Prompt chaining: breaking big tasks into steps

A single giant prompt ("research, plan, write and format a whole marketing campaign") often gives a shallow result: some parts are good, some are skipped and you can't tell which step went wrong. Prompt chaining means splitting the job into several small prompts, where the output of one becomes the input of the next.

In this tutorial you will plan a weekend promotion for the Brew & Bean coffee shop with a four-step chain, first by pasting between chat windows, then as a Python script. You need a free chat tool (ChatGPT, Gemini, Claude or Copilot). Python 3 is needed only for the last part, and the script runs without any AI account. You should have read *Be specific* and *Getting structured output*.

Chaining is not always better. For a simple task, one clear prompt is faster. Chain when the task has distinct stages (gather, decide, write, check), when the output of one stage needs your judgement, or when one prompt becomes so long that parts get ignored.

The owner wants a weekend promotion that gets more people ordering ahead on the website. The chain has four steps. Each has one job and a clear output.

Brew & Bean is a small pickup-only coffee shop in Indiranagar, Bengaluru, open 8 am to 10 pm. Menu: espresso ₹120, cappuccino ₹180, cold brew ₹220, filter coffee ₹80, masala chai ₹90, butter croissant ₹150, banana bread ₹130, veg grilled sandwich ₹160. Customers order ahead on our website.

Give me 6 ideas for a weekend promotion that gets more people to order ahead on the website. Each idea: a short name and one sentence. Ideas must be cheap for a small shop to run and must not need seating or delivery. Number them 1 to 6 and write nothing else.

What to expect: You should get six numbered one-line ideas, such as a combo (chai + banana bread) or an early-morning offer. Reject ideas that need things the shop doesn't have, such as a seating area or delivery.

This is a human checkpoint. Read the ideas and pick one. This is where chains beat one giant prompt: your judgement goes in the middle. Copy your chosen idea into the next prompt.

Here is the chosen promotion idea for Brew & Bean (pickup only, 8 am to 10 pm, orders on our website):

---
[paste the idea you chose]
---

Turn it into a plan with these headings: Offer (exact items and prices in ₹), When it runs (days and hours), How customers claim it, How we'll measure success (one number we can read from our order records), Risks. Use only items from the menu below. Keep it under 150 words.

Menu: espresso ₹120, cappuccino ₹180, cold brew ₹220, filter coffee ₹80, masala chai ₹90, butter croissant ₹150, banana bread ₹130, veg grilled sandwich ₹160.

What to expect: Check the arithmetic of any combo price yourself, and that the measure of success is something the shop can actually count (for example, orders placed between 8 am and 10 am on Saturday and Sunday).

Using only the plan below, write two things:
1. A WhatsApp message to regular customers: under 40 words, one emoji at most, ends with "Order ahead: [website link]".
2. An Instagram caption: under 60 words, with 3 hashtags.

Use only the facts in the plan. Do not invent offers, prices, times or discounts.

Plan:
---
[paste the plan from step 2]
---

What to expect: Count the words. Check every price and time against the plan.

Models review text better when this is a separate job. Asking the same prompt to write and check its own work rarely finds much; a fresh prompt with a reviewer's role and the plan to check against does better.

You are a careful editor. Compare the messages to the plan and list every problem you find: any price, time, item or offer in the messages that is not in the plan, any word limit exceeded, and anything misleading. If there are no problems, say "No problems found" and explain what you checked.

Plan:
---
[plan]
---

Messages:
---
[messages from step 3]
---

What to expect: A good reply is a short list of specific problems ("The WhatsApp message says 10 am but the plan says 9 am") or an explanation of what was checked. If it finds problems, rerun step 3 with the list added, not the whole chain.

Once a chain works by hand, you can automate it. The script below runs the same four steps. It does not need an AI account: the function ask is a stand-in that returns a fixed reply so you can see how data flows between steps. To use a real model, replace the body of ask with an API call (see *Your first LLM API call (free Gemini API)* in the AI Engineering track). Save it as chain.py and run python chain.py.

FACTS = (
    "Brew & Bean is a pickup-only coffee shop in Indiranagar, Bengaluru, "
    "open 8 am to 10 pm. Menu: espresso 120, cappuccino 180, cold brew 220, "
    "filter coffee 80, masala chai 90, croissant 150, banana bread 130, "
    "veg sandwich 160 (rupees)."
)


def ask(prompt):
    """Stand-in for a real model call. Replace this body with an API call."""
    print("--- PROMPT ---")
    print(prompt)
    print()
    return "[model reply goes here]"


def ideas():
    return ask(
        f"{FACTS}\n\nGive 6 ideas for a weekend promotion that gets more people "
        "to order ahead on the website. One short sentence each, numbered."
    )


def plan(idea):
    return ask(
        f"{FACTS}\n\nTurn this promotion idea into a plan under 150 words with "
        f"headings: Offer, When, How to claim, How to measure.\n\nIdea: {idea}"
    )


def write(plan_text):
    return ask(
        "Using only the plan below, write a WhatsApp message under 40 words "
        f"and an Instagram caption under 60 words.\n\nPlan:\n{plan_text}"
    )


def check(plan_text, messages):
    return ask(
        "List every price, time or offer in the messages that is not in the "
        f"plan.\n\nPlan:\n{plan_text}\n\nMessages:\n{messages}"
    )


idea_list = ideas()
print("Ideas:", idea_list)
chosen = input("Paste the idea you choose: ").strip()   # human checkpoint

plan_text = plan(chosen)
messages = write(plan_text)
report = check(plan_text, messages)
print("Review:", report)

Each function does one step and returns text. The output of one function is passed as an argument to the next, which is all a chain is. The input() line is the human checkpoint.

In AI engineering interviews you may be asked how to build a reliable multi-stage LLM feature, such as summarising a support ticket, classifying it and drafting a reply. A strong answer describes separate steps with clear inputs and outputs, validation between steps, logging of each step's result and a human or rule-based check before any action. Chains with fixed steps are different from agents that decide their own steps; that difference is covered in *AI agents: tools, loops and planning* in the AI Engineering track.