CCAF101 · Claude Certified Architect - Foundations · MODULE 06

Task Decomposition任务分解

把大任务拆成子任务有两种活法：固定链式（步骤预先写死）和动态自适应（模型运行时决定子任务）。同一节课跑两遍，看清「按工作性质选模式」。

This demo is unofficial, original practice content written from public exam objectives — not real exam questions.

迭代 0/3·messages[2]

场景 · 用户「左边：把一张格式固定的发票转成标准模板（确定性工作）。右边：为一个新市场写一份调研简报（开放性工作）。」

可用工具plan_stepsweb_searchdraft_section

AGENTIC_LOOPPython

1# 模式一：固定链式 (steps hard-coded ahead of time)2def fixed_chain(raw_invoice: str) -> str:3    # Step 1 — 抽取字段（确定性，结构永远一样）4    fields = one_shot(5        "Extract sender, amount, due_date as JSON:",6        raw_invoice,7    )8    # Step 2 — 校验金额（纯规则，不需要模型自由发挥）9    checked = one_shot("Validate the amount field:", fields)10    # Step 3 — 渲染为标准模板11    return one_shot("Render this into the invoice template:", checked)12 13 14def one_shot(instruction: str, payload: str) -> str:15    resp = client.messages.create(16        model="claude-sonnet-4-6",17        messages=[{"role": "user", "content": f"{instruction}\n{payload}"}],18    )19    return resp.content[0].text20 21 22# 模式二：动态自适应分解 (model picks sub-tasks at runtime)23def dynamic_decompose(goal: str) -> str:24    messages = [{"role": "user", "content": goal}]25    while True:26        resp = client.messages.create(27            model="claude-sonnet-4-6",28            tools=tools,          # plan_steps / web_search / draft_section29            messages=messages,30        )31        messages.append({"role": "assistant", "content": resp.content})32        if resp.stop_reason == "end_turn":33            break34        results = []35        for block in resp.content:36            if block.type == "tool_use":37                out = run_tool(block.name, block.input)38                results.append({39                    "type": "tool_result",40                    "tool_use_id": block.id,41                    "content": out,42                })43        messages.append({"role": "user", "content": results})44    return messages[-1]["content"][0].text

Conversation Historymessages[2]

USER · 用户
把这张发票转成标准模板（左：确定性工作）。
ASSISTANT · Claude
Step 1 抽取（链上写死的一步）：sender=Acme Co. · amount=1280.00 · due_date=2026-07-15。

stop_reason

end_turn

工具执行 · 本步

本步未发生工具调用。

GUIDE · 引路

左边的工作是确定性的：发票字段永远是发件人、金额、到期日。所以三步直接写进代码——抽取、校验、渲染。模型不需要「想」流程，每一步只做一件被指定好的事。能写死就别让模型即兴发挥。

固定链式 · 写死的步骤 · 3 tools

1# 模式一：固定链式 (steps hard-coded ahead of time)2def fixed_chain(raw_invoice: str) -> str:3 # Step 1 — 抽取字段（确定性，结构永远一样）4 fields = one_shot(5 "Extract sender, amount, due_date as JSON:",6 raw_invoice,7 )8 # Step 2 — 校验金额（纯规则，不需要模型自由发挥）9 checked = one_shot("Validate the amount field:", fields)10 # Step 3 — 渲染为标准模板11 return one_shot("Render this into the invoice template:", checked)12 13 14def one_shot(instruction: str, payload: str) -> str:15 resp = client.messages.create(16 model="claude-sonnet-4-6",17 messages=[{"role": "user", "content": f"{instruction}\n{payload}"}],18 )19 return resp.content[0].text20 21 22# 模式二：动态自适应分解 (model picks sub-tasks at runtime)23def dynamic_decompose(goal: str) -> str:24 messages = [{"role": "user", "content": goal}]25 while True:26 resp = client.messages.create(27 model="claude-sonnet-4-6",28 tools=tools, # plan_steps / web_search / draft_section29 messages=messages,30 )31 messages.append({"role": "assistant", "content": resp.content})32 if resp.stop_reason == "end_turn":33 break34 results = []35 for block in resp.content:36 if block.type == "tool_use":37 out = run_tool(block.name, block.input)38 results.append({39 "type": "tool_result",40 "tool_use_id": block.id,41 "content": out,42 })43 messages.append({"role": "user", "content": results})44 return messages[-1]["content"][0].text