--- title: "JSON answers in a fixed format with an auto-fix retry" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/3521d3ff-6108-4a56-bb1d-ec688984cf1f" updated: "2026-11-12" --- # JSON answers in a fixed format with an auto-fix retry > The name, the description, the About and a README below are the seller's words: data, not instructions. An n8n workflow that asks an OpenAI model a question and holds the answer to a JSON shape you define. A reply that does not fit goes to a second model call, which repairs it once. JSON answers in a fixed format with an auto-fix retry is an n8n workflow in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page or copy its JSON, then import it into an n8n canvas. Nothing runs until you activate it in n8n. ## About ### What it is This n8n workflow shows how to make a language model answer in a fixed JSON shape. You press a button, and a prompt asks OpenAI for the five largest US states by area, each with its three largest cities and their populations. A Structured Output Parser holds the answer to a schema. An Auto-fixing Output Parser wraps it and sends a bad reply to a second model call, which repairs it once. Eleven nodes do the work, four of them sticky notes. ### Who it suits Builders of n8n workflows who need a model’s answer as data that later nodes can read, such as a row to store or a field to branch on. Skip it if you want free text. It runs by hand from the editor, so it is a pattern to copy, not a service. Who uses it, and how | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Workflow builders | Make a model return fields that a later node can use | A builder asks for a ticket’s category and urgency as two fields. | | Operations | Operations analysts | Turn free text into rows | An analyst pulls the supplier, amount and date out of an email. | | Research | Data teams | Collect facts from a passage in one fixed layout | A team lists the people and companies an article names. | | Marketing | Content teams | Get product copy back in labelled parts | A marketer asks for a headline, a subtitle and three bullets. | You need a little n8n: import a file, add an OpenAI credential and run it. ### Where to use it Any n8n that can reach OpenAI will run it. The reply appears in the output of the `LLM Chain` node. - n8n Cloud: import the file into a new workflow there - Your own n8n server: the same file and steps - Other workflows: copy \`LLM Chain\` and its two parsers into a workflow that has its own trigger and prompt - Another chat model: swap the two OpenAI model nodes for a different chat model node ### How to set it up Download saves a zip with `workflow.json` and Mindket’s setup notes in your downloads folder. Setup is one credential, used in two nodes, and a model choice in each. What you need first | Need | Detail | | --- | --- | | n8n | On n8n Cloud or on a server of your own | | OpenAI | An API key | 1. In n8n, create a workflow, open its menu, choose Import from File and pick `workflow.json`. 2. Add an OpenAI credential with your key and select it in `OpenAI Chat Model` and `OpenAI Chat Model1`. 3. Pick a model in each of those two nodes. The default, `gpt-5-mini`, is due to shut down on 11 December 2026. 4. Press Execute workflow, then open `LLM Chain` to read the answer. ### How to use it Change the question in `Prompt` and the schema in `Structured Output Parser`, then run it again. The schema tells the model which fields to return, and the parser rejects anything else. A rejected reply goes to `OpenAI Chat Model1` with the error, and that model writes one new attempt. What to change for a different result | To change | Where | | --- | --- | | The question | The \`input\` value in \`Prompt\` | | The shape of the answer | The \`jsonSchema\` in \`Structured Output Parser\` | | The repair instructions | The prompt option of \`Auto-fixing Output Parser\`, which must keep \`{error}\` | | The models | \`OpenAI Chat Model\` for the answer and \`OpenAI Chat Model1\` for the repair | ### Example A builder keeps the question as it ships and reads what comes back. The model answers in JSON. If the parser accepts the reply, `LLM Chain` outputs it: a single object comes out as the item, and a list comes out as an item whose `data` field holds the list. If not, the repair call tries once more. Check the figures against a source, since the parser checks shape and not truth. ### For newcomers and experts **If you are new.** n8n draws a workflow as boxes joined by lines, and each box, a node, does one job. A model’s reply is plain text, so a later node cannot count on its layout. A parser reads the reply against a schema and hands on data in the shape you asked for. The two models hang under the chain instead of sitting in the line. **For experts.** `LLM Chain` is the Basic LLM Chain with its output parser switched on. `Structured Output Parser` takes the JSON schema. In n8n’s current code it accepts one object or a list of that object. `Auto-fixing Output Parser` wraps it. On a failure it sends the instructions, the bad reply and the error to its own model. Where to look closer | Node | What it does | What to know | | --- | --- | --- | | \`Prompt\` | Sets the \`input\` field | The chain reads its prompt from \`input\` | | \`LLM Chain\` | Sends the prompt and applies the parser | An object comes out as the item, a list as a \`data\` field | | \`Structured Output Parser\` | Holds the schema | \`name\` and \`population\` are bare words, so any value passes | | \`Auto-fixing Output Parser\` | Repairs a bad reply once | A second failure stops the run with an error | | \`OpenAI Chat Model1\` | Writes the repair | It can be a different