The short answer
Prompt chaining means breaking one large job into a sequence of smaller prompts, where the answer to each prompt becomes the input to the next. Instead of asking an AI assistant to do everything in one go, you ask for step one, check the result, feed it into step two, and so on until the job is done.
That is the entire technique. There is no tool to install and nothing to configure. It works in any chat assistant you already use, and it is one of the most reliable ways to get better results on anything longer or more complicated than a quick question.
Why one giant prompt goes wrong
When you stuff a whole job into a single prompt ("summarise these notes, then write a client email about them, make it friendly but firm, and keep it short"), you are asking the model to juggle several tasks at once, making every decision in one pass with no chance for you to steer.
Chaining fixes three problems at once.
Each step gets the model's full attention. A prompt that asks for one thing tends to get that one thing done properly. A prompt that asks for five things tends to get five things done adequately, with the hardest one done worst.
You get checkpoints. Between steps, you can read the output, fix an error, cut something irrelevant, or add a fact the model could not know. The next step then builds on corrected material rather than on a mistake.
A wrong turn costs one step, not the whole job. If step two goes sideways, you rerun step two. With a mega-prompt, one misread instruction can derail the entire output, and your only option is to start again.
A useful mental model: you would not ask a colleague to research, decide, draft, and polish in a single breath. You would review the research before the decision and the draft before the polish. Chaining gives you those same review points with an AI.
A worked example: meeting notes to client email
Here is a chain most office workers will recognise. You have a page of messy meeting notes and you owe the client an update email. Three prompts, run one after the other in the same conversation.
Step 1: extract, do not write. First, pull out only what matters.
Here are my raw notes from today's client meeting:
[paste your notes]
List only: the decisions made, the actions agreed and who owns each one, and any open questions. Do not summarise the discussion. If an owner or date was not stated, write "not stated" rather than guessing.
Now pause. Read the list. Fix anything the model misread, delete anything the client should not see, and add anything it missed. This checkpoint is the whole point of chaining.
Step 2: draft from the corrected list.
Using the corrected list above, write an update email to the client.
Audience: the client's project lead, who was in the meeting.
Structure: one sentence of thanks, the decisions as bullets, the actions with owners, then the open questions as things we will confirm.
Under 180 words. Do not add anything that is not in the list.
Step 3: tighten the tone. The draft is usually right in substance and slightly off in voice, so fix only that.
Keep the structure and every fact exactly as they are. Rewrite it to sound warmer and less formal, the way you would write to someone you have worked with for a year. Cut any sentence that exists only to be polite.
Three small prompts, two chances to catch problems, and an email you can send after a light edit. The mega-prompt version of this task routinely invents a deadline, buries a decision, or writes in a voice you then have to fight.
Chaining is not chain-of-thought
These two techniques have confusingly similar names and get mixed up constantly. They are different tools for different jobs.
Chain-of-thought is a single prompt in which you ask the model to reason step by step before answering, so its final answer can build on its own written reasoning. Prompt chaining is several separate prompts, with you standing between them, checking and correcting each output before it feeds the next.
| Chain-of-thought | Prompt chaining | |
|---|---|---|
| How many prompts | One | Several, in sequence |
| Who does the steps | The model, inside one answer | You and the model, taking turns |
| Where the checking happens | After the answer, by reading the reasoning | Between every step, before the next one runs |
| Best for | Maths, logic, decisions with several constraints | Long or multi-part jobs: research then draft then edit |
You can happily use both at once: one link in your chain can ask the model to think step by step. Our Prompt Engineering course has a full lesson on chain-of-thought if you want that side of the story.
When not to bother
Chaining has a cost: more prompts, more reading, more of your attention. For most everyday requests, that cost is not worth paying. If one clear prompt gets the job done ("rewrite this paragraph in plainer English", "give me five title ideas"), chaining just adds friction.
Reach for a chain when the job has natural stages, when the output is long enough that a single wrong assumption would poison all of it, or when the first attempt at a mega-prompt came back muddled. That muddle is usually the model telling you the job wanted to be a chain.
❌ Weak prompt
Prompt
Here are my meeting notes: [notes]. Summarise them, write a client email about the key points, make it friendly but professional, keep it short, and also list any risks I should flag internally.
Output
A single block that blurs the client email and the internal risk list together, summarises the wrong points, and adopts a tone somewhere between friendly and legal. Fixing it means untangling four tasks at once.
Four jobs, two different audiences, one prompt. The model had to make every decision in one pass, and you have no checkpoint to catch the moment it picked the wrong key points.
✅ Good prompt
Prompt
Here are my meeting notes: [notes]. First, list only the decisions made, the actions agreed with owners, and any open questions. Do not write the email yet.
Output
A short, checkable list of decisions, actions and open questions, with nothing invented and nothing else attached.
This is just the first link of a chain. It does one job, produces something you can verify in thirty seconds, and gives every later step a clean foundation. The email and the risk list become their own steps, each aimed at the right audience.
Where to go next
If chaining makes sense to you, you already understand the idea behind it: what you get out depends on how you structure what you put in. That idea is the whole of prompt engineering, and our free Prompt Engineering course teaches it properly, with interactive examples and no account required. For the wider picture, start with What Is Prompt Engineering?.