Output modes

When a recording finishes, Zerm has to decide between two things you cannot have at
once: text in front of you immediately, or text that the AI has already improved. The
Output setting, at the top of Enhancement, is where you make that trade explicitly.

It is also the only switch for AI enhancement. Instant means no AI; the other two use
it. There is no separate on/off toggle.

Instant

The transcript is pasted at your cursor the moment transcription finishes. No language
model is involved at any point — nothing is sent to a provider, nothing is waited on.

This is the fastest path in the app. Filler-word removal, your
dictionary replacements, and formatting still apply, because those
run locally and cost nothing.

Choose Instant if you dictate short messages, if you are offline, or if you would rather
fix the occasional word yourself than wait.

A fresh install starts on Instant, because the on-device enhancement model is not
downloaded yet. Once you set up a provider, Instant + Refine is the everyday choice.

Instant + Refine

The transcript is pasted immediately — byte-for-byte the same path as Instant, with the
same latency. The enhancement then runs in the background and, when it comes back,
quietly replaces what was already typed.

The paste never waits for the model. Refine has a short budget of its own, about eight
seconds. If the enhancement fails, times out, or you have already moved on, what you
have is the raw transcript, which is exactly what Instant would have given you.

Where in-place replacement works, and where it does not

Rewriting text that has already been typed into another application means writing
through the macOS Accessibility API, and not every application implements the parts
required to do that.

It works in native macOS text views — Mail, Notes, TextEdit, Xcode, most AppKit and
SwiftUI apps, and simple form fields.

It does not work in:

• Electron apps — Slack, Visual Studio Code, Cursor, Discord, Notion, and most of
  the category. They expose a value and a selection but no working setter.
• Web content in browsers. contenteditable surfaces nothing that can be set, which
  covers most web apps.
• Terminals — Terminal, iTerm, Ghostty, WezTerm, kitty, Alacritty, Hyper, Warp, and
  the rest. The shell owns the line buffer; accessibility only mirrors a read-only
  screen, so writing to it would be meaningless.

In all of those, Refine still runs — but instead of changing text you have already
typed, the refined version is offered to you, and what is on screen is left alone.
Nothing is silently rewritten, nothing is copied a second time, and nothing is
silently lost. Replacement is only allowed through a direct accessibility write;
Zerm never pastes the refinement via the clipboard.

Zerm does not guess. Terminals are on a fixed deny-list, and every other application is
probed once for a working setter; the answer is cached per app, so the second dictation
into a given app costs nothing to decide.

Refine is also skipped entirely while a password field is focused anywhere on the
system — Zerm will not so much as read the focused element in that state.

Enhanced

Zerm waits for the enhancement to return, then pastes once. Nothing appears until the
final text is ready.

It is the slowest of the three by exactly the round-trip time of your provider, and it
is the right choice when you want the pasted text to be final the moment it appears —
into a document, or into an app where Refine could only offer rather than replace.

Enhanced is also the only mode that can include what is on your screen as context. See
contextual awareness.

The wait is bounded by the enhancement timeout, which defaults to 15 seconds. If it
expires, Zerm pastes the raw transcript rather than losing what you said, and tells you
it timed out. See AI enhancement for the timeout and retry settings.

Choosing

	Text appears	AI improves it	Good for
Instant	immediately	never	speed, offline, short messages
Instant + Refine	immediately	afterwards, in native apps	everyday dictation
Enhanced	after the model	before you see it	when the first paste must already be final

The mode is a global setting. A Power Mode can choose a different
output for one app or website, and ⌘E in the recorder switches between Instant and your
AI mode for a single recording — see
enhancement shortcuts.

A Power Mode with Auto Send uses Enhanced instead of Instant + Refine, because the
message is sent before a refinement could replace it.