# Output the next step can use > Ask a chat assistant for action items as named fields with allowed values, check the result with the program that has to read it, and fix the prompt until it passes. Page: https://schubergphilis.github.io/ai-training/concepts/structured-output/ License: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/ In the lesson on instruction, context, and format you asked for three bullets under 15 words, and a person read them. In this lesson the reader of the answer is a program. We ask a chat assistant for the action items in the notes of a project meeting, and we paste each answer into a small checker that accepts the same items as a task tracker's import. The first answer is prose and the checker can't read it. By the end the answer passes, and the difference is what the prompt now says about the format. You need any chat assistant, and the checker is on this page. Every response shown here is illustrative: the author wrote it to show the pattern, and the component says so. Your model gives a different text, and the checker tells you whether that text passes. ## Prose that someone retypes Here are the meeting notes. They're fictional, and every round uses them unchanged. > Project meeting, 29 September. Present: Dana, Jonas, Maria, Sam. > > Maria shows the new login page. Jonas found that the password reset email > still goes to the old address. Jonas fixes that before the test round on > Friday 3 October. Everyone agrees this is the most urgent item. > > The contract for the search box ends in December. Dana asks Sam to get > two quotes for a replacement, by 17 October. Nobody sees it as urgent > yet. > > The help texts still use the old system name. Maria takes that, target 10 > October, low priority. > > Sam notes that the accessibility check isn't booked yet. Dana books it > with the agency by 8 October. Medium priority. The team's tracker imports items from a file, and each item needs four fields: an owner, a task, a due date written as `YYYY-MM-DD`, and a priority that's `high`, `medium`, or `low`. Paste the notes into your chat tool with a one-line request above them. #### Prompt (illustrative, not a recorded transcript) ```text List the action items from the notes between the triple quotes. """ [the notes, pasted in full] """ ``` #### Response ```text Here are the action items from the meeting: 1. **Jonas** will fix the password reset email so that it goes to the new address. This needs to be done before the test round on Friday 3 October and is the most urgent item. 2. **Sam** will get two quotes for a new search box by 17 October, since the current contract ends in December. 3. **Maria** will update the help texts to use the new system name, with a target of 10 October. 4. **Dana** will book the accessibility check with the agency by 8 October. ``` This response is illustrative, like every one on this page, and yours reads a little differently. It is a good answer for a person. For the tracker it is four paragraphs that someone retypes. While retyping they decide what the priority of item 2 is, and they decide which year "17 October" is in. Each decision is a place for a mistake, and a second person retyping the same notes would decide some of them the other way. The prompt said nothing about the format, so the model chose the format it uses for a general reader, as it did in the lesson on instruction, context, and format. ## One line per item **Structured output** is an answer in a fixed format that the next step can read without a person reformatting it. The simplest such format is one line per item, with the fields named in the line. Ask for exactly that, and name the allowed values where a field has a fixed set of them. #### Prompt (illustrative, not a recorded transcript) ```text List every action item from the notes between the triple quotes. Write one line per item and nothing else, in this exact format: owner: first name | task: what they do, under 12 words | due: YYYY-MM-DD | priority: high, medium, or low Use the first names as written in the notes. The year is 2026. Take the priority from what the notes say about urgency. """ [the notes, pasted in full] """ ``` #### Response ```text owner: Jonas | task: fix the password reset email address | due: 2026-10-03 | priority: high owner: Sam | task: get two quotes for a new search box | due: 2026-10-17 | priority: low owner: Maria | task: replace the old system name in the help texts | due: 2026-10-10 | priority: low owner: Dana | task: book the accessibility check with the agency | due: 2026-10-08 | priority: medium ``` Now check the answer. The checker below accepts the same items as the tracker's import. It reads the text in its box and shows one line per item, `ok` or `FAIL`, then a summary line. It reads the one-line format from this round. The next round asks for the same fields as JSON (JavaScript Object Notation, a common text format for data that programs exchange), and the checker takes that format too. The box already holds the answer shown above, so you can try the checker before you paste your own. Press Check. The report should read: ```text ok 1: Jonas, fix the password reset email address, 2026-10-03, high ok 2: Sam, get two quotes for a new search box, 2026-10-17, low ok 3: Maria, replace the old system name in the help texts, 2026-10-10, low ok 4: Dana, book the accessibility check with the agency, 2026-10-08, medium 4 of 4 items pass ``` Now check your own answer. Replace the text in the box with the answer from your chat tool, and press Check again. The checker skips a line that starts with three backticks, so you can leave a code fence around the answer in place. If an item fails, the line tells you which field is missing or which value is outside the allowed set. That failure is about your prompt, so change the prompt, run it again in a fresh chat, and check again. ## Named keys and allowed values The one-line format is easy to read and easy to break: a task that contains a `|` splits into an extra field, and there is no agreed way to write a value that's empty or a list. When the next step is a program, ask for the format the program reads. For many trackers and scripts that's JSON. The request has the same parts as before: the type, the fields, and the allowed values of each field. #### Prompt (illustrative, not a recorded transcript) ```text List every action item from the notes between the triple quotes. Return a JSON array with one object per item. Each object has exactly these four keys, and no others: - "owner": the first name as written in the notes - "task": what they do, under 12 words - "due": the date as "YYYY-MM-DD", in the year 2026 - "priority": one of "high", "medium", or "low", taken from what the notes say about urgency Return only the JSON, with no text before or after it. """ [the notes, pasted in full] """ ``` #### Response ````text ```text [ {"owner": "Jonas", "task": "fix the password reset email address", "due": "2026-10-03", "priority": "high"}, {"owner": "Sam", "task": "get two quotes for a new search box", "due": "2026-10-17", "priority": "low"}, {"owner": "Maria", "task": "replace