Apply task decomposition techniques to structure complex requests
Break a large or multi-part request into an ordered sequence of smaller prompts, each with one clear goal and a defined output that feeds the next, with human checkpoints between steps.
Key points
- 1
The help centre's core advice for complex work is to break the request into substeps. One prompt asking for extraction, analysis, a risk register and an email at once produces a shallow mix; four focused prompts produce four good outputs.
- 2
Order the steps so each produces what the next needs: gather or extract first, then analyse, then decide, then draft. For a course: outline and objectives, then lessons, then assessments aligned to the objectives.
- 3
Define the handoff. Each step should end with a concrete, named output (a numbered list of action items, a table of bottlenecks) that the next prompt references explicitly, rather than relying on a vague earlier conversation.
- 4
Put a human review between steps. Errors early in a chain propagate silently into everything downstream, so check intermediate outputs before feeding them forward.
- 5
Do not drop the input-gathering step because Claude "can infer" it. Claude does not know your current process, policy or data unless you supply it.
- 6
Do not isolate dependent steps in unconnected chats. If step three depends on step two's output, that output must be carried into step three's prompt.
- 7
For multi-day or multi-person work, create the shared foundation first (win themes, style guide, objectives), store it in a Project as instructions and knowledge, and run each bounded step in its own chat inside the Project.
- 8
Match each step to the right tool. Language models are strong at categorising, summarising and writing; counting hundreds of items, arithmetic and exact lookups are deterministic work for a spreadsheet or Claude's analysis tool. Verify these intermediate results and re-run only the steps that depend on them.
- 9
Decomposition also helps diagnosis: when the final output is wrong, the intermediate outputs show which step failed, so you fix that step instead of rewriting everything.
- 10
Get a plan first. For an open-ended task, ask Claude to propose the steps and outputs, approve or adjust the plan, then execute it step by step.
- 11
Common distractors: one giant prompt "so everything stays consistent"; re-running the same overloaded prompt with "be thorough"; collapsing steps to hide a failing intermediate; assigning the most capable model to step one and assuming later steps are just formatting.
Read the source
Test yourself on Apply task decomposition techniques to structure complex requests
Ten questions, with the answer and explanation after each one.