Smart Prompter

I originally meant Smart Prompter to be a little demo of how much useful work Apple's on-device speech and language models can do, even on relatively low-end hardware. Then my NLP background got involved and the scope got a little out of hand--separate microphone and call-audio capture, multilingual transcription, fuzzy topic matching, model-assisted coverage tracking and a floating cue window trampled out before I knew it.

The app listens during video calls, follows a script of talking points and suggests a short next response. Its separate cue panel floats above the call while the main window handles setup, transcripts and manual corrections. Audio and transcripts stay on the Mac are not stored or handled beyond each session.

Requirements

• macOS 26 and Xcode 26
• Apple Intelligence enabled for contextual coaching
• an on-device Speech recogniser for the selected locale

No cloud API or account is required. Audio and transcripts are not written to disk.

Build and run

make test
make build
make run

make build selects the first valid Apple Development identity in the user keychain. The identity gives macOS a stable designated requirement and preserves privacy permissions when the executable changes. The build stops if no identity is available; it does not fall back to ad-hoc signing.

Set CODE_SIGN_IDENTITY when an Apple Development or Developer ID identity is already available:

CODE_SIGN_IDENTITY="Apple Development: Your Name (TEAMID)" make build

On first use, macOS asks for Speech Recognition, Microphone, and Screen & System Audio Recording access. If you deny one, enable it later in System Settings → Privacy & Security.

Enter one topic per line. Earlier topics are preferred when several points fit equally well, but conversation relevance can select a later point. Select a downloaded speech language and choose Start listening. The language menu keeps installed models in its short top-level list. Get More Languages… opens the full downloadable list; Smart Prompter installs a selected model when listening starts. Drag the semi-opaque cue panel by its background and resize it from any window edge or corner. Show Cue restores it after it has been hidden. Click a script item to mark or unmark it manually. Unchecking an automatically covered item keeps it manually uncovered for the rest of that listening session.

How speaker labels work

ScreenCaptureKit supplies two distinct audio outputs:

• You — the selected microphone;
• Call — system audio, excluding Smart Prompter itself.

This signal-level split is fast and deterministic for calls. It does not identify individual remote participants. A meeting with several remote speakers will label all of them as Call.

Speech recognition is configured to run on-device. Topic coverage has a deterministic keyword fallback; when the system language model is available, Apple Intelligence adds contextual coaching. If Apple Intelligence is disabled, the app continues to show the next uncovered script item.