Technique for pretraining ChatGPT on documentation by chat share made with custom instruction

Overview: An AI’s pretraining might give it an idea of how to answer, but loading documentation into the context or chat history can make it perform even better at a task, warming it up to the subject.

Previously: Just ask the AI for documentation on the topic, and a few more questions. The AI now has more subject knowledge present in its immediate attention.

New Technique:

  1. (Generate documentation with AI or clean up yours in text editor)
  2. Place custom instruction to respond with only “documentation received”. (using new ChatGPT feature)
  3. Paste your documentation to ChatGPT (in chat-sized pieces to minimize the chance of backend summarization).
  4. Share the chat, and then you can bookmark it, and continue a conversation with different custom instructions.

Justification: ChatGPTs backend chat management seems to prefer recalling user inputs when choosing past conversation that is relevant to understanding. Any unnecessary prompting or responses is omitted by this technique.

Custom instruction prompt:

ChatGPT is now placed into a special internal mode for receiving documentation. Do not produce a response for a human. The only output is “I understand this documentation” with no other response.

Example: Chat Share Link created with technique, primed to answer questions on a specific subject.

Happy prompting!

1 Like

Update: OpenAI has gone out of their way to break the technique described above.

A chat share that has been loaded with context and style, using techniques such as custom instructions, can no longer be continued by another user or even yourself.

That usefulness had to be destroyed.

Now, instead of “continue conversation”, OpenAI put in an ad.

image