the old system name in the help texts", "due": "2026-10-10", "priority": "low"}, {"owner": "Dana", "task": "book the accessibility check with the agency", "due": "2026-10-08", "priority": "medium"} ] ``` ```` The checker reads an answer that starts with `[` as JSON and applies the same four checks: every key present, no extra keys, a date the import can parse, and a priority from the allowed set. The box below contains the same answer, laid out one key per line. Press Check, and compare the report with this one: ```text ok 1: Jonas, fix the password reset email address, 2026-10-03, high ok 2: Sam, get two quotes for a new search box, 2026-10-17, low ok 3: Maria, replace the old system name in the help texts, 2026-10-10, low ok 4: Dana, book the accessibility check with the agency, 2026-10-08, medium 4 of 4 items pass ``` Paste your own JSON answer into the box and check it too. Both rounds show the same report, because the checker looks at the fields and the values, whatever the packaging. Naming the fields is what made the two answers comparable, and it is also what makes a missing value visible: an item without a due date fails a check instead of disappearing into a paragraph. Vendors' developer tools add ways to force a format on the output, and the public course on the Claude API describes one of them (Building with the Claude API, Claude Academy). The rule in this lesson is older than those features and works in any chat tool: say what the format is, in enough detail that a program can check it (ChatGPT Prompt Engineering for Developers, DeepLearning.AI). ## Check the first result A usable format request names the output type (a JSON array, or one line per item), the fields of each item, and the allowed values of every field that has a fixed set of them, including how dates and numbers are written. Leave one out and the model fills the gap with its own default, and that default can change from one run to the next. #### Pitfall: Return JSON A learner types "Return the action items as JSON" and gets an array of four objects with `assignee`, `action`, `deadline` and `priority` keys, where the deadline is "Friday 3 October" and the priority is "urgent". The import rejects every item. They run the same prompt again and get a JSON object with an `action_items` key holding objects with `who`, `what` and `when` keys. Both answers are valid JSON, and neither matches the other or the tracker. Rule: name the keys and the allowed values, because "return JSON" fixes the packaging and leaves the contents to the model. The box below holds the first of those two answers, with the code fence the chat tool put around it. Check it to see what the checker reports when the type is right and the fields are not. Press Check. The report should read: ```text FAIL 1: missing owner, task, due (found assignee, action, deadline) FAIL 1: priority "urgent" is not one of high, low, medium FAIL 2: missing owner, task, due (found assignee, action, deadline) FAIL 2: priority "normal" is not one of high, low, medium FAIL 3: missing owner, task, due (found assignee, action, deadline) FAIL 3: priority "normal" is not one of high, low, medium FAIL 4: missing owner, task, due (found assignee, action, deadline) 0 of 4 items pass ``` The last part of the rule is about order of work. Check the first result against the program that has to read it before you produce the rest. A format that looks right on the screen can still fail on the one detail the consumer is strict about, such as a date format or an empty field, and finding that on item 1 costs one prompt change. Finding it after 200 items costs 200 corrections. The checker in this lesson is a stand-in for that consumer. At work the consumer is the import button or the script that reads the file, and you use that instead. #### Checkpoint: Which request can the tracker import? A colleague shows you three requests for the same meeting notes. Which one gives the tracker a file it can import without anyone editing it? - "Put the action items from these notes in a table with owner, task, due date and priority, so I can import them into the tracker." - "Return the action items from these notes as JSON, one object per item with the usual fields. I will import the file into our tracker." - "One line per item, fields separated by tabs, in this order: owner (first name), task (under 12 words), due (YYYY-MM-DD, year 2026), priority (high, medium or low)." #### Checkpoint: The first result looks right The first result looks right. What do you do next? - Run the other 39 sets of notes with the same prompt, then import every result at once. - Import the first result into the tracker now, and change the prompt if the import rejects anything. - Ask the assistant to check its own answer for mistakes, then run the other 39 sets. ## Exercise Take the sample notes from this lesson, and write your own format request from scratch without looking at the ones on this page. Ask for the four fields with their allowed values, in the one-line format or as JSON. Run the prompt in your chat tool, paste the answer into the checker below and press Check. If any item fails, change one thing in the prompt, run it again in a fresh chat, and check again. Stop when the report says `4 of 4 items pass`. Fifteen minutes is enough. Doing this once against a checker you can't argue with is what makes you name the fields the next time, before the first run. A good result is a passing report and a prompt that names the type, the four fields and the allowed values of `due` and `priority`. Which part of your first request did the checker catch, and would a person retyping the answer have caught it? Stretch: Add a fifth field to your request, such as a status that is one of open, blocked or done, and change the prompt until every item has a status with one of those values. The checker knows only the four fields, so it reports the status as unexpected on each item. The tracker's import would reject the fifth field too, so at work you would add the field to the import first. Here, check that no other failure is left, and read the status values yourself. ## Recap 1. Prose is for a person to read. When the next step is a program, ask for the format that program reads, and the retyping and the decisions made while retyping disappear. 2. A usable format request names the output type, the fields of each item, and the allowed values of every field that has a fixed set of them, including how dates and numbers are written. 3. "Return JSON" fixes the packaging and leaves the keys and values to the model, which may choose differently on each run. Named keys make runs comparable and make a missing value visible. 4. Check the first result against its consumer, the import or the script that reads the file, before you produce the rest. Fixing the prompt after item 1 costs one change. ## References - Building with the Claude API, Claude Academy, https://academy.claude.com/courses/building-with-the-claude-api - ChatGPT Prompt Engineering for Developers, DeepLearning.AI, https://www.deeplearning.ai/courses/chatgpt-prompt-eng