One prompt that summarizes a newsletter, writes five hooks, and adapts the best one for social tends to return a draft that's passable at all three jobs and sharp at none. The fix is usually a shorter prompt, repeated: one to plan, one to draft, one to check, each starting from the previous step's output.
What prompt chaining is
Prompt chaining is breaking a multi-part AI job into a sequence of smaller prompts, where each step's output becomes the next step's input. The connection is what makes the chain: step two starts from what step one produced, so the job's context carries forward without any single message carrying all the instructions. The simplest version runs as messages in one chat window. Saved steps in a workflow, or code passing files between tools, come later, when a chain repeats.
A worked chain
A newsletter shows the shape. One prompt covering the whole job, summarize, write hooks, and adapt for other platforms, has to juggle three instruction sets at once. Split across steps, the same job runs as four links:
- Paste the newsletter and ask: "Summarize the three main points, one sentence each."
- Read the three sentences against the original. They're short; a check takes a minute.
- Ask: "Draft five opening hooks for a post based on point two."
- Pick the hook that sounds most like you and ask for the full post from it.
Step two is the checkpoint. If the three sentences describe the newsletter as "three ways to use AI at work" when the piece is mostly about checking AI output before sharing it, the hooks built on that summary inherit the wrong emphasis. Fixing the summary costs half a minute; fixing a finished post built on it costs the post. Each link has one job and one instruction set, and the intermediate results stay visible, where a single-prompt draft buries the same miss in a wall of text.
Why the steps beat one big prompt
Three things improve when a job splits into links.
Instructions stop competing. A prompt doing five jobs carries five instruction sets in one message, and each addition pulls the model's attention across all of them. A chain gives each job its own message with its own instructions.
Each link is checkable. The three-sentence summary from step one is easy to read against the source, and a wrong step gets caught before it becomes the foundation for three more.
You can steer between links. One prompt commits you to its whole plan up front. Between links, the next instruction can change based on what came back, and that correction is the part a single prompt can't do.
Where chains break
The failure modes are predictable, and each has a check.
Mistakes travel. Each step's output is the next step's input, so an error in link one doesn't get corrected downstream, it gets built on. The check between links is the fix, and so is fixing the right end: re-run the weak step with a sharper instruction rather than hand-editing its output, because the hand-edit repairs one draft while the sharper instruction repairs the runs after it.
Setup has a cost. Each link adds a prompt to write and an output to check. A job that fits one prompt, like tightening a paragraph, shouldn't become four links. Chains earn their keep when the job has distinct phases with distinct instructions, or when one message keeps missing part of the task no matter how it's reworded.
Long conversations drift. A chain running as one long chat carries everything said before it, and earlier instructions can bleed into later steps. When a chain runs past a few links, start a fresh conversation per step and paste in the previous output. The chain survives the move, because the connection between links is just the output itself.
From chain to workflow
A chain you run once is a technique. A chain you run every week is a workflow waiting to be written down. This site's publishing pipeline runs as a chain: collect ideas, draft from an outline, check the facts and the voice, one pass each with its own check. When the steps stay stable for a few runs, each link's instructions get saved so the chain runs the same way without re-explanation, and a chain on a schedule becomes a candidate for automation.
Break the job into links, check each one, fix instructions rather than outputs, and write the chain down when it repeats. The pieces, when you need them:
- What Is an AI Workflow? Building Your First One: the container a chain becomes once the steps stabilize.
- How to Build an AI Content System: a chain at production scale, link by link.
- How to Automate Repetitive Tasks with ChatGPT: when the chain runs on a schedule instead of you.
FAQ
Does prompt chaining require coding?
No. The common version runs as successive messages in one chat: each message handles one job, and its output feeds the next request. Code shows up only when a chain repeats on a schedule or hands files between tools, which is optional and comes later.
What's the difference between prompt chaining and a prompt template?
A template is one reusable prompt with blanks to fill, like "write a [format] for [audience]". A chain is a sequence of prompts connected output-to-input. They combine easily: a chain's first link can be a template you reuse every week.
How long should a prompt chain be?
As short as the job allows. Most chains we run are two to four links, and each link needs a job the previous one can't do, because each link costs a prompt to write and a check to run.
What's the difference between prompt chaining and an AI agent?
In a chain, you run each step and decide when to move on. An agent plans and runs several steps on its own and shows you the result at the end. Chains are the version where you hold the steering wheel; agents hand the driving to software, which pays off only after the steps are proven.