A single prompt can do a lot. But some tasks are too complex, too long, or require too many distinct cognitive steps for a single prompt to handle reliably. Prompt chaining solves this.
Prompt chaining is the practice of breaking a large task into a sequence of smaller, focused prompts — where the output of each prompt becomes the input for the next.
Why Single Prompts Break Down
Consider asking a model to: "Research our competitors, analyze their pricing strategies, identify gaps we can exploit, and write a competitive positioning document for our board."
In a single prompt, the model:
- Can't actually research (it has no real-time web access)
- Has to juggle four distinct cognitive modes simultaneously
- Produces mediocre output on all four because it's spread too thin
- Creates a document you can't easily review or correct mid-process
Chaining solves each of these problems.
The Core Pattern
[Input] → [Prompt 1] → [Output 1] → [Prompt 2] → [Output 2] → [Final Output]
Each step is a focused, single-purpose prompt. The output of step N becomes part of the input for step N+1.
Example: Writing a Research Report
Single prompt (bad):
Research the state of electric vehicles in 2026 and write a comprehensive
10-page report covering market size, key players, technology trends, and
regulatory landscape.
Results in a vague, superficial, hallucination-prone document.
Chained approach (good):
Step 1 — Extract structure:
You are a research analyst. Given the topic "Electric Vehicles in 2026",
create a detailed outline for a 10-page report.
Include: section titles, what each section should cover, key questions
each section should answer, and suggested data sources to look for.
Output as a numbered outline.
Step 2 — Research each section (with provided data):
Using the outline below and the research data I'm providing,
write Section 2: Market Size and Growth.
[Outline from Step 1]
[Data you've gathered from actual sources]
Write 400-500 words. Be specific, cite the data points I provided.
Do not invent statistics.
Step 3 — Review and synthesize:
Here are the five sections of our EV report:
[Sections from previous steps]
Write a 3-paragraph executive summary that:
1. States the single most important finding
2. Summarizes the key trends
3. Gives 2 specific recommendations for our business
Keep it under 250 words.
Each step produces reviewable, high-quality output that can be corrected before moving to the next.
Common Chaining Patterns
Extract → Transform → Format
Step 1: Extract all relevant data points from this document
Step 2: Analyze the extracted data and identify key insights
Step 3: Format the insights as an executive slide deck outline
Draft → Critique → Revise
Step 1: Write a first draft of [content]
Step 2: Critique the draft — what's weak, what's missing, what should change?
Step 3: Revise the draft based on the critique
Classify → Route → Respond
Step 1: Classify this customer message (billing / technical / shipping / other)
Step 2: Based on classification, select the appropriate response template
Step 3: Personalize the template for this specific customer message
Summarize → Compare → Decide
Step 1: Summarize document A in bullet points
Step 2: Summarize document B in bullet points
Step 3: Compare the two summaries and recommend which approach to take
Managing State Between Steps
The key challenge in prompt chaining is passing context cleanly between steps. Best practices:
Pass only what's needed. Don't dump the entire previous step's output into the next prompt — extract and pass only the relevant parts.
Use structured formats for handoffs. If Step 1 produces a list and Step 2 needs to process it, make sure Step 1's output is in a clean, parseable format (JSON, bullet points, numbered list).
Add context headers. When passing output between steps, label it clearly:
Here is the research outline created in the previous step:
[output]
Using this outline, now write Section 3...
Store intermediate results. If you're building a pipeline programmatically, save each step's output to a variable or file before moving on — so you can debug or restart from any point.
When to Chain vs When Not To
Use chaining when:
- The task has distinct phases (research → analyze → write)
- The output of one step needs human review before the next
- The task is too long for a single context window
- Accuracy is critical and you want to validate each step
- Different steps benefit from different model settings (temperature, model size)
Don't chain when:
- A single well-structured prompt handles it fine
- The overhead of managing multiple steps isn't worth it
- The task is fast and disposable (drafting a quick email)
Prompt Chaining in Code
If you're building with the API, chaining looks like this (Python pseudocode):
# Step 1: Extract outline
outline = llm.call(
prompt=f"Create a report outline for: {topic}",
temperature=0.3
)
# Step 2: Write each section
sections = []
for section in parse_outline(outline):
content = llm.call(
prompt=f"Write section: {section}\nData: {research_data}",
temperature=0.5
)
sections.append(content)
# Step 3: Synthesize
report = llm.call(
prompt=f"Write executive summary for:\n{join(sections)}",
temperature=0.3
)
Each step is independent, auditable, and can be retried or modified without redoing the whole task.
Key Takeaway
Prompt chaining is how you take AI from "useful toy" to "reliable tool." Complex tasks that fail in a single prompt often succeed when broken into focused, sequential steps. Design your chain so each step is simple, verifiable, and produces clean output that feeds cleanly into the next.
Next: Learn Prompt Evaluation Frameworks — how to measure and improve prompt performance scientifically.