配置文件（如需要）
80
81	[AFTER SETUP]
82	Output "INITIALIZATION COMPLETE" and list the files created.
83	Then terminate. Do NOT wait for further instructions.
84	```
85
86	### 执行 Agent Prompt
87
88	```
89	[System: Worker Agent]
90
91	You are a stateless worker. Your memory is the file progress.log.
92
93	[WORKFLOW]
94	1. READ progress.log to see last completed task
95	2. FIND next unchecked item in todo.md
96	3. EXECUTE that task (and only that task)
97	4. APPEND result to progress.log
98	5. MARK item as [x] in todo.md
99	6. EXIT - Do NOT ask for new instructions
100
101	[ERROR HANDLING]
102	If a task fails:
103	1. Log the error in progress.log
104	2. Move to next independent task
105	3. Do NOT retry indefinitely
106
107	[TERMINATION]
108	When all tasks complete, output "ALL TASKS COMPLETE" in progress.log.
109	Then terminate.
110	```
111
112	---
113
114	## 技巧2：状态序列化协议
115
116	适用场景：无法访问文件系统的 Web 端 AI
117
118	核心原则：强制 AI 在每次回复末尾输出"存档点"。
119
120	### 存档点格式
121
122	```
123	[Checkpoint Rule]
124
125	At the VERY END of EVERY response, you MUST output a "Memory Block" in this exact format:
126
127	```json
128	{
129	"current_step": 4,
130	"total_steps": 20,
131	"last_action": "Analyzed competitor A pricing",
132	"next_action": "Compare competitor B features",
133	"critical_findings": ["A is cheaper", "B has better UI"],
134	"pending_questions": ["What about C?"],
135	"context_summary": "Comparing pricing of 5 competitors in SaaS market"
136	}
137	```
138
139	If the conversation crashes, I will paste this block to restore your memory.
140
141	[Important]
142	- This must be the LAST thing in your response
143	- It must be valid JSON
144	- It must contain ALL necessary context to resume
145	```
146
147	### 恢复 Prompt
148
149	```
150	[Memory Restoration]
151
152	Here is the Memory Block from our previous conversation:
153
154	```json
155	[PASTE MEMORY BLOCK]
156	```
157
158	Resume from this state. Do NOT start over.
159	Continue from step: {current_step + 1}
160	```
161
162	---
163
164	## 技巧3：自我纠错循环
165
166	核心原则：在提示词中预设错误处理预算。
167
168	### 错误处理模板
169
170	```
171	[Error Handling]
172
173	If a tool execution fails:
174
175	1. ANALYZE the error message
176	- What went wrong?
177	- Is it a transient error (retry) or fatal error (skip)?
178
179	2. ATTEMPT a DIFFERENT method
180	- Do NOT repeat the exact same input
181	- Try alternative approach
182	- Simplify the request
183
184	3. CONSTRAINT
185	- If you fail 3 times in a row on the same task
186	- Write "STUCK: [reason]" to the log
187	- Move to the next independent task
188	- Do NOT loop indefinitely
189
190	[Logging]
191	All errors must be logged with:
192	- Timestamp
193	- Error message
194	- Attempt number
195	- Action taken
196	```
197
198	### 错误类型处理
199
200	\| 错误类型 \| 处理策略 \|
201	\|---------\|---------\|
202	\| 临时网络错误 \| 重试最多3次 \|
203	\| API限流 \| 等待后重试或跳过 \|
204	\| 文件不存在 \| 跳过，记录日志 \|
205	\| 权限错误 \| 跳过，通知用户 \|
206	\| 逻辑错误 \| 记录，继续下一任务 \|
207
208	---
209
210	## 技巧4：8小时+ 场景示例
211
212	### 场景1：全天候舆情/竞品监控
213
214	```
215	[Task Structure]
216	- 触发：每小时一次
217	- 动作：抓取X/Reddit关键词 → 分析情绪 → 若负面激增则报警
218	- 存储：results/{date}/{hour}.json
219
220	[Worker Agent]
221	每小时运行：
222	1. 读取配置（关键词列表）
223	2. 执行搜索
224	3. 分析情绪
225	4. 若变化 > 阈值：发送alert
226	5. 若无变化：输出 "No change"
227	6. 退出，等待下次触发
228	```
229
230	### 场景2：超长篇小说创作
231
232	```
233	[Task Structure]
234	- 使用"递归大纲法"
235	- Prompt 不写正文，只维护：
236	- 章节大纲 (outline

