Prompts

Prompts & AI Actions

Prompt actions run your dictated or selected text through a configured LLM provider for AI-driven text processing. They power the cleanup workflow and can be bound to profiles or triggered from a palette.

Prompt actions

A prompt action is an LLM-driven text-processing step. You define what the action should do, and TypeWhisper sends the relevant text to an LLM provider to produce the processed result.

In the Linux branch, prompt actions support:

• LLM-driven text processing
• Provider overrides — point an action at a specific LLM providers provider and model rather than the default
• Action plugin routing — send the result to an action plugin
• A per-action hotkey, and a manual-only flag

Seeded prompt actions

On a fresh install, TypeWhisper seeds these preset actions (all enabled):

Action	What it does
Translate to English	Translate the text to English, returning only the translation.
Write Email	Rewrite the text as a polished, professional email body.
Format as List	Convert the text into a clean bullet-point list.
Action Items	Extract action items and tasks as a numbered list.
Reply	Write a concise, professional reply matching the original tone.

A separate Auto Clean Up Text action is also seeded, but disabled — it drives the flagship cleanup workflow. See Text cleanup.

The Prompts screenshot above shows the maintainer's own renamed prompts and differs from these shipped defaults.

Per-action options

Each prompt action carries these options in the editor:

Option	Effect
Provider override	Run the action against a specific [[LLM providers
Action plugin	Route the result to an action plugin instead of inserting it normally. See [[Other plugins
Hotkey	A per-action hotkey that triggers the action directly. Assign it in the editor.
Manual only	When on, the action runs only when invoked manually (from the palette or its hotkey); it is hidden from the profile prompt-action picker and is never auto-applied by a profile during dictation.

Provider overrides

By default a prompt action uses your configured LLM provider. You can override the provider per action so one action runs against a specific LLM providers while others use the default. LLM responses stream token by token into the dictation overlay and the prompt palette when the provider's Stream responses toggle is on.

Action plugin routing

A prompt action can route its result to an action plugin. See Other plugins for the available action plugins.

Prompt palette hotkey

An optional prompt palette hotkey lets you trigger prompt actions on demand. Assign it on the shortcuts panel — see Global hotkeys.

Spoken commands

Spoken command mode lets you steer TypeWhisper by voice instead of only dictating text into apps. When it is on, every dictation is checked for a spoken keyphrase at the start (default "TypeWhisper"). If the keyphrase is present, the rest of what you said is treated as a command rather than text to type: TypeWhisper decides whether you meant to edit your current selection or create new text, then transforms the highlighted text or writes fresh text at the cursor — streaming the result into the overlay.

Spoken commands ship disabled. Enable and configure them under Settings → Prompts:

Control	Effect
Enable spoken commands	Turns the keyphrase check on. Off by default.
Keyphrase	The word that marks a dictation as a command (default TypeWhisper). Matching is fuzzy and speech-to-text tolerant, so a slight mishearing still works — but it tightens as more words fold in, so ordinary dictation isn't swallowed as a command.
Command model	The LLM used to interpret and run commands. Leave on Use default provider to follow your default [[LLM providers

Edit vs create

A lightweight classifier decides intent from your phrasing and routes the command accordingly:

• Edit / transform — operates on your highlighted selection (for example "TypeWhisper, make this more formal"). If a saved prompt action fits the request, that action is used. If nothing is highlighted, the command tells you so rather than acting on stale clipboard contents.
• Create — generates new text at the cursor (for example "TypeWhisper, write a thank-you note"). A create command never triggers a copy, so it is safe to run in a terminal.

Plugin-backed prompt actions are excluded from spoken-command matching, because their output is routed to a plugin rather than to the cursor.

Terminals: an edit command copies your selection with Ctrl+Shift+C in terminal apps (plain Ctrl+C there is SIGINT, which would interrupt whatever is running). TUI editors such as Neovim keep the visual selection internal, so no OS-level copy can read it — an edit command in that case reports that it couldn't read the selection instead of guessing.

Spoken commands are recorded in History like dictations, tagged with a "Command" badge, and their LLM request/response is visible in the History Inspect panel when provenance capture is enabled.

Relationship to the cleanup pipeline

Prompt actions are the LLM half of the dictation cleanup workflow. The deterministic cleanup levels and per-profile formatting choices live in the cleanup pipeline; the Medium and High cleanup levels route through your configured LLM provider — the same providers prompt actions use. The flagship setup binds the seeded Auto Clean Up Text prompt action to a profile so dictation flows straight through cleanup before insertion. For the formatting and cleanup specifics, see Text cleanup.

Long-term memory

When a memory storage plugin is enabled alongside a configured LLM provider, durable facts extracted from eligible transcriptions can be recalled as context by future prompt actions. See Long-term memory.

Related pages

• LLM providers — the providers prompt actions and cleanup run against.
• Long-term memory — durable facts that prompt actions can recall as context.
• Global hotkeys — assign the optional prompt palette hotkey.
• Text cleanup — the cleanup and formatting pipeline that prompt actions drive.
• Profiles — bind a prompt action as a per-context override.
• History — where spoken commands are recorded, and the Inspect panel that shows what was sent to the LLM.

---

Changelog

Date	Change
2026-06-17	Initial version.
2026-06-17	Added Prompts list and editor screenshots.
2026-06-17	Added prompt-palette screenshot.
2026-06-17	Documented the seeded preset actions and disabled Auto Clean Up Text, noted the screenshot shows renamed prompts, and added per-action provider/model override, hotkey, and manual-only options.
2026-07-08	v0.12.0: documented spoken command mode — the keyphrase, edit-vs-create routing, the command-model picker, terminal/Neovim behavior, and history recording.