JSON Prompt Extractor

A JSON prompt extractor for chat exports, API files and AI images

This JSON prompt extractor does two jobs. Extract prompts reads the files that hold your AI conversations and requests — a ChatGPT or Claude data export, an OpenAI, Anthropic or Gemini request body, a Batch API JSONL file, or a PNG made by Stable Diffusion or ComfyUI — and lists every system, user and assistant message in one searchable table you can save as CSV, JSONL or Markdown.

Extract JSON from AI output finds the JSON inside a model’s answer, even when it is wrapped in chatter, fenced as code, or cut off half-way, repairs the usual mistakes, tells you exactly what it fixed, and can check the result against a JSON Schema.

How to extract prompts or JSON

1. Choose Extract prompts or Extract JSON from AI output.
2. For prompts, open a file (an export .zip, conversations.json, .json, .jsonl or .png) or paste JSON; the detected format is shown above the table.
3. Search the text or filter by role, then export the rows as CSV, JSONL or Markdown.
4. For JSON, paste the AI’s answer; pick a block if several are found and read the list of repairs.
5. Open Check against a JSON Schema to validate the result, then copy or download it.

JSON Prompt Extractor features

Formats it understands

Export formats are not formally documented and change from time to time; this tool follows the field names found in current exports. API shapes follow the OpenAI, Anthropic and Google API references, checked 14 October 2026. Any other JSON still gives up strings stored under keys such as prompt, system, instructions, content or text.

Worked example: rescuing JSON from an AI answer

The example answer has a comment, single and curly quotes, an unquoted key, a trailing comma, Python’s True, and it stops in the middle of the second order. This tool reports:

• Removed 1 comment
• Changed 4 single-quoted strings to double quotes
• Straightened curly quotes around 1 string
• Quoted 1 unquoted key
• Removed 1 trailing comma
• Changed 1 Python True/False/None to JSON
• Dropped 1 incomplete key or value at the cut-off point
• Closed 3 brackets left open at the end (the output was cut off)

and returns this valid JSON:

{
  "name": "Asha Rao",
  "email": "asha@example.com",
  "tags": [
    "vip",
    "beta"
  ],
  "active": true,
  "orders": [
    {
      "id": 1,
      "total": 25.5
    },
    {
      "id": 2
    }
  ]
}

The incomplete "total": is dropped rather than guessed — and the example schema then reports exactly that: $.orders[1] is missing the required “total”. Valid syntax is not the same as complete data, which is why the schema check is worth running. To format larger JSON, use the JSON formatter; to turn rows into a spreadsheet, the JSON to CSV converter.

Where your prompts are inside exports and images

In ChatGPT you request an export under Settings → Data controls; the e-mailed ZIP contains conversations.json. Claude’s export, from Settings → Privacy, also arrives as a ZIP with conversations.json. Open either ZIP here directly.

Stable Diffusion front-ends store the prompt in the PNG file itself, which is why sharing an image can share your prompt. Read it here, and remove it before posting with the EXIF remover. JPEG and WebP images keep prompts in EXIF instead — try the EXIF viewer for those.