Break big jobs into steps: prompt chaining explained
One giant prompt gives you something that looks finished but is hard to check. Break the job into steps with checkpoints and you get better work you can trust and reuse.
In 30 seconds
- Split a big job into steps, such as research, outline, draft, critique and revise.
- Each step’s output feeds the next, so check at every checkpoint before errors travel.
- Save chains that work as templates or projects, and automate the ones that never change.
Ask an AI for a finished 2,000-word report in one go and you’ll get something that looks done: headings, a conclusion, confident figures. Look closer and it’s often thin where it matters, padded where it doesn’t, and hard to check.
Jargon busterPrompt chaining: Splitting a big task into a series of prompts, where each step’s output becomes the next step’s input, with a check in between. is the fix. You split the job into steps, run each as its own prompt, and feed each step’s output into the next, checking it on the way.
What a chain looks like
- ResearchGather sources and notes
- OutlineAgree the structure
- DraftWrite it a section at a time
- CritiqueScore it, then revise
Anthropic, which makes Claude, sums up the trade-off: a chain gives up a little speed for accuracy, by making each step an easier task. It suits jobs that split cleanly into fixed stages. Newer models do more multi-step work on their own, especially in thinking mode, so chain when you want control: a set structure, a checkpoint, a person signing off.
Why it beats one giant prompt
- Better qualityEach step is one clear job, so the model has less to juggle and does each part better.
- Easier to checkA flawed outline is quick to fix. The same flaw buried in a finished report is not.
- Easier to fixWhen something goes wrong, you can see which step caused it and rerun just that one.
- ReusableOnce a chain works, next month’s run is the same prompts with new material.
Three chains to try
Each step still needs a clear prompt, with the task, context and format: see how to write a good prompt. Run the steps in one chat, pasting in the checked output from the step before, or in fresh chats if you want each step to start clean.
- A report. Gather sources into notes, then agree an outline, then draft a section at a time, then critique against your criteria, then revise.
- A monthly newsletter. List the month’s updates, then pick the five that matter most to readers, then draft in your house style, then check every fact and link, then suggest three subject lines.
- Tidying data. Describe the columns, then list the problems, such as duplicates and mixed date formats, then agree the fixes, then apply them to a copy, then compare row counts and list every change.
I’m writing a report on [topic] for [audience], to help them decide [decision]. Don’t write the report yet. From the sources below, pull out the key facts, figures and arguments as bullet notes, each tagged with its source. Mark anything uncertain or contradictory. [paste or upload your sources]
Make it critique against criteria
The most useful link in many chains is a critique: draft, review against criteria, then revise. Anthropic calls this pattern self-correction. The criteria do the work. “Make it better” gets vague tweaks; “every figure has a source” gets a list of the figures that don’t have one.
Run the critique in a fresh chat, with just the draft and the criteria, so the earlier conversation doesn’t colour it. And keep your opinion out of it: if you sound pleased with the draft, it may tell you what you want to hear, a habit called Jargon busterSycophancy: When an AI tells you what you want to hear, agreeing with you instead of giving an honest view..
Review the draft below against these criteria: [1. A busy reader can act on it in two minutes. 2. Every figure has a source. 3. Under 800 words. 4. No jargon.] Score each from 1 to 5. For anything under 5, quote the passage and say exactly what to change. Don’t rewrite it yet. [paste the draft]
Save it, automate it, check it
Once a chain works, save it. Keep the prompts and criteria in a document, or set up a project (ChatGPT and Claude have projects, and Gemini has Gems), so each run starts with the same instructions and files.
If the steps never change and the material arrives on its own, such as a weekly export, an Jargon busterAutomation: Getting software to do a repeated task for you, triggered by an event such as a new email arriving. tool like Zapier or Make can run the chain for you, AI steps included. If each run needs judgement about what to do next, that’s closer to an AI agent, with you approving the important moves.
Check yourself
3 quick questions nothing is savedTools in this guide
- CChatGPTOpenAI’s general-purpose assistant for writing, questions, analysis and images.
- CClaudeAnthropic’s assistant, strong at long documents, careful writing and code.
- GGeminiGoogle’s assistant, built into Gmail, Docs and Android.
- ZZapierConnects thousands of apps and automates tasks, including AI steps.
- MMakeVisual builder for complex, multi-step automations.
Sources (3)
- Building effective agentsAnthropic, December 2024
- Prompting best practicesAnthropic
- What are projects?Anthropic (Claude Help Center)
Spotted a mistake? Tell us and an editor will check it.
