Riffwise：打造 AI 回覆工具，難在界定它可查看和執行的事

I build Riffwise, a Chrome extension that drafts replies from the conversation already open. This is a product-building note, not a benchmark: the question I keep coming back to is not “Can AI write a sentence?” It is “What should an AI writing tool be allowed to see and do?”

A reply assistant sits close to somebody else’s words. That makes its trust boundary part of the product, not a footnote.

1. Make the request explicit

The user starts a reply or quote-draft action. Riffwise reads the relevant post or email context for that request. The writing request is sent to OpenAI to generate the draft. That is a cloud AI call, and users should be able to understand that before they use it.

A useful product test: after a quick look at the screen, can a new user say what context is sent, who receives it, and why? If not, the interface needs to explain it more clearly.

2. Keep “draft” separate from “send”

The extension returns words to review. The user can edit, insert, copy, or ignore them; the user decides whether to post or send. Inserting text into a composer is not the same as publishing it, and the interface should keep that distinction obvious.

That boundary also makes it easier to say no to a bad draft. A generated reply is a suggestion, not an action taken on someone’s behalf.

3. Give people controls that change the result

Tone, length, language, and personal writing instructions matter more than a dozen novelty modes. They let the person shape a draft and make the tool useful across different conversations. They also make the limits visible: the model can help with a first pass, but the person still owns the words.

4. Be candid about the tradeoff

Human review does not make a cloud tool private by magic. Riffwise sends the writing-request content and relevant attached post media to OpenAI for generation. Its privacy policy says ordinary prompts and generated replies are not intentionally stored in the application database. If you would not be comfortable sending a conversation for processing, do not use it for that conversation.

I’m still working on making the first-run explanation clearer. If you build browser tools or AI features, I’d be interested in your answer: where do you draw the line between helpful context and too much access?

Disclosure: I built Riffwise. I used AI assistance to prepare this article and reviewed the final text.

If you want to inspect the product, Riffwise is in beta for desktop Chrome. The Free plan has 40 successful generations per calendar month, with no card required; an account is required. Nothing posts or sends automatically. Try Riffwise in the Chrome Web Store.