Task statement 1.2 · Prompting and task execution

1.2 · Task Decomposition

In Wrenfield Housing, the course example, the communications team writes an annual tenant report from three files of figures. Asked for in one message, the report comes back polished, with a wrong figure hidden in eight pages. Task decomposition, which splits a complex request into smaller steps, lets the team check each stage.

Why it's tested

The tempting move is to put the whole request into one long, detailed prompt, because one message feels faster than four. That prompt asks Claude to read the sources, choose the figures, plan the structure and write, all in one reply. The in-between work stays inside the reply, so the first thing you can check is the finished document. By then, an early error is buried in polished prose. The opposite habit fails too, because splitting a simple request into many messages only adds time.

KnowledgeTwo ways to break a request down

Task decomposition means splitting a complex request into smaller steps, each with one job. The Claude Help Center puts it briefly: break down complex requests into substeps. The first technique works inside one prompt, where you write the steps as a numbered list. Anthropic's prompting guide recommends a numbered list when the order or completeness of the steps matters. Claude still answers in one reply but follows your order.

The second technique is prompt chaining: you split the work into separate prompts, each handling one stage, and the output of one becomes the input of the next. In the Claude apps you chain by hand, sending each step's message after reading the last reply. A chain that must run automatically is a job to escalate to a developer or architect. Chaining takes longer, but it makes each step easier and often makes the result more accurate. It also shows you each in-between result. Anthropic's guide gives this as a reason to chain: you can inspect the intermediate outputs. This rules out any answer that only lengthens one prompt when the need is to check the work partway through.

To structure a complex request as a chain, work in this order:

  1. The result: what the finished piece is, who reads it, and who checks it.
  2. The stages: the jobs the work passes through, such as gathering facts, planning the structure, writing and reviewing.
  3. One job per message: each prompt asks for one stage and names the output it starts from.
  4. A check between steps: read each output, and correct it before the next step builds on it.
  5. A review at the end: ask Claude to review the draft against your criteria, then to revise it.

Interactive · where you can check the annual report

The annual tenant report as a chain of four steps, read from top to bottom. Three files feed step 1, a figures table. The checked table feeds step 2, an outline. The checked outline feeds step 3, a draft written one section at a time. The draft feeds step 4, where Claude reviews and revises it. Beside each step is the check you make on its output. When the whole report is asked for in one message, the four steps are dashed and labelled hidden in one reply, and the checks are dashed and labelled not shown, except the final check of all eight pages at once. INPUTS WHAT YOU CHECK repairs summary survey results finance summary THREE FILES 1 · Figures table key figures from each file, with its source NOW CHECKED HIDDEN IN ONE REPLY SHOW YOU CHECK NOT SHOWN each figure against its file CHECKED TABLE 2 · Outline the sections and the point each makes NEXT NOW CHECKED HIDDEN IN ONE REPLY SHOW YOU CHECK NOT SHOWN every topic has a section CHECKED OUTLINE 3 · Draft by section one section per message, from steps 1 and 2 NEXT NOW CHECKED HIDDEN IN ONE REPLY SHOW YOU CHECK NOT SHOWN each section as it arrives FULL DRAFT 4 · Review and revise Claude checks the draft against the outline NEXT NOW NOT ASKED FOR SHOW YOU CHECK the final read-through all eight pages at once REVISED REPORT Each check comes before the next step builds on the result.

Choose how to ask for the annual report, and see where you can check the work.

One message: every stage happens inside one reply, so the first thing you can check is eight finished pages.

Step 1: the key figures come back as a short table. A wrong number is one row to fix, not a sentence on page five.

Step 2: from the checked table, Claude proposes the sections. You fix the outline before any prose depends on it.

Step 3: one section per message, short enough to read properly as it arrives.

Step 4: Claude reviews the draft against the outline and revises it. The communications lead's read-through comes last.

Dashed, labelled NOT SHOWN or NEXT: nothing to check yet. Solid, labelled YOU CHECK: you see the output before the next step uses it.

The last step is what Anthropic's guide calls self-correction, its most common chaining pattern: generate a draft, have Claude review it against criteria, then have Claude refine it. Anthropic's blog shows it with a research summary: one prompt asks for a summary, the next asks Claude to review it for accuracy, clarity and completeness, and the third asks it to improve the summary from that feedback. Paste any output you corrected into the next message, wrapped in tags such as <figures>, so Claude builds on your version, not its first one.

