Getting structured output: tables, JSON and checklists

By default an AI replies in flowing paragraphs. That is fine for a chat, but not when you want to compare options, paste results into a spreadsheet or read the answer in a program. For that you ask for a structure: a table, a checklist or JSON.

In this tutorial you will turn the same Brew & Bean order message into a table, a checklist and a JSON object, then read the JSON with a short Python script. You need a free chat tool (ChatGPT, Gemini, Claude or Copilot). Python 3 is needed only for the last part. You should have read *Anatomy of a good prompt*: format is one of its six parts.

Tables work best when you have several items and the same few questions about each. Name the columns yourself, in order. If you don't, the model picks columns for you, and they change from run to run.

Compare Python, Java and JavaScript for a first-year B.Tech student who wants to get a campus placement.

Format: a table with exactly these columns: Language | Easy to learn (1-5) | Used for | Placement relevance (one sentence).
One row per language. After the table, add one line recommending a starting language and why. No other text.

What to expect: You should get a clean table with three rows and four columns, then one recommendation line. Scores are opinions, so treat them as a starting point, not data.

Most chat tools show a table as a grid. If you want to paste it into a spreadsheet, add: "Give the table in CSV format inside a code block." CSV is plain text with commas between columns, which Excel and Google Sheets can import.

Brew & Bean is a pickup-only coffee shop open 8 am to 10 pm. Write the closing checklist for the person who closes the shop at 10 pm.

Format: a checklist of 10-12 items. Each item starts with a verb and can be ticked off in under 5 minutes. Group the items under three headings: Counter, Kitchen, Cash and orders. Do not add explanations.

What to expect: Check that every item is a single action you can tick off ("Wipe the espresso machine steam wand"), not a vague goal ("Make sure everything is clean").

JSON is a text format that programs can read. An object is written in curly braces with "key": value pairs. Text values are in double quotes, numbers are not, and lists are in square brackets. Almost every language can parse it.

Below is a real message a customer might send to the shop. Our goal is to turn it into an order object matching the Brew & Bean order shape: customer_name, items (each with item_id and qty) and pickup_time.

Extract an order from the customer message below.

Return ONLY a JSON object, with no explanation and no code fences, in exactly this shape:
{"customer_name": string or null, "items": [{"item_id": string, "qty": integer}], "pickup_time": string or null}

Rules:
- item_id must be one of: espresso, cappuccino, cold-brew, filter-coffee, masala-chai, croissant, banana-bread, veg-sandwich.
- If a value is not in the message, use null. Do not guess.
- pickup_time stays as the customer wrote it.

Message:
---
Hi, this is Divya. 2 cappuccinos and one banana bread please, I'll pick up around 6:30 in the evening.
---

What to expect: Expected: {"customer_name": "Divya", "items": [{"item_id": "cappuccino", "qty": 2}, {"item_id": "banana-bread", "qty": 1}], "pickup_time": "around 6:30 in the evening"}. Notice there is no phone number in the message, and the shape had no phone field, so nothing is invented.

Copy the model's JSON reply into a file called reply.json (just the JSON). Then save the script below as read_order.py in the same folder and run python read_order.py. It parses the text and checks it, which is what any program using an AI reply should do. Models sometimes wrap JSON in code fences even when told not to, so the script strips them.

import json

VALID_ITEMS = {
    "espresso", "cappuccino", "cold-brew", "filter-coffee",
    "masala-chai", "croissant", "banana-bread", "veg-sandwich",
}


def clean(text):
    """Remove ``` fences if the model added them."""
    text = text.strip()
    if text.startswith("```"):
        lines = text.splitlines()[1:]
        if lines and lines[-1].strip().startswith("```"):
            lines = lines[:-1]
        text = "\n".join(lines)
    return text.strip()


def parse_order(text):
    order = json.loads(clean(text))
    if not isinstance(order.get("items"), list) or not order["items"]:
        raise ValueError("items must be a non-empty list")
    for item in order["items"]:
        if item.get("item_id") not in VALID_ITEMS:
            raise ValueError(f"unknown item_id: {item.get('item_id')}")
        qty = item.get("qty")
        if not isinstance(qty, int) or qty < 1:
            raise ValueError(f"bad qty for {item['item_id']}: {qty}")
    return order


with open("reply.json", encoding="utf-8") as f:
    reply = f.read()

try:
    order = parse_order(reply)
    print("Order OK for", order.get("customer_name") or "unknown customer")
    for item in order["items"]:
        print(f"  {item['qty']} x {item['item_id']}")
    print("Pickup:", order.get("pickup_time") or "not given")
except (json.JSONDecodeError, ValueError) as err:
    print("Problem with the reply:", err)

json.loads turns the text into Python dicts and lists. The checks after it reject anything that is not a real menu item or a positive whole number.

If the script prints Problem with the reply, you have found the reason programs need checks. Paste the error and the reply back to the model and ask it to fix the JSON (next prompt). APIs from some providers can also enforce a JSON schema for you; that is covered later in *Structured output with JSON schemas* in the AI Engineering track.

The JSON below failed to parse with this error: [paste error message].

Fix it so it is valid JSON. Do not change any values. Return only the corrected JSON.

---
[paste the JSON]
---

What to expect: Typical causes: a trailing comma, single quotes instead of double quotes, or text before or after the object. The fixed version should differ only in punctuation.