Prompt Chaining for Writing – A Practical Workflow
Prompt chaining for writing means splitting a writing job into a fixed sequence of smaller prompts (for example brief, outline, draft, critique and revision) where the output of each step becomes the input of the next. It works because each call has one clear task you can inspect before moving on, and Anthropic describes the trade plainly: you accept more latency in exchange for higher accuracy. Use it for long or high-stakes pieces; for a short caption or a single email, one well-written prompt is usually enough.
This guide explains where the technique comes from, what the model vendors recommend, a chain you can use for articles, and how to tell whether the extra steps are paying off. The method is tool-neutral and works in any chat model or API.
What prompt chaining is
Anthropic's engineering article Building effective agents, credited to Erik S. and Barry Zhang and dating from December 2024, defines prompt chaining as a workflow that "decomposes a task into a sequence of steps, where each LLM call processes the output of the previous one". It adds that you can insert programmatic checks, which it calls gates, between steps "to ensure that the process is still on track". Two of its examples are writing tasks: generating marketing copy and then translating it, and writing an outline, checking it against criteria, then writing the document from the outline.
Google's Gemini prompting strategies describe the same pattern among its ways to break down complex prompts: "make each step a prompt and chain the prompts together in a sequence", with the output of the last prompt being the final result. The same page lists two relatives worth knowing. Breaking down instructions means one prompt per instruction, chosen according to the input. Aggregating responses means running different tasks on different parts of the material in parallel and combining the results.
Chaining is distinct from chain-of-thought prompting. The chain-of-thought paper by Jason Wei and colleagues showed that prompting a model with examples of intermediate reasoning steps improved performance on arithmetic, commonsense and symbolic reasoning tasks. That happens inside a single prompt and response. Prompt chaining happens across several calls, and you can read, edit or reject each intermediate output yourself.
Why one big prompt struggles
A single prompt that asks for research, structure, tone, SEO, fact-checking and a final draft at once asks the model to balance every instruction in one pass. When the result is weak, you cannot tell which instruction failed.
The research case for chaining began with exactly that problem. In AI Chains, Tongshuang Wu, Michael Terry and Carrie J. Cai argued that the breadth, lack of transparency and limited controllability of large models make them less effective on complex tasks. In their 20-person user study, chaining "not only improved the quality of task outcomes, but also significantly enhanced system transparency, controllability, and sense of collaboration". Participants also debugged unexpected outputs by testing individual parts of a chain, much as a programmer tests a function.
Anthropic's current prompting best practices add a caveat for newer models: with built-in thinking, models handle most multistep reasoning internally, so explicit chaining is "still useful when you need to inspect intermediate outputs or enforce a specific pipeline structure". For editorial work, inspection is usually the point. An editor wants to approve the outline before anyone drafts from it.
A chain for articles and long posts
This six-step chain covers most blog posts, guides and reports. Each step is a separate prompt, and each output is saved.
- Brief. Turn your notes into a structured brief: reader, search query, angle, evidence you hold, constraints. Ask the model to list anything missing rather than guess.
- Questions. "List the questions this reader is likely to have, in the order they would ask them." This surfaces scope before structure.
- Outline. Feed in the brief and approved questions. Ask for headings that each make one claim, with a note of the evidence each section needs.
- Draft, section by section. Give the model the approved outline, the brief and the section to write. Drafting one section per call keeps each request small and lets you stop when a section goes wrong.
- Critique. In a fresh prompt, paste the draft and a checklist: accuracy against sources, one idea per section, reading level, banned phrases, spelling convention. Ask for problems only, not a rewrite.
- Revise. Paste the draft and the critique; ask for a revision that addresses each point and changes nothing else.
Steps five and six are what Anthropic calls the most common chaining pattern, self-correction: generate a draft, review it against criteria, refine based on the review. The Self-Refine paper by Aman Madaan and colleagues tested a similar loop, where one model generates, critiques and refines its own output. Across seven tasks, outputs from Self-Refine were preferred by humans and automatic metrics over single-step generation, with an average improvement of about 20 percent absolute in task performance.
