A window only you can see
You are showing your software on a call. The client asks something you did not prepare for. GhostCue has heard the question, and one sentence to say, with the points behind it, is already on your screen. Nobody else can see it: the window is left out of every screen share, recording and screenshot by Windows itself.
Which is also why this is a drawing: a screenshot of the overlay comes out without it.
Swipe sideways to see the whole window →
CLIENT What would fifty drivers cost us on the annual plan?
On the annual plan it comes to 510 dollars a month for fifty drivers, billed as 6,120 for the year.
The pricing page lists 12 dollars per driver per month
Annual billing takes 15 percent off, which the same page shows
Not in brief: whether a fleet this size gets a volume discount
- State. Listening, starting, off, or an error, which then fills the window with what went wrong.
- What it is doing. The devices it listens to, or "Thinking…", "Reading the window…" while an answer is on its way.
- The capture check. Says the window is hidden from screen shares, and turns into a red
VISIBLE TO SCREEN SHAREif Windows would not hide it. - Buttons. The same as the hotkeys. Clicking them never takes the keyboard from the app you are demoing;
historylists the call's earlier answers. - The question it is answering.
ASKEDin place ofCLIENTis the question as the assistant understood it, on Urdu calls. - What to say, in your own voice, ready to read out.
- The points behind it: facts, figures and where they came from.
- Not in brief: your material did not cover this, so the sentence above is a holding answer, not a promise.
- What it heard last, or here, the window you captured with Ctrl+Alt+W, and whether the whole page came with it.
- The hotkeys, all with Ctrl+Alt, so you never have to reach for the overlay mid-call.
- Drag anywhere to move it; drag this corner to resize. Size, position, text size and opacity are kept.
The assistant
Listens to the call and tells you what to say when the client asks a question, drawing on your brief and your own documents.
For recorded demosThe teleprompter
Walks you through a timed script while you record a demo video: what to show, what to do, and the words to say.
From the client's question to your answer in seconds
An answer starts to appear about two to four seconds after the client stops speaking: the question, one sentence to say, and up to three supporting points.
- ContextWrite the brief: what you are showing, who the client is, what they care about, and what you must not promise.
- KnowledgeAdd FAQs, pricing or specification files for the answers to draw on, and a project folder to dig into.
- AudioPlay a sound and speak, and check that both level bars move.
- Start listeningPress it before the call. The first start takes about ten seconds while Claude Code starts.
Answer from what is on screen
Asked about a web page, an email or what the client is showing you? Press Ctrl+Alt+W on that window. A page in your browser goes with its whole text, scrolled or not; anything else, such as the client's shared screen, goes as a screenshot, so the answer knows what you can see and no more.
Answers that know their limits
A point marked Not in brief: means your material did not cover the question, and the sentence above it is a holding answer rather than a fact.
As long as you need
Answer length runs from one sentence to a full explanation. A question about a set of things gets every member, and a request for code gets the code.
Look back through the call
Every answer stays until the next call. Step back with Ctrl+Alt+, or open the history and pick a question. A new answer waits for you instead of pushing away the one you are reading.
Urdu and English
For Urdu-speaking clients, answers come in the client's language, with Urdu in Roman letters. The overlay shows the question as it was understood, so you can tell when it misheard.
Hotkeys, all with Ctrl+Alt
On your computer, with your own Claude
Both sides of the call
The client is whatever your speakers or headphones play, captured with WASAPI loopback. You are the microphone. Wear headphones: GhostCue drops microphone speech that overlaps the client's.
Transcribed locally
Speech is turned into text on this computer with faster-whisper. In English the base model takes about 0.7 seconds a sentence on a six-core CPU.
Answered by your Claude Code
Answers come from your own Claude Code, kept running for the length of the call so each answer starts in under a second. The transcript, and any window you capture, is sent to Anthropic to be answered and counts towards your plan's usage.
Kept to yourself
Settings and call transcripts stay in %APPDATA%\GhostCue, and transcripts can be turned off. No Claude Code session is saved.
Record a demo without losing your place
Write the demo as a plain-text script of timed cues. The overlay shows each cue's directions on the left and the words to say on the right, and moves on by the clock or when you press a key.
[0:12] Sign in DO: Type the email and password, then click "Sign in". SHOW: The dashboard, with the three summary cards. SAY: I'll start by signing in with a demo account. This is the dashboard. NOTE: If the login is slow, keep talking.
In manual mode, cues change only when you press Next or Previous, and the countdown turns red when you fall behind the script. The teleprompter uses only the Python standard library.
Check it once
- Share your screen in a test meeting and confirm on a second device that the overlay is absent. Its header reads
hidden from capturewhile the exclusion is active, and a redVISIBLE TO SCREEN SHAREif it could not be applied. - The overlay is hidden but your mouse pointer is not, so use the hotkeys.
- Speech recognition mishears names. List yours under Context, in "Names and jargon".
- Transcribing a call may require telling the people on it, depending on where they are.
Get started
Needs 64-bit Windows 10 version 2004 or later and, for the assistant, Claude Code, installed and signed in.
GhostCue is not code-signed yet, so Windows SmartScreen says "Windows protected your PC". Click More info, then Run anyway. The speech model downloads once, the first time you start listening.
Or run it from source, with Python 3.10 or later
# get the code git clone https://github.com/Blankscreen-exe/ghostcue.git cd ghostcue # install python -m venv .venv .venv\Scripts\python -m pip install -r requirements.txt # run: the assistant lives in the system tray GhostCue.bat
The teleprompter is on the tray menu and in the Start menu, or run python -m ghostcue prompter my_script.txt.