Break multi-step work into Claude-friendly chunks
Goal: Stop dumping a whole project into one prompt—chunk into plan → approve → execute step → check.
Best for: Reports, multi-email campaigns, study plans, small automations you still supervise.
“Cowork” here means working *with* the model across steps, not letting an agent roam unsupervised. Each chunk has a done check before the next. This pairs with human-in-the-loop checkpoints and blast-radius thinking for anything that sends or deletes.
Multi-step “cowork” modes need a written Done and pause points. Without them, the agent keeps going past your comfort.
Write the pause phrase you will paste when nervous: “Stop. Show plan and wait.” Keep it in your notes app.
Why beginners skip this
Skipping the hard part of “Break multi-step work into Claude-friendly chunks” feels fine until a live thread pays the bill. Hold: Plan, approve, one step, check—never the whole project in one gulp—prove it on a real artifact, or skip the badge Mid-week cost shows up as rework, not as a missing bookmark.
Step 1 — Plan first
Do this step with a real artifact open—not from memory.
Task: [big goal]
Return a numbered plan of ≤7 steps.
Mark which steps need my OK before continuing.
Do not execute yet. Flag missing inputs as [TODO].
Step 2 — Execute one step
Example
Filled prompt
Goal: 3-email onboarding sequence for product X.
5-step plan ONLY. I approve before execute step 1. Ban one-gulp full copy.
What good output looks like
- Pulls from Project/file named in the ask for “Stop dumping a whole project into one prompt—chunk”
- Standing instructions respected; no silent scope creep
- Decisions quoted or labeled Unclear — never “sounds good”
- One-minute check: does output match claude-cowork-multi-step-tasks Done?
Step 3 — Check, then unlock next
After each chunk: run a 60-second eval (claims, fluff, next step). Freeze facts. High blast-radius steps stay human-gated.
Step 2 done. Critique against: clarity, invented facts, missing [TODO]s. Scores + fix list only. Do not start step 3.
One Artifact or one thread
Keep the plan and outputs in one place. Six parallel chats for one project recreate the mess you tried to escape.
Restate the approved plan steps 1–N and which are done. Then wait.
Example
Filled prompt
Goal: ship onboarding email sequence (3 emails) for product X.
Return a 5-step plan ONLY (no copy yet). I will approve before step 1 execute.
Each step needs Done + check. Ban one-gulp full draft.
What good output looks like
- Standing instructions respected; no silent scope creep
- Decisions quoted or labeled Unclear — never “sounds good”
- One-minute check: does output match claude-cowork-multi-step-tasks Done?
- Pulls from Project/file named in the ask for “Stop dumping a whole project into one prompt—chunk”
Save this card
Card: Plan, approve, one step, check—never the whole project in one gulp—prove it on a real artifact, or skip the badge. Keep this beside the tool; read before you start.
1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)
- [ ] Tools/writes paused
- [ ] Refuse-line dated
- [ ] Artifact path saved
- [ ] Next link chosen (or stop)
Before you start
Open the Claude Project you will keep. Confirm standing instructions + the knowledge file for “Stop dumping a whole project into one” are current. Guide: claude-cowork-multi-step-tasks. If you only have a blank practice box, stop and fetch the real job first.
- Goal in one line: ___
- Allowed tools: ___
- Forbidden: ___
- Human approve needed for: ___
How to use this week
This week in Claude: break one real multi-step job into plan → approve → one step → check. Refuse a single mega-prompt. Keep the step card on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed claude-cowork-multi-step-tasks, put 25 minutes on the calendar and redo once.
Pause phrase ready
Keep “Stop. Show plan and wait.” in notes. Paste it when nervous.
Ship the practice
Close the tab only after Practice now is done on a real artifact.
Do not bookmark-only
Finish Practice now, then tick this guide on your path. Titles unread are not skills.
Close the loop
After Practice now, write one line: what you will refuse next time. That line is the habit.
Check the unknown path
What should the model do when a field is missing—null, TBD, or ask you? Write it once in Boundaries.
Example
Filled prompt
Project: translate + localize 3 help articles EN→ZH-HK.
Plan 5 steps with approve gates. Execute step 1 only (glossary freeze).
Ban one-gulp all three.
What good output looks like
- Pulls from Project/file named in the ask for “Stop dumping a whole project into one prompt—chunk”
- One-minute check: does output match claude-cowork-multi-step-tasks Done?
