Quick answers to what people usually ask. For data specifics, the privacy page is the long-form. For rules of the road, see terms.
What is RawReply?
A small chat product plus two sidekick tools (Caption & Roast). The chat helps you write replies that sound like you, not like an AI. The tools take a screenshot you upload and generate captions or roast comebacks for it. Both run on Claude (Anthropic) with AWS Bedrock as a backup for the vision side.
Do I need to sign in?
For the chat product — yes. Google sign-in. We get back your email + name; that's all.
For the Caption & Roast tools on the landing page — no. They're anonymous and rate-limited so people can try without committing.
Is it free?
Yes, for now. If that changes, anyone with an account gets an email before anything is charged. We're not going to surprise you with an invoice.
What does the system “remember” about me?
Facts you mention in conversation that look useful for future replies — your role, the project you're building, decisions you've made, preferences you've stated. A reflection step writes these to a memory store after each chat turn. The next time you open any chat, the assistant has that context and replies in your voice from the jump.
Memory is per-account and never shared with other users. Facts that haven't been useful in 90 days quietly age out. You can see + edit + delete every fact yourself at /settings/memory.
How do I turn memory off / delete what's stored?
Open /settings/memory. Every fact has a delete button. We don't have a global “memory off” toggle yet — if that matters to you, email privacy@rawreply.com and we'll prioritize the feature.
What are Projects?
A folder for related chats, with two extras:
- Custom instructions — a system-prompt prefix that applies to every chat in the project (e.g. “always reply in lowercase, no em-dashes, talk like a senior IC reviewing my code”).
- Files — upload docs, screenshots, PDFs, code. The assistant reads them when relevant to your question.
Projects are private to your account. Other people can't see them. To delete: open the project, kebab (⋮) → Delete. Chats inside move back to the main list; files and project memory are wiped.
Why are the roasts in Hinglish for me but not for my friend in NYC?
The Roast tool reads two signals to decide what voice to use: your IP-derived country (from Cloudflare) and your browser language. If either says India / South Asia, the full 8-card deck leans into Hinglish + desi humor. If you're in the US/UK/etc., 6 of the 8 cards stay in English; two cards (Punjabi + Bollywood) always include some desi flavor by design.
You can override neither today. If you want a Western roast on an Indian IP (or vice versa), that toggle is on our list.
Are my chats used to train AI?
No. RawReply doesn't train its own model. Anthropic and AWS Bedrock — the providers behind the replies — also don't train on API traffic by default; you can verify that in their public policies (linked from the privacy page).
What if I share a chat with someone — what do they see?
A read-only snapshot of that one chat. They can't see your other chats, your projects, or your memory facts. If you delete the original chat, the share link stops working. Share links are opaque random tokens; search engines don't index them.
What if I share a Roast card?
Card pages are indexed by search engines — they exist to be shared. Don't put anything on a card you wouldn't want public. Cards auto-delete after 7 days.
Why does it sometimes take 15-20 seconds to generate?
The Caption & Roast tools generate up to 14 things in parallel — base candidates, a 8-card voice deck, an aura label, a sample group-chat reaction. Each is a vision model call. They run concurrently but the slowest one sets the total wait. Speed is on our list; we'd rather give you 8 good options than 1 fast bad one.
Why does the reply sound a little “AI” sometimes?
Large language models default to a polite, hedging, corporate-safe register. We tune the prompts to push past that — quoting your actual words back, naming the archetype you're trying to be, using mockery instead of polite sarcasm. It's a sliding scale; some inputs land harder than others. If a specific reply feels off, screenshot it and email us — that's how the prompts get sharper.
Is there a mobile app?
Not yet. The web app works on iOS and Android browsers and is built for one-handed use. Add it to your home screen (Safari → Share → Add to Home Screen) for an app-like launcher.
Something not covered here?
Email privacy@rawreply.com. Privacy, billing, bugs, feature requests, “your roast was wrong about my friend” — all the same inbox, one real person on the other end.