PromptsRush：Claude Sonnet 5 生产级 Agent 任务分解与系统提示词模板

一句话结论

提供专为 Claude Sonnet 5 设计的 Agent 任务拆解、可复用系统提示词生成、自主任务简报以及根因排错提示词，通过结构化契约和明确验收条件最大化模型的 Agent 执行效率。

适用场景

• 适合的任务：复杂工作流拆解、自主 Agent 任务简报制定、系统提示词反向提炼、线上故障根因分析与代码自测。

• 不适合的任务：无明确边界的开放式闲聊、无需结构化输出的一句话日常问答。

• 适用的模型版本：Claude Sonnet 5（API 及各大客户端）。

• 适用的客户端、Agent 或 API：Claude Code、Cursor、Cline、Claude Web/Desktop 及自建 Agent 框架。

• 推荐的推理档位和参数：任务分解与架构分析推荐 high；常规代码修复与测试生成推荐 medium。

可直接使用的内容

1. 复杂工作流拆解与 Agent 化提示词（Workflow Decomposer）

Here is a workflow I do manually: [DESCRIBE THE STEPS, TOOLS, FREQUENCY, TIME COST].
Decompose it: which steps are deterministic (automate with code/rules), which need judgment but follow patterns (delegate to an AI agent with a prompt), and which genuinely need me. For the agent-able steps, write the reusable prompt. For the code-able steps, name the simplest tool. Estimate hours saved per month and the maintenance cost I am signing up for.

2. 可复用系统提示词生成器（Reusable System Prompt Generator）

I do this task repeatedly: [DESCRIBE TASK + PASTE 1-2 EXAMPLES OF GOOD OUTPUT].
Write a reusable system prompt that would make an AI produce output like my examples every time. Include: role, the quality bar defined by observable properties (not adjectives), format contract, edge-case handling rules, and the banned failure modes you can infer from what my examples deliberately avoid. Keep it under 250 words so it stays maintainable.

3. 自主 Agent 任务简报提示词（Agent Autonomous Task Brief）

Write a task brief for an AI agent that will work autonomously on: [THE TASK].
Include: the goal with a measurable done-condition, the inputs and where they live, hard constraints (what it must never do), decision rules for the 3 most likely ambiguities it will hit, when to stop and ask versus proceed, and the exact format of the final report. Assume the agent is capable but literal — every unstated assumption becomes a wrong guess.

4. 故障根因排查提示词（Bug Root-Cause Hunt）

You are a senior [LANGUAGE/STACK] engineer debugging a production issue.
Symptom: [WHAT HAPPENS — error message, wrong output, timing]
Expected: [WHAT SHOULD HAPPEN]
Code: [PASTE THE RELEVANT FUNCTION(S) ONLY]
Reason about the failure path before proposing anything. List the 3 most likely root causes ranked by probability, the one-line check that would confirm each, and the fix for the top candidate. Do not suggest rewriting the whole module.

测试/工作流步骤

1. 确定任务契约：明确任务目标、确定输入输出格式及禁绝的失败模式。

2. 复制提示词模板：替换模板中的 [BRACKETS] 占位符为具体业务上下文。

3. 首轮执行：将提示词输入 Claude Sonnet 5，要求先输出执行计划与推理路径。

4. 验证与迭代：根据 Sonnet 5 给出的结构化检查清单验证输出，必要时触发后续子步骤。

原始证据与数据

• 来源指出：在 Sonnet 5 上直接套用泛化提示词容易产生冗余思考或过于发散的回答，而结构化模板（明确角色、上下文约束、可观测质量门禁、输出契约）能让 Sonnet 5 保持高生产级首轮采纳率。

• 强调 Sonnet 5 拥有更强的指令遵循与自动纠错能力，但在面对模糊指令时倾向于过度生成辅助代码，因此必须在提示词中显式加入约束（如“Do not suggest rewriting the whole module”）。

适用边界

• 提示词中的约束规则必须具体明确；若使用主观形容词（如“写得更好”），Sonnet 5 可能会花费额外 token 循环自检却无法提升核心指标。

• 在超长上下文场景下，需将核心约束置于提示词开头或结尾，避免中间信息衰减。

来源摘录或观察（仅做合规短引）

• 来源指出：“Claude Sonnet 5 is the model we route 80% of our work through... near-frontier output at a speed and cost that turns 'should I ask the AI?' into a non-question.”