Gates between steps
A gate is a check that must pass before the next prompt runs. In an automated pipeline it can be code; in an editorial workflow it is usually a person with a short checklist. Useful gates for writing:
- After the brief: does it name a specific reader and the exact query? If not, stop.
- After the outline: does every heading answer part of the query, and is there an evidence note for each? Remove sections that fail.
- After each section: are all figures, names and quotes traceable to the sources in the brief? Anything else is deleted or sourced.
- After revision: has anything changed that the critique did not ask for?
Gates matter most for facts. Google's guidance on generative AI content warns that models "predict a likely sequence of words based on their training data", that outputs may contain hallucinations, and that it is "critical to manually factcheck and review all AI-generated content" before publishing, including titles, meta descriptions and image alt text.
Prompts that reduce errors inside the chain
Anthropic's page on reducing hallucinations lists techniques that slot neatly into a chain:
- Allow "I don't know". Explicitly permit the model to admit uncertainty; the page says this can drastically reduce false information.
- Extract quotes first. For long source documents (the page suggests over 20,000 tokens), ask for word-for-word quotes in one step, then write from those quotes in the next.
- Verify with citations. Ask the model to support each claim with a quote; if it cannot find one, it must retract the claim.
- Restrict to supplied material. Instruct the model to use only the documents provided, not its general knowledge.
- Iterative refinement. Use outputs as inputs for follow-up prompts that check or expand previous statements.
The quote-extraction step has a public worked example. The Prompt Engineering Guide's chaining page shows a two-prompt chain for document question answering: the first prompt extracts relevant quotes into tagged output, and the second composes an answer from those quotes plus the original document.
Format the hand-offs so each step can read the last one cleanly. OpenAI's prompt engineering guide recommends Markdown and XML tags to show where one piece of content begins and ends, and Anthropic recommends wrapping documents and inputs in their own tags. Ending step three with an outline inside <outline> tags makes it easy to paste into step four.
Where to chain and where not to
Anthropic's agents article frames chaining as ideal "for situations where the task can be easily and cleanly decomposed into fixed subtasks". Writing tasks that fit:
- long guides and reports with several sections
- content that must match a source document closely, such as summaries of research or policy
- translation or localisation after drafting
- pieces where an editor must sign off a plan before drafting begins
Tasks that rarely need it: short social posts, quick replies, a single product line, or a headline brainstorm. The same article warns that agentic systems often trade latency and cost for better performance, and that for many applications optimising a single call "with retrieval and in-context examples is usually enough". Every extra step costs time and, on paid APIs, tokens.
The same page describes a related workflow, the evaluator-optimizer, where one call generates and another evaluates in a loop. It says this fits best when there are clear evaluation criteria and when "LLM responses can be demonstrably improved when a human articulates their feedback", likening it to the iterative process a human writer goes through. For writing, that is a sign your critique step needs explicit criteria, not "make it better".
Common chain mistakes
- Passing too little forward. Each prompt starts without memory of your intentions unless you include the brief. Carry the brief and constraints through every step.
- Passing everything forward. Pasting all earlier outputs into every prompt buries the current task. Pass what the step needs.
- No gate. A chain without checks simply automates errors faster. An invented statistic in step four becomes "fact" by step six.
- Vague critique prompts. "Improve this" returns a rewrite. "List places where a claim lacks a source" returns something you can act on.
- Too many steps. If a step produces nothing you would review or change, merge it with its neighbour.
Measuring whether the chain helps
Compare the chain against your old single-prompt method on the same three or four topics. Track what matters to your team: editing time per piece, number of factual corrections at review, and how often the outline had to be rebuilt. OpenAI's guide recommends building tests and evaluation suites because prompt behaviour can change between models and model snapshots; a fixed set of test topics does the same job for an editorial team.
Keep the chain documented: the prompt for each step, the gate checklist and the model used. When output quality drops, change one step at a time so you can see which change made the difference. Our sister site AI Disraeli UK covers related AI writing workflows for British audiences.