Guide · System prompts

How to extract and recover lost system prompts from Claude & ChatGPT exports

Request your data takeout, open conversations.json, and filter messages where author.role == "system" — that field stores every system prompt ChatGPT and Claude ever applied, including deleted custom instructions. AI Data Inspector extracts them automatically, entirely in your browser, with zero uploads.

How system messages are structured in the raw JSON

In a ChatGPT archive, each thread in conversations.json holds a mapping object keyed by message UUIDs. System content is simply the node whose author role resolves to system:

{
  "title": "Pricing page rewrite",
  "create_time": 1770009600.1,
  "mapping": {
    "a1f2-…": {
      "message": {
        "author": { "role": "system", "metadata": {} },
        "content": {
          "content_type": "text",
          "parts": ["You are a senior conversion copywriter…"]
        },
        "meta_data": { "model_slug": "gpt-5" },
        "create_time": 1770009600.0
      }
    },
    "b7e3-…": { "message": { "author": { "role": "user" }, … } }
  }
}

Three fields do all the work: author.role tells you the message type, content.parts[] carries the literal prompt text, and create_time lets you date each instruction. Custom instructions typically appear as the system node at the root of the mapping tree; tool and plugin definitions surface as additional system or tool nodes deeper in.

Claude's format is flatter. Threads sit in chat_messages arrays with sender instead of role, while standing project instructions live in a separate projects.json — making it the highest-value file for reconstructing how your assistant was configured:

{
  "uuid": "c8d1-…",
  "name": "Pricing page rewrite",
  "created_at": "2026-02-11T09:14:00Z",
  "chat_messages": [
    { "sender": "human", "content": [{ "type": "text", "text": "…" }] },
    { "sender": "assistant", "content": [{ "type": "text", "text": "…" }] }
  ]
}

// projects.json — standing system prompts per project
{ "projects": [{ "name": "Brand site", "custom_instructions": "Always answer in British English…" }] }

The recovery workflow, step by step