Prompt chaining: how to handle long or complex work with AI
Break big tasks into a sequence of focused prompts — research, outline, draft, review — to get higher-quality reports, proposals and plans from AI.
By the AiTrain editorial team · Reviewed under our editorial standards
Ask an AI assistant to "write a 3,000-word market report" in a single prompt and you will usually get something long, smooth and shallow. Split the same job into a chain of smaller steps and the quality jumps. This is prompt chaining.
Key takeaways
- Each prompt in a chain does one job and produces an output you check before moving on.
- Chains give you control points where your judgement improves the result.
- A standard four-step chain — gather, outline, draft, review — covers most writing tasks.
Why single prompts struggle with big tasks
A long request asks the model to plan, research, structure, write and self-edit all at once, in a single pass, with no chance for you to correct direction. Errors in the plan flow straight into the draft. By the time you read the output, fixing it means starting over.
The four-step writing chain
Step 1 — Gather
Here are my notes and three source documents. List the 10 most important points for [audience] about [topic]. For each, say which source it comes from.
Check: Are the points right? Is anything important missing? Add or remove before continuing.
Step 2 — Outline
Using these approved points, propose an outline for a [length] [document type]. Give section headings and one sentence per section describing its purpose.
Check: Is the order logical for the reader? Does each section earn its place?
Step 3 — Draft
Write section 2 only, following the outline. Use the tone of the example below. Keep it to about 300 words.
Drafting section by section keeps quality high and lets you redirect early.
Step 4 — Review
Act as a sceptical [manager/client]. Review the full draft below. List the five weakest points, any unsupported claims, and anything confusing. Don't rewrite it.
Then fix the issues yourself or with targeted prompts.
Other useful chains
| Task | Chain |
|---|---|
| Proposal | Client needs → objections → solution outline → pricing table → draft → critique |
| Job posting | Role outcomes → must-have skills → draft → bias check → final |
| Data analysis | Describe dataset → questions to answer → formulas/code → interpret results → summary |
| Lesson plan | Learning goals → activities → assessment → differentiation → handout |
| Research brief | Questions → sources (yours) → extract findings → compare → conclusions |
Tips
- Carry forward only what's needed. Paste the approved outline into the draft step rather than relying on a long chat history.
- Name the steps. "This is step 3 of 4" helps the model stay in its lane.
- Save the chain. A chain that works once will work again; store it in your team prompt library.
- Automate later. Once a chain is stable, it is a natural candidate for automation.
Chaining feels slower the first time. By the second or third use, it is faster than single prompts — because you stop rewriting outputs that went wrong at step one.