Which chat screenshots give the best reading

Pick screenshots where you are doing most of the talking, in your everyday voice, across a few different situations. Five to ten clear, varied screenshots of your own messages give Flash Forge far more to work with than three near-identical ones, and legibility matters more than anything else.

Flash Forge does not judge what you say; it listens for how you say it. The rhythm of your sentences, whether you lead with feelings or facts, how you handle a disagreement, what you check twice. Your screenshots are its only evidence, so a little care in choosing them pays off.

What the AI is looking for

During the reading, the app picks up your rhythm and word choice, looks at signals such as tone, clarity, pacing and how you handle conflict, then infers motives, fears and desires. That works best when the screenshots show you making choices, not just reacting with a sticker. Useful signals include:

• How you make a decision: do you list options, ask others, or simply decide?
• How you disagree: directly, diplomatically, with humour, or by going quiet?
• How you help: practical steps, reassurance, or a question back?
• What you notice first: risks, people’s feelings, quality, efficiency, possibilities?
• Your natural length: short bursts or long, considered paragraphs.

Screenshots that work well

• A group chat where you are organising something: a trip, a dinner, a weekend plan.
• A work thread where you weigh options and land on a decision.
• A conversation where you disagreed with someone and held or changed your view.
• A friend asking for help and your reply.
• A relaxed exchange that shows your humour and the phrases you always use.

Aim for a spread. Five screenshots from five different situations usually tell a clearer story than ten from the same argument.

Scroll so that each screenshot captures a full exchange, not just a single line. The AI understands you better when it can see what you were replying to and how the conversation moved after your message.

Screenshots to skip

• Threads where other people do almost all the talking.
• Forwarded articles, memes, voice-message bubbles and long pasted quotes you did not write.
• Photos of a screen, heavy crops, or text shrunk so small it is hard to read.
• Anything with passwords, addresses, bank details, health information or other people’s private matters.
• Messages you wrote in a role, such as customer service scripts, unless that voice is genuinely yours.

If the AI cannot find readable chat text, the reading stops with “We couldn’t read any chat content in these screenshots”. Nothing is kept, and you can try again with clearer images.

Privacy habits worth keeping

The consent screen is explicit: the screenshots you choose, including any text and names visible in them, are sent to Google’s Gemini API to produce your reading. AgentSoul resizes them on the iPhone, processes them in memory and never stores them, keeping only the reading. Even so, the kindest approach to the people in your chats is to crop out their names and profile photos with the iPhone’s built-in editor before you pick the images. It does not change the result, because the AI is reading you.

Languages and mixed chats

You can use chats in English, Chinese or any mix of the two. Many bilingual people switch languages depending on who they are talking to, and that switch is itself part of your voice. If you want an agent that works mainly in one language, it can help to lean your selection towards chats in that language, but you do not have to.

Still unsure? Run the sample chats first to see what a full result looks like, then come back with your own screenshots.