Fix Broken JSON from ChatGPT, Claude, and LLM Output

Why LLMs Generate Broken JSON

LLMs generate text by predicting the next token, not by ensuring valid syntax. This leads to five common failure modes:

• Truncated output — the response is cut off mid-object, missing closing brackets and braces.
• Python booleans — True and False instead of true and false.
• Python None — None instead of null.
• Markdown code fences — JSON wrapped in ```json ... ``` blocks.
• Commented or mixed content — explanatory text around the JSON like "Here is the data:" before the JSON.
• Unescaped special characters — quotes inside strings without backslash escaping.

The repair tool handles all of these in a single click — just paste the raw ChatGPT output and let the tool fix it.

1. Fixing Truncated JSON

Truncation is the most common LLM JSON error. The output just stops — missing closing brackets, braces, or even the end of a string. Consider this ChatGPT output:

{
  "users": [
    {"id": 1, "name": "Alice", "role": "admin},
    {"id": 2, "name": "Bob", "role": "user"}

Notice: "admin has a missing closing quote, the array is not closed (] missing), and the object is not closed (} missing). The repair tool detects the open structures and closes them:

{
  "users": [
    {"id": 1, "name": "Alice", "role": "admin"},
    {"id": 2, "name": "Bob", "role": "user"}
  ]
}

Valid JSON in milliseconds — no manual bracket counting needed.

2. Python Booleans and None Values

LLMs trained on Python-heavy data often output Python-style values:

{
  "name": "Widget",
  "inStock": True,      // Python boolean — should be true
  "discount": None,     // Python None — should be null
  "isPremium": False    // Python boolean — should be false
}

The repair tool finds and replaces all Python-style values:

{
  "name": "Widget",
  "inStock": true,
  "discount": null,
  "isPremium": false
}

Valid JSON, ready for any JavaScript, Python, or Java parser.

3. Markdown Code Fences and Commentary

LLMs often wrap JSON in ```json code blocks with explanatory text around them. The repair tool strips everything except the JSON content:

Here is the JSON data you requested:
```json
{
  "status": "success",
  "count": 3,
  "items": ["a", "b", "c"]
}
```
Let me know if you need modifications.

The tool extracts only the JSON block, ignoring the surrounding commentary. If there are multiple code fences, it picks the one that contains valid JSON-like content.

Best Practices for LLM JSON Generation

• Ask for raw JSON only. In your prompt, add "Return only valid JSON, no markdown formatting, no explanations." This reduces but doesn't eliminate errors.
• Always repair before using. Even if the JSON looks correct, run it through the repair tool. Edge cases like trailing commas or unescaped quotes are hard to spot visually.
• Validate the result. After repair, use the JSON Editor to validate and format the output before integrating it into your application.
• Set output length limits. Truncation happens more often with large JSON outputs near the model's token limit. Request smaller chunks if possible.
• Use JSON mode when available. Some APIs (like OpenAI's JSON mode) constrain the model to return valid JSON. Enable it if your API supports it, but still repair as a safety net.

Related Tools

Frequently Asked Questions