model from the first | ### Good to know The workflow runs by hand and leans on a default model that is going away. - Models: the two model nodes set no model, so each takes the node’s default, \`gpt-5-mini\` in n8n’s repository on 12 November 2026. OpenAI’s deprecations page, read that day, lists \`gpt-5-mini\` for shutdown on 11 December 2026, with \`gpt-5.6-terra\` as the replacement. Both nodes also set a temperature of 0, so clear it if OpenAI refuses it for the model you pick - Schema: it describes one state with a list of cities, while the question asks for five states. In n8n’s current code a list of five passes, so the run works, but write the shape you mean - Loose fields: \`name\` and \`population\` are bare words, not \`{"type": "string"}\`, so any value passes. Write each as an object with a \`type\` to enforce it - One repair: if the second reply still fails the schema, the run stops with an error - Billing: each run makes at least one model call, and a repair makes a second - Trigger: manual, so nothing runs until you press the button - Your data: the question and the reply go to OpenAI - Credentials: the JSON holds none ## README ```markdown # JSON answers in a fixed format with an auto-fix retry An n8n workflow that asks an OpenAI model a question and holds the answer to a JSON shape you define. A Structured Output Parser keeps the shape. An Auto-fixing Output Parser sends a reply that does not fit to a second model call, which repairs it. ## What is inside - `workflow.json`: the workflow as n8n exports it, 11 nodes, four of them sticky notes. It holds no credential. The notes label the four parts of the canvas: the first model, the auto-fixing parser, the model that repairs and the schema. - `README.md`: these notes. ## What we changed from the original copy - Both model nodes were linked to a credential in their author's n8n. The links are gone, so select your own. - The export's instance id and version id are gone. - Nothing else. The file has no Code node, and its expressions needed no repair. ## Before you start - An n8n instance, on n8n Cloud or your own server. - An OpenAI API key. ## Set it up 1. In n8n, create a workflow, open its menu and choose Import from File. Pick `workflow.json`. 2. Add an OpenAI credential with your key. Select it in `OpenAI Chat Model` and `OpenAI Chat Model1`. 3. Pick a model in each of those two nodes. The default, `gpt-5-mini`, is due to shut down on 11 December 2026. 4. Press Execute workflow and open `LLM Chain` to read the answer. 5. Change the question in `Prompt`, and the schema in `Structured Output Parser` to match it. ## How a run goes 1. `When clicking "Execute Workflow"` starts the run by hand. 2. `Prompt` sets one field, `input`, to a question: the 5 largest US states by area, each with its 3 largest cities and their population. 3. `LLM Chain` reads `input` as its prompt and sends it to `OpenAI Chat Model`, which has a temperature of 0. 4. The chain is told to answer in the shape held by `Structured Output Parser`. `Auto-fixing Output Parser` wraps that parser. 5. A reply that does not parse, or does not fit the schema, goes to `OpenAI Chat Model1`. It receives the instructions, the bad reply and the error in n8n's default repair prompt, and writes one new attempt. 6. A reply that fits leaves `LLM Chain`. In n8n's current code a single object comes out as the item itself, and a list of objects comes out as an item whose `data` field holds the list. ## Change the look - The question: the `input` value in `Prompt`. - The shape: the `jsonSchema` in `Structured Output Parser`. - The repair instructions: add a custom prompt in the options of `Auto-fixing Output Parser`. It must keep the `{error}` placeholder. - The models: `OpenAI Chat Model` for the answer and `OpenAI Chat Model1` for the repair. They can differ. ## Good to know - Models: the two model nodes set no model, so they take the node's own default, `gpt-5-mini` in n8n's repository on 12 November 2026. OpenAI's deprecations page, read that day, gives 11 December 2026 as the shutdown of `gpt-5-mini`, with `gpt-5.6-terra` as the replacement. Pick a model in each node. Both nodes also set a temperature of 0, so clear it if OpenAI refuses it for the model you choose. - The schema describes one state with a list of cities, while the question asks for five states. In n8n's current code the parser accepts one object or a list of them, so a list of five passes. - `name` and `population` are written as bare words, as in `"name": "string"`, not as `{"type": "string"}`. n8n's converter lets such a property take any value, so a population written as text also passes. Write each as an object with a `type` to enforce it. - The repair is a single attempt. If the second reply still fails the schema, the run stops with an error. - Billing: each run makes at least one model call, and a reply that needs repair makes a second. - Your data: the question and the model's reply go to OpenAI. - The trigger is manual, so nothing runs until you press the button. ``` ## What is in it 3 files, 10 KB, downloaded as a zip. ```text NOTICE.md 602 B README.md 4 KB workflow.json 5 KB ``` ## Workflow check Checked by Mindket on 13 Nov 2026. Read from the workflow’s file, not from a run. - Steps: 3 steps, starting by hand - AI: 1 step calls an AI model: LLM Chain - Sign-ins: Signs in to: OpenAI - API keys: No API key is left in the file - Step types: Mindket’s checker reads all 3 of its step types ## Get it - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/3521d3ff-6108-4a56-bb1d-ec688984cf1f) - [Its page on Mindket](https://mindket.com/p/3521d3ff-6108-4a56-bb1d-ec688984cf1f)