OpenAI has released a new prompting guide aimed at everyday users, simplifying the process of interacting with its AI models. The guide focuses on practical advice, including four optional building blocks for prompts, rather than technical details like API parameters or model tuning. It reflects a shift in focus from developer-centric documentation to a more user-friendly approach. The guide is part of a broader effort to streamline how users engage with OpenAI's products, such as ChatGPT Work, which is built on Codex technology and the GPT-5.6 model. The guide also outlines how to structure prompts around a goal, context, output format, and boundaries, with the emphasis on starting small and adding rules only when necessary.
The guide recommends leading with the result rather than a sequence of steps, encouraging users to leave room for the AI to search, compare, and adjust its approach. It highlights that a target audience or format shapes the output more than detailed instructions. OpenAI also suggests using constraints rather than step-by-step scripts, with examples like keeping approved dates and budget figures unchanged or preparing a message as a draft without sending it. The same less-is-more logic applies to context, with only relevant sources attached to influence the answer. The guide includes options like spreadsheets, PDFs, images, and shared project files, along with plugins for Google Drive, Gmail, Slack, and GitHub.
The guide differentiates between ChatGPT for quick tasks and ChatGPT Work for more complex projects that require multiple sources or produce larger deliverables. Work tasks consume more credits but offer time savings and support for important decisions. For recurring tasks, OpenAI suggests refining prompts manually before automation. Users are encouraged to refine output through follow-ups rather than perfecting the first attempt. Preferences that carry across sessions are stored in 'Settings > Personalization' as 'Custom Instructions,' while task-specific details remain in the prompt. For Codex, OpenAI introduces features like 'Steer' and 'Queue' to influence tasks mid-run, with sandbox mode restricting file and network access unless approved. Two slash commands, '/plan' and '/goal,' help with multi-step projects, while '/review' supports local or GitHub-based code reviews.
Source: thedecoder