KnowledgeSteps in one prompt, or a chain of prompts

The deciding question is whether you need to see the work in between. Numbered steps in one prompt keep the order without slowing you down, which suits a request such as summarising a meeting and then listing its actions. A chain suits a request where an early mistake would spread, or where several sources must be read before anything is written. Anthropic's blog adds two more signs: the checks in between add value, or a single prompt gives inconsistent results. The two techniques compare like this:

One reply

Numbered steps in one prompt

Claude follows your order inside one reply: fast, and enough when you only need to check the result.

Fits: a short request where the order of the work matters.

One message per stage

A chain of prompts

Each stage is its own message, read before the next: slower, but each stage is easier to do and to check.

Fits: complex work where an early error would spread.

SkillWorked example: the annual report in one message, then in four

Wrenfield's annual tenant report for 2025–26 runs to about eight pages on repairs, the tenant satisfaction survey, rent and spending, and next year's plans. It is built from an anonymised repairs log summary, the survey results and the year's finance summary, with last year's report as the style guide. The first attempt asked for everything in one message:

# before: one message
Using the three attached files, and last year's report for style,
write our 2025–26 annual tenant report: repairs, survey, spending, plans.
→ eight polished pages: four paragraphs on repairs, two sentences on the survey,
  "81% of tenants were satisfied with repairs" (the survey says 71%)
# after: one message per step, each output checked before the next
1  List the key figures from each file in a table, with the file each came from.
2  From the checked table, propose the sections and the point each one makes.
3  Write the survey section from the table and outline. (Then each other section.)
4  Review the draft against the outline and last year's style, then revise it.
→ 71% corrected in the step 1 table; the survey gets its own section at step 2

The report read well but was uneven, with a wrong figure on page five. The cause was five jobs in one reply: reading three files, choosing figures, planning, writing and matching a style. None of that work was visible, so the first chance to check anything was eight finished pages. The chain does not stop Claude misreading a figure, but it puts the figure in a short table where the error is quick to spot. The outline gives the survey its own section. The communications lead still reads the final report, as Wrenfield's AI policy requires, but now as the last check, not the only one.

Not every request needs a chain. Anthropic's prompting guide notes that Claude handles most multistep reasoning on its own, and chaining adds time. A one-page letter needs one good prompt.

Ask one question: will you need to check or correct a result partway through, before the rest of the work builds on it? Yes → chain it: one message per stage, and check each output before the next. No → send one prompt, with numbered steps if the order matters.
Takeaway
  • Task decomposition splits a complex request into smaller steps with one job each: numbered steps in one prompt, or a chain of separate prompts.
  • A chain trades time for accuracy and shows you each in-between result, so an early error is caught before it spreads.
  • End a chain with self-correction: Claude reviews its draft against your criteria, then revises it.

check yourself

0 / 3 answered

A marketing manager asked Claude, in one message, to compare five competitors' product brochures and recommend which product the company should match. The recommendation reads well, but the manager cannot tell which features came from which brochure. A director will act on it, so the manager must check the comparison before any recommendation is built on it.What is the most effective way to restructure the request?

Select one.

An HR coordinator is asked for a two-page guide for new starters, drawn from a 40-page staff handbook. The guide must cover only what applies in a new starter's first week, and the HR manager will review it. Asking for the guide in one message gave a draft that wandered into pensions and long-service leave.Which TWO steps best structure this request?

Select two.

A charity's finance officer sent Claude one message with twelve monthly expense exports, asking for yearly totals by category, a comparison with the budget and a short note for the trustees. The note is clear, but its category totals do not add up to the year's total. The trustees must get figures the officer has checked personally.What is the most effective first step?

Select one.

sources

  1. Claude Platform Docs, "Prompting best practices": sequential steps as a numbered list when the order or completeness of steps matters; "Chain complex prompts": Claude handles most multistep reasoning internally, explicit chaining is still useful to inspect intermediate outputs or enforce a specific pipeline, and self-correction (generate, review against criteria, refine) is the most common chaining pattern. Checked 2026-10-15.
  2. Claude Blog, "Best practices for prompt engineering": chaining breaks complex tasks into smaller sequential steps with separate prompts, each output feeding the next; the prompts can be given by hand after each response; it trades latency for accuracy; when to use it; the three-prompt research summary example. Checked 2026-10-15.
  3. Claude Help Center, "My prompt isn't giving me a helpful answer": break down complex requests into substeps. Checked 2026-10-15.
  4. Wrenfield's annual tenant report, its three files, the 71% survey figure and the one-message draft are this course's example.