РусскийRUEnglishEN

Real-time AI suggestions

Overview

AI suggestions are answers that Whisperer generates right during a call, based on the live transcript, your questions, and your knowledge base. The answer arrives as a stream into the overlay's AnswerArea: text appears as it's generated, supports Markdown, reasoning disclosure, and page pagination (1/N). Suggestions work in several modes — from a plain answer to analyzing a screenshot and designing System Design with diagrams.

All requests go through our AI provider with models from the Claude, GPT, DeepSeek, Qwen, Gemini, Grok, and Llama families. Which models are available depends on your plan (tier gating): the top frontier models are included only on Max.

When to use

  • Interview: quickly phrase an answer to a behavioral question or sketch out a System Design solution.
  • Sales and negotiations: find an argument, respond to an objection, double-check a fact.
  • Technical interview / pair programming: get code in coding mode.
  • Making sense of what's on screen: send a screenshot of a whiteboard, diagram, or code in vision mode.

Step-by-step

  1. Ask a question as text. Type the question into the CommandBar input field and press Enter — the answer starts streaming into the AnswerArea.
  2. Or ask "from what was heard." Press Enter on an empty field — Whisperer takes the latest transcript fragment (the other person's question) and answers it. Handy when the question was just spoken aloud.
  3. Or attach a screenshot. Take a screenshot (with the captureScreenshot hotkey) — the frame goes to vision mode, and the model analyzes the image.
  4. Read the streaming answer. Text appears as it's generated. You can expand the model's reasoning if present.
  5. Page through. Long answers and a series of requests are split into pages — switch with the 1/N pagination in the AnswerArea.

Answer modes

  • Plain answer (role responses) — a text answer to a question, taking context and the knowledge base into account.
  • Coding (role coding) — code generation and explanation, a model optimized for programming.
  • Vision / multi-screenshot (role vision) — analysis of a single image or several accumulated frames (the array of screenshots goes to the vision endpoint). Useful for long tasks that don't fit on one screen.
  • System Design (role system_design) — a structured answer with Mermaid diagrams (components, data flows, architecture diagram).

How the prompt is assembled (in brief)

The system prompt is assembled by priority, top to bottom:

  1. Role prompt — your personal template from Prompt Studio.
  2. User context — the data you set about yourself and your role.
  3. RAG block — relevant fragments from the knowledge base, selected dynamically for each request.

Different model roles handle different tasks: responses (chat), coding (code), vision (images), system_design (diagrams), generation (mind map), transcription (Whisper), embedding (vectorization for RAG).

Model tier gating

On Free, basic models are available; on paid plans, a wider set. Frontier models (the top ones: Claude Opus, flagship GPT, etc.) are available only on Max plans. More in Plans and what's included.

Screenshots

📸 [Screenshot: AnswerArea with a streaming answer and 1/N pagination]

📸 [Screenshot: a "from what was heard" request via an empty Enter]

📸 [Screenshot: a System Design answer with a Mermaid diagram]

📸 [Screenshot: vision mode with an attached screenshot]

Common mistakes

  • Enter sent nothing. If the field is empty, Whisperer takes the question from the transcript — make sure the other person is heard (the "Screen Recording" permission is required) or type the question as text.
  • A frontier model isn't available. The top models are enabled only on Max — on Free/Start/Pro the nearest available one is selected.
  • Context "didn't work." User context is reference information about you, not a command; ask the actual question in the input field.
  • Screenshot wasn't analyzed. Multi-screenshot and vision aren't available on all plans — check the capabilities in Plans.

Best practices

  • Ask short, specific questions — streaming starts faster and the answer is more accurate.
  • For a long on-screen task, accumulate several frames and send them as a multi-screenshot rather than one at a time.
  • Build up your knowledge base — RAG mixes your facts into the answer and reduces "hallucinations."
  • Set up a personal role prompt tailored to your task (interviews / sales / teaching).
  • For architectural questions, use System Design mode — you get a ready Mermaid diagram instead of a text-only description.

Related articles