Draft support replies from your own policy doc

Gather the current refund policy, shipping times, plan limits, and your top help articles. Delete old versions so the assistant can't quote last year's rules. Strip customer names, order numbers, and payment details out of anything you upload.

A project is a workspace that keeps your files and instructions attached to every chat inside it. Claude's help center says projects are available on free accounts, with a limit of five projects. In ChatGPT, select New project in the sidebar, then add files; free accounts can attach up to five files per project, per OpenAI's help page.

In the project instructions, tell the assistant to answer only from the uploaded documents, quote the line it relied on, and reply 'not covered in our docs' when it can't find one. Anthropic's guide to reducing made-up answers recommends this mix: permission to say it doesn't know, plus a quote for every claim.

For each draft, find the quoted line in the source document. If the quote is missing, reworded, or from the wrong policy, fix the reply yourself. Anything marked 'not covered' goes to a teammate who owns the policy, not back to the bot for a second guess.

Check: Every reply you send has a quote you found in the current document.

OpenAI's help page says that on ChatGPT Free, Plus, and Pro, project content may be used to train its models if the 'Improve the model for everyone' setting is on, while Business, Enterprise, and Edu workspace content is not used for training by default. Anthropic's privacy center says Claude Free, Pro, and Max users can switch training off under Settings, then Privacy, then Help Improve our AI models, and that its commercial products, like Team and Enterprise plans and the API, aren't used for training by default. Review that setting, or your workspace's equivalent, before you upload internal documents.