- Standing instructions respected; no silent scope creep
- Decisions quoted or labeled Unclear — never “sounds good” [claude-cowork-multi-step-tasks#3]
Example
Filled prompt
Localize 2 help articles.
Plan 5 steps with gates. Execute glossary freeze only today.
Context: real workstream this week (Asia/Shanghai). Name actual files/people.
Constraints: no invented metrics/owners/prices; label Unknowns; stop before send/delete/pay.
Done check: I can mark pass/fail in one minute on the output.
What good output looks like
- Decisions quoted or labeled Unclear — never “sounds good”
- Pulls from Project/file named in the ask for “Stop dumping a whole project into one prompt—chunk”
- One-minute check: does output match claude-cowork-multi-step-tasks Done?
- Standing instructions respected; no silent scope creep
Red flag
Starting chunk two before chunk one is accepted; goal drift between chunks; current chunk undefinable—stop.
Pitfalls
| Pitfall | Fix |
|---|---|
| One prompt for the whole project | Plan → chunks |
| Skipping OK gates | Explicit approve lines |
| Letting steps invent data ahead | [TODO] rule |
| Parallel chats for the same plan | One thread or Artifact |
Practice now
Live card: one Claude artifact/diff/email pass for “Break multi-step work into Claude-friendly chunks” with Blind before share (Asia/Shanghai).
Today: split one multi-step job into 3–5 chunks with an acceptance line each. Finish and accept chunk one before starting chunk two (Asia/Shanghai).
Today card (copy) — claude-cowork-multi-step-tasks: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___
- [ ] Artifact path written
- [ ] Refuse example would say "yes"
- [ ] Dated Asia/Shanghai
One sentence takeaway
Plan, approve, one step, check—never the whole project in one gulp—prove it on a real artifact, or skip the badge
- You can point to a dated refuse-line
- You have one artifact path from tonight
- You can teach the red flag in one sentence
Next
HITL checkpoints · Chaining · Blast radius · Claude Projects knowledge
Pair with today: open the first Next link (agents-human-in-the-loop-checkpoints) only after you dated a refuse-line for this guide (Asia/Shanghai).
把多步工作拆成 Claude 好消化的块
目标: 停止把整个项目倒进一条提示——拆成计划 → 批准 → 执行一步 → 检查。
适合: 报告、多封邮件活动、学习计划、你仍监督的小自动化。
这里的「协作」指跨步骤与模型共事,不是让智能体无人看管乱走。每一块在下一块前有完成检查。凡涉及发送或删除,配合人在回路与爆炸半径思考。
多步骤「协作」模式需要写明的完成标准与暂停点。没有它们,智能体会冲过你的舒适区。
写好紧张时要粘贴的暂停句:「停下。显示计划并等待。」放进笔记应用。
为什么新手跳过
一个超长提示词塞进整项目显得果断,却会把拐错弯拖到很晚才发现。改成「计划→你批准→执行→检查」分块,才能在第一块就拦住跑偏。
