718: ChatGPT Custom Instructions: A Major, Easy Hack for Data Scientists

Episode Summary

Executive Summary: John Crohn shares a practical ChatGPT custom-instructions hack from Sean Khosla: instruct the model to first draft an outline, then critique it, then produce the best final answer. The approach is meant to improve response quality via self-critique, while also enforcing a clear, professional, direct tone suited to technical users.

Main Topics: Custom instructions as a quality-control tool: The episode explains how ChatGPT's custom instructions can be used to systematically improve outputs without manually prompting each time. The self-critique prompt structure: Sean Khosla's prompt asks the model to outline its response, critique that outline, and then synthesize the best final solution, creating an internal revision loop. Tone optimization for technical users: The second part of the instruction tells ChatGPT to avoid overly bubbly or salesy language and instead respond clearly, professionally, and directly. Audience fit for data/AI professionals: Crohn emphasizes that listeners are likely data scientists, ML engineers, AI researchers, and software engineers who prefer concise, useful answers over model caveats. Practical setup steps in ChatGPT: The episode briefly walks through where to find custom instructions in ChatGPT and notes that the hack uses only the 'how would you like ChatGPT to respond' field. Sharing and iterating on prompt ideas: Listeners are encouraged to copy the prompt from the show notes, adapt it, and share variations publicly for communal improvement.

Key Arguments: A structured self-critique prompt can improve output quality by forcing the model to revise its first inclination before responding. Custom instructions are a low-friction way to apply this improvement across many conversations automatically. Technical users often do not need generic persona padding or repeated model disclaimers; directness is more valuable. A clear, professional, and direct tone better matches the needs of data science and engineering audiences. The hack is portable: users can copy the prompt, tailor it, and potentially apply it to GPT-4 and possibly GPT-3.5.

Data Points: Podcast format: 5 Minute Friday - The episode is presented as a short, tactical segment. Model references: GPT-4 - Crohn says the hack improves responses from OpenAI's GPT-4 model. Possible additional model: GPT-3.5 - He notes it may also work with GPT-3.5, though untested thoroughly.

Pivotal Quotes: "first come up with a detailed outline of how you think you should respond, then critique the ideas in this outline, mention the advantages and disadvantages, and ways it could be improved, then use the original outline and the critiques you made to come up with your best possible solution" — Sean Khosla: The core custom-instructions prompt intended to trigger iterative self-critique and improve final answers. "Overall, your tone should not be overly dramatic. It should be clear, professional, and direct." — Sean Khosla: Instruction to shape ChatGPT's style for a technical audience. "we don't need ChatGPT to kind of have its regular persona" — John Crohn: Crohn explains why a more restrained tone is preferable for professional listeners.

Implications: Listeners can use custom instructions to make ChatGPT more useful for technical workflows by improving answer quality and reducing fluff. The broader implication is that prompt structure and tone constraints can materially shape model behavior at scale.