Four dictation apps launched yesterday. Most are competing on the wrong axis.

Yesterday's board had SKI, Yap and Phantom Voice all launching, with Wispr Flow promoted alongside them. Four takes on getting speech into a machine, one day.

They are mostly racing on the same three things: accuracy, latency, and whether it runs locally. All real, all table stakes within a year.

Here is what I think the category is missing.

Dictating prose and dictating to an AI are different problems, and almost everything is built for the first one.

When a human reads the output, polish matters. Strip the filler, fix the punctuation, tidy the sentence. Every one of these ships some version of that and they are right to.

When a model reads the output, polish is a liability. My messy sentence carries hesitation and self-correction that the model actually uses. A cleaned-up version can quietly change what I asked for, and I will not notice until the agent does the wrong thing confidently.

The second gap is vocabulary. Half of what I say to a coding agent is npx, pnpm, useEffect, MCP, branch names, repo names. General speech models mangle those confidently and identically every time, and none of these let me teach them. whisper.cpp supports prompt biasing. Apple's API supports contextual strings. The hooks exist and nobody is exposing them.

So the thing I would build is boring: a per-project word list, and a toggle that turns cleanup off when the destination is an agent rather than a person.

Curious whether people who dictate into Claude Code or Cursor daily see this the same way, or whether I am over-indexing on my own workflow. Genuinely open to being wrong about the cleanup part in particular.