步骤 1——先计划
用真实材料做这一步——不要凭记忆。
任务:[大目标]
返回 ≤7 步编号计划。
标出哪些步需要我 OK 才能继续。
先不要执行。缺输入标 [待办]。
步骤 2——只执行一步
实例
填好的提示词
规划多步任务:收集竞品定价页链接 → 起草对比表 → 任何邮件前停下来等我批准。写出完成标准 + 暂停点。
像样的输出大概长这样
- 完成标准明确
- 发信前暂停
- 步骤有序
- 贴合填例里的约束
- 不编造事实、负责人或指标
- 新手能看懂;可直接交给同事
步骤 3——检查,再解锁下一步
每块之后:跑 60 秒评估(主张、空话、下一步)。冻结事实。高爆炸半径步骤保持人闸。
第 2 步完成。按清晰度、编造事实、未填 [待办] 批评。只要分数 + 修改清单。不要开始第 3 步。
一个 Artifact 或一个线程
计划和产出放一处。一个项目开六个并行聊天,等于重回你想逃离的混乱。
复述已批准计划步骤 1–N 及哪些已完成。然后等待。
实例
填好的提示词
运行已起草邮件。写一段我可粘贴的暂停话术:停下、显示差异、等待。
像样的输出大概长这样
- 干净停下
- 显示差异
- 等待
- 贴合填例里的约束
- 不编造事实、负责人或指标
- 新手能看懂;可直接交给同事
保存这张卡
卡面:块列表 + 每块验收句。开下一块前看卡;上一块未验收不许开下一块;验收句写不清就停,不硬开。
1. 我交出了:___ 2. 我拒绝了:___ 3. 我永不再做:___(日期:Asia/Shanghai)
- [ ] 工具/写入已暂停
- [ ] 拒做句已标注日期
- [ ] 成品路径已存
- [ ] 已选下一步(或停)
开始前
打开你真正会用的账号/项目。准备好与「把多步工作拆成 Claude 好消化的块」相关的真实材料(不要空白练习框)。
默认姿态:在人确认发送/写入之前只出草稿。授予新工具前,先写一条带日期的拒做句(Asia/Shanghai)。
- 一句话目标:___
- 允许的工具:___
- 禁止:___
- 需要人批准:___
本周怎么用
本周一件多步活:拆成 3–5 块,每块写验收句,只批准当前块。把过/不过写进笔记并注明日期(Asia/Shanghai)。
暂停话术备好
笔记里留「停下。显示计划并等待。」紧张时粘贴。
把练习做完
实卡:为「把多步工作拆成 Claude 好消化的块」做一次 Claude 产物/diff/邮件;分享前盲读(Asia/Shanghai)。
今天把「把多步工作拆成 Claude 好消化的块」练完:真实材料、过/不过、一句永不做。注明日期(Asia/Shanghai)。
不要只收藏
做完「现在就练一次」,再在路径上打勾。没练过的标题不是技能。
闭环
「现在就练一次」后写一行:下次你要拒绝什么。那一行才是习惯。
检查未知路径
字段缺失时模型该怎样——null、TBD,还是问你?在边界里写一次。
红旗
上一块还没写出可勾选的验收句,就开下一块;块与块之间目标悄悄漂移。说不清「当前块交付是什么」——立刻停。每块先写验收句再开工。
常见坑
| 坑 | 改法 |
|---|---|
| 一条提示做完全项目 | 计划 → 分块 |
| 跳过 OK 门 | 明确批准行 |
| 步骤提前编数据 | [待办] 规则 |
| 同一计划开并行聊天 | 一个线程或 Artifact |
实例
填好的提示词
任务:Stop dumping a whole project into
场景要点:Claude Project 或长文材料打开;练习编号 3/4。
边界:仅草稿;未知 [?];禁止悄悄发送;禁止编造指标。
返回:可检查产出 + 过/不过理由一行。
像样的输出大概长这样
- 约束贴着「把多步工作拆成 Claude 好消化的块」与 claude-cowork-multi-step-tasks
- 未知标 [?];不编造负责人/数字
- 不可逆动作前停下
- 一分钟内能打过/不过(claude-cowork-multi-step-ta
- 约束贴着「把多步工作拆成 Claude 好消化的块」与 claude-cowork-multi-step-tasks
- 未知标 [?];不编造负责人/数字
- 不可逆动作前停下
- 一分钟内能打过/不过(claude-cowork-multi-step-tasks #4)
- 过/不过:≤60 秒能打勾(claude-cowork-multi-step-tasks #4)
现在就练一次
今天(Asia/Shanghai)把一件真实任务拆成≥3 块,写下第一块验收句并跑完第一块。过:验收句可勾选且材料在手;不过:块边界含糊。只收藏标题算失败。
今日卡片(可复制)— claude-cowork-multi-step-tasks: 1. 我要产出的成品:___ 2. 通过/失败标准:___ 3. 若想跳过核验/发送的拒做句:___(日期:Asia/Shanghai) 4. 本篇一条永不做:___
- [ ] 成品路径已写
- [ ] 拒做例句会说「可以」
- [ ] 已标注 Asia/Shanghai 日期
一句话带走
多步协作先切成可单独验收的小块,上一块勾选通过再开下一块,比一次抛出整包更稳,也更早拦住跑偏。本周拆一件真活,只完成并验收第一块就停,记下下一块何时再开
- 能指出一条带日期的拒做句
- 今晚有一条成品路径
- 能用一句话说清红旗
下一步
人在回路 · 链式 · 爆炸半径 · Claude Projects knowledge
今日搭配:先为本篇写上拒做句日期(Asia/Shanghai),再打开下一步里的第一篇(agents-human-in-the-loop-checkpoints)。
← Back to home · Learn · Tools · Compare · Start