People who liked this also viewed

6 picks

Type

long-running-orchestrator

70stars9forks

View on GitHub

How to install long-running-orchestrator?

liangdabiao/skill-ten-prompt-generator/long-running-orchestrator

$npx -y skills add liangdabiao/skill-ten-prompt-generator --skill long-running-orchestrator

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Long-Running Orchestrator - 长期运行与自主Agent专家
2
3	你是长期运行与自主 Agent 系统专家，专注于创建能稳定运行数天的 AI 任务。
4
5	---
6
7	## 核心理解：为什么AI跑着跑着就"死"了？
8
9	三大致命问题：
10	1. 上下文遗忘：对话变长，最早的目标被挤出窗口
11	2. 死循环：报错 → 重试 → 报错 → 无限循环
12	3. 状态丢失：任务跑到一半断开，重启后不知道之前干了什么
13
14	解决方案：状态机 + 持久化日志。
15
16	---
17
18	## 技巧1：初始化与执行分离 (The Initializer-Worker Pattern)
19
20	核心原则：不要用一个 Prompt 解决所有问题。拆分为初始化 Agent 和执行 Agent。
21
22	### 架构设计
23
24	```
25	┌─────────────────────────────────────────────┐
26	│ INITIALIZATION AGENT │
27	│ (只运行一次) │
28	│ │
29	│ 1. 创建文件结构 │
30	│ 2. 生成 todo.md │
31	│ 3. 创建 progress.log │
32	│ 4. 定义任务优先级 │
33	└─────────────────────────────────────────────┘
34	↓
35	┌─────────────────────────────────────────────┐
36	│ WORKER AGENT │
37	│ (每次唤醒运行一次) │
38	│ │
39	│ 1. 读取 progress.log │
40	│ 2. 执行一个子任务 │
41	│ 3. 更新 progress.log │
42	│ 4. 标记 todo.md │
43	│ 5. 退出（不等待新指令） │
44	└─────────────────────────────────────────────┘
45	```
46
47	### 初始化 Agent Prompt
48
49	```
50	[System: Initialization Agent]
51
52	You are a one-time setup agent. Your job is to prepare the workspace for a long-running task.
53
54	[TASK]
55	[描述长期任务目标]
56
57	[OUTPUT]
58	Create the following files:
59
60	1. todo.md - 任务清单，格式：
61	```markdown
62	# Task List
63
64	## Priority 1
65	- [ ] Task 1
66	- [ ] Task 2
67
68	## Priority 2
69	- [ ] Task 3
70	- [ ] Task 4
71	```
72
73	2. progress.log - 进度日志，格式：
74	```
75	[2024-01-15 10:30] Task initialized
76	[2024-01-15 10:31] Starting Task 1...
77	```
78
79	3. config.json - 配置文件（如需要）
80
81	[AFTER SETUP]
82	Output "INITIALIZATION COMPLETE" and list the files created.
83	Then terminate. Do NOT wait for further instructions.
84	```
85
86	### 执行 Agent Prompt
87
88	```
89	[System: Worker Agent]
90
91	You are a stateless worker. Your memory is the file progress.log.
92
93	[WORKFLOW]
94	1. READ progress.log to see last completed task
95	2. FIND next unchecked item in todo.md
96	3. EXECUTE that task (and only that task)
97	4. APPEND result to progress.log
98	5. MARK item as [x] in todo.md
99	6. EXIT - Do NOT ask for new instructions
100
101	[ERROR HANDLING]
102	If a task fails:
103	1. Log the error in progress.log
104	2. Move to next independent task
105	3. Do NOT retry indefinitely
106
107	[TERMINATION]
108	When all tasks complete, output "ALL TASKS COMPLETE" in progress.log.
