Prompt chaining means breaking one long request into a short sequence of prompts where each step produces something the next step can use. Multi step prompting works because each prompt only has to do one job, which makes the output easier to inspect and far easier to fix when something goes wrong at step three.
This is the pattern most major prompt engineering guides describe when they advise decomposing complex tasks, and it is the mechanism behind every multi step agent setup. Below is the reasoning for cutting a task, a copy-ready chain you can fill in today, and the checks that catch a broken hand-off.
Why one giant prompt underperforms
A single prompt that says research this, summarise it, argue both sides, and write the recommendation makes four demands at once. When the output is wrong you cannot tell which demand produced the failure.
Splitting the same work into steps has a second benefit that matters more: each step can be checked. A wrong research summary is caught before it becomes the basis of an argument.
| One giant prompt | Five small prompts |
|---|---|
| Hard to locate the failure | Failure is visible at one step |
| Rewrites everything each time | Rewrites touch one step |
| Long prompt, drifting constraints | Each prompt holds a few constraints |
| One answer to judge | Five checkable outputs |
The cost is coordination. Each hand-off loses some information, so the chain has to be explicit about what the next step receives.
Choose where to cut the chain
Cut where the output format changes
Research notes, then an outline, then prose. Each change of format is a natural cut, because the next step consumes a defined artefact rather than a vague description of one.
Cut where a decision belongs to you
Position, pricing, scope, and audience choices are yours. Put a hand-off before each one so the model asks rather than assumes.
Cut where the model would have to guess
If a step needs information the request does not contain, that is a cut point. Put the gap at the end of a step, ask for it explicitly, and resume.
A copy-ready five-step chain, step by step
Step 1: gather
Task: collect the raw material for [deliverable]. Topic: [topic]. Sources I can give you: [paste source text or note what you may and may not use]. Extract, do not interpret: [the specific items you need, for example claims, dates, and quoted lines]. Output as a numbered list where each item keeps the source label [label format]. Do not add commentary. If a required item is missing from the sources, write [MISSING: item].
Step 2: select
From the material below, keep only the items that support [the decision or argument]. Drop anything that is off-topic, duplicated, or unsupported by the item beside it. Output the kept items with their source labels and one line each on why they survived. Then list the items you dropped and the reason. Do not write the final answer yet.
Step 3: structure
Using only the kept items below, produce an outline for [deliverable]. Format: [headings and word budget per section]. Mark each section with the item labels it relies on. If any section has no supporting item, write [NO SUPPORT: section] instead of filling it. Do not write the prose.
Step 4: draft
Write the section [section name] from the outline below. Use only the item labels listed for that section. Word budget: [budget]. Tone: [tone in three words]. Do not introduce a claim that has no item label. Where you must stretch past the evidence, write [STRETCH: claim] so I can judge it.
Step 5: check
Review the draft below against the outline and the item list. Report: claims with no supporting item, numbers or names that do not appear in the sources, instructions from my brief that were not followed, and anything promised in the outline but missing from the draft. Output a table with the issue, the location, and the fix. Do not rewrite the draft.
Fill in these details
- Deliverable: one artefact, named, so every step knows what it is building toward.
- Source labels: how you mark provenance. Without labels, step two cannot show its reasoning and step four cannot stay grounded.
- Word budget: per section, not total, so one section cannot eat the whole piece.
- Drop reasons: a short reason per dropped item, which is where you catch a wrong selection.
- Missing markers: the bracket convention for gaps. Keep it identical across all five prompts so you can search for it in one pass.
Worked example: turning research into a decision brief
A builder needs to choose a database for an internal tool and has five candidate options documented.
| Step | Input | Output | Why it is separate |
|---|---|---|---|
| 1. Gather | Documentation for five options | Numbered list of claims with labels | Interpretation at this stage contaminates the facts |
| 2. Select | The claim list | Kept claims plus dropped claims with reasons | You can check the selection before it shapes the brief |
| 3. Structure | Kept claims | Outline with labels per section | Catches unsupported sections early |
| 4. Draft | One section plus its labels | Prose with no unsourced claims | Drafting one section keeps quality high |
| 5. Check | Draft plus outline | Issue table | A review pass needs the original labels to work |
The value is at step two. It is the cheapest place to catch a wrong filter, and it happens before any prose exists.
Check the joins, not just the outputs
- Trace one item forward. Pick a claim from step one and find where it appears in the final draft. If it stops anywhere, you know the exact join that dropped it.
- Compare counts across steps. Five claims in, three sections supported, one claim unused. A large drop means step two was too aggressive.
- Read the gap markers. Every [MISSING] or [NO SUPPORT] is a question only you can answer. Resolve them before step five.
- Re-run a single step, not the chain. Fixing step three should not require regenerating steps one and two.
Common prompt chaining mistakes
- Cutting where there is no real change of task. Every cut adds coordination cost. Cut at decisions, format changes, and missing information.
- Handing over prose when a list would do. Structured input gives the next step less to misread.
- Letting step four invent support. The label requirement only works if the draft prompt forbids new claims explicitly.
- Running all five steps before checking any. The point of the chain is early detection.
- Saying chain prompts when you have not checked the join. A broken hand-off produces a confident, wrong final answer.
Frequently asked questions
What is prompt chaining in simple terms?
Running several prompts in sequence where each one takes the previous output as input. The work is decomposed so each prompt has a single job and each output can be inspected before the next step relies on it.
How many steps should a chain have?
Enough to make each step checkable. Five is a useful default for research, structure, drafting, and review work, but a two-step chain is right for simple tasks and longer chains only earn their length when each cut is justified. Chaining also costs more tokens, since each output is passed forward and read again, so the extra steps earn their place when a late failure would be expensive or the work must be repeatable.
Can I automate a chain?
Yes, in agent tools that support sequential steps or tool calls. The design rules are the same: explicit hand-offs, one job per step, and a check before the next step depends on the result.
Run one chain end to end today
Pick a task you have already done twice by hand, break it into five steps with the prompts above, and run it once with checks between steps. When you want the reusable instruction patterns behind each step, browse the system prompt topic collection.