Using Custom GPTs

A custom GPT is a version of ChatGPT that operates according to a specific task or workflow, without needing to be explained from scratch every time. By pre-loading your team's context, desired format, and recurring rules, you can get more consistent results.

There are three core elements. You define behavior with instructions, reinforce knowledge with uploaded files, and enable tools such as web search, data analysis, and connected actions. As a result, you can reduce repeatedly entering the same prompt or re-uploading the same files over and over.

Regular chat is suited to spontaneous questions or one-off tasks. Custom GPTs, on the other hand, are better suited to highly repetitive work like the following.

• Automating repeat tasks: Locking in frequently used prompts as a workflow
• Adding tools or integrations: File analysis, external app integration, and leveraging deeper context
• Maintaining consistent context: Repeatedly applying the same tone, structure, and rules

Examples made by the OpenAI team include ChatGPT Use Cases for Work, Professional Writing Coach, Data Analyst, Coding Assistant, and Visual Designer. Each focuses on discovering work ideas, polishing writing, summarizing and charting uploaded data, generating and debugging code, and creating brand imagery, respectively.

The creation process begins by pressing Create in GPTs. In the Create tab, you converse with the GPT Builder to make a draft, and in the Configure tab, you set details such as name, description, instructions, conversation starters, knowledge, capabilities, and custom actions. Since writing instructions is especially the hardest part, it's faster to first make a draft with ChatGPT and then refine it while looking at real examples.

Validating with evals before sharing is also important. Prepare 10-15 representative questions and correct answers to check accuracy and reliability, and adjust the instructions or knowledge based on the results. It's also easy to miss pressing Update in the top right when changing settings, so make sure to save.

There's no need to start complicated. Pick one task you're already repeating, build a first version, and validate it with a small test set. Once it starts working reliably, share it with your team to get the same quality of results with less effort.