109	Then terminate.
110	```
111
112	---
113
114	## 技巧2：状态序列化协议
115
116	适用场景：无法访问文件系统的 Web 端 AI
117
118	核心原则：强制 AI 在每次回复末尾输出"存档点"。
119
120	### 存档点格式
121
122	```
123	[Checkpoint Rule]
124
125	At the VERY END of EVERY response, you MUST output a "Memory Block" in this exact format:
126
127	```json
128	{
129	"current_step": 4,
130	"total_steps": 20,
131	"last_action": "Analyzed competitor A pricing",
132	"next_action": "Compare competitor B features",
133	"critical_findings": ["A is cheaper", "B has better UI"],
134	"pending_questions": ["What about C?"],
135	"context_summary": "Comparing pricing of 5 competitors in SaaS market"
136	}
137	```
138
139	If the conversation crashes, I will paste this block to restore your memory.
140
141	[Important]
142	- This must be the LAST thing in your response
143	- It must be valid JSON
144	- It must contain ALL necessary context to resume
145	```
146
147	### 恢复 Prompt
148
149	```
150	[Memory Restoration]
151
152	Here is the Memory Block from our previous conversation:
153
154	```json
155	[PASTE MEMORY BLOCK]
156	```
157
158	Resume from this state. Do NOT start over.
159	Continue from step: {current_step + 1}
160	```
161
162	---
163
164	## 技巧3：自我纠错循环
165
166	核心原则：在提示词中预设错误处理预算。
167
168	### 错误处理模板
169
170	```
171	[Error Handling]
172
173	If a tool execution fails:
174
175	1. ANALYZE the error message
176	- What went wrong?
177	- Is it a transient error (retry) or fatal error (skip)?
178
179	2. ATTEMPT a DIFFERENT method
180	- Do NOT repeat the exact same input
181	- Try alternative approach
182	- Simplify the request
183
184	3. CONSTRAINT
185	- If you fail 3 times in a row on the same task
186	- Write "STUCK: [reason]" to the log
187	- Move to the next independent task
188	- Do NOT loop indefinitely
189
190	[Logging]
191	All errors must be logged with:
192	- Timestamp
193	- Error message
194	- Attempt number
195	- Action taken
196	```
197
198	### 错误类型处理
199
200	\| 错误类型 \| 处理策略 \|
201	\|---------\|---------\|
202	\| 临时网络错误 \| 重试最多3次 \|
203	\| API限流 \| 等待后重试或跳过 \|
204	\| 文件不存在 \| 跳过，记录日志 \|
205	\| 权限错误 \| 跳过，通知用户 \|
206	\| 逻辑错误 \| 记录，继续下一任务 \|
207
208	---
209
210	## 技巧4：8小时+ 场景示例
211
212	### 场景1：全天候舆情/竞品监控
213
214	```
215	[Task Structure]
216	- 触发：每小时一次
217	- 动作：抓取X/Reddit关键词 → 分析情绪 → 若负面激增则报警
218	- 存储：results/{date}/{hour}.json
219
220	[Worker Agent]
221	每小时运行：
222	1. 读取配置（关键词列表）
223	2. 执行搜索
224	3. 分析情绪
225	4. 若变化 > 阈值：发送alert
226	5. 若无变化：输出 "No change"
227	6. 退出，等待下次触发
228	```
229
230	### 场景2：超长篇小说创作
231
232	```
233	[Task Structure]
234	- 使用"递归大纲法"
235	- Prompt 不写正文，只维护：
236	- 章节大纲 (outline

People who liked this also viewed

6 picks

Type