Your dictation should type what you said, and nothing else

Dictation into a coding agent has a failure mode that dictation into a text message does not. If your phone mistypes a word in a message to a friend, they read past it. If it mistypes a word in a prompt, the agent does the wrong thing, competently, for twenty minutes.

So the bar for speech in this product is not "usually close enough". It is that the words which reach the agent are the words you said.

The rule we build to

The only change permitted between the speech model's output and the text the agent receives is a find-and-replace of terms you listed by hand. Nothing else. That sounds austere until you consider what the alternatives do.

• No model that can rewrite you. A model that writes text is not a transcriber. A cleanup pass that tidies your grammar is also free to change your meaning, and it will, on the day you dictate something unusual - which is most days, if you work on anything specific.
• No substitution you did not ask for. A fuzzy matcher that decides "sure" was probably your name, or that "focus" was probably a model, is not correcting you. It is guessing, and a guess in a prompt is an instruction you never gave.
• No hidden vocabulary hints. Feeding the speech model a list of words you did not configure changes what it hears. That is a thumb on the scale, and you cannot see it from the transcript.

A wrong word from you is your problem, and you will spot it. A wrong word from us is worse, because it looks exactly like something you meant.

The dictionary is the part you do control

Every project has words the speech model has never heard: your product name, your repository, the internal library, the two colleagues whose names you say forty times a day. Those are the words that come out wrong, and they are the words that matter most in a prompt.

The dictionary fixes that permanently, and it takes about five minutes once. Open the Cockpit in a browser, choose Dictionary in the left rail, and add each term spelled the way it should be spelled.

The important detail is where that list lives. It belongs to your Gateway, not to the browser you added it in and not to one machine. Teach it once and every Director connected to that Gateway has it - along with the transcription of voice notes you record on your phone. That holds whether you use the Gateway we host or one you run yourself.

Why this is written down rather than assumed

Because it is the kind of rule that erodes quietly. Every one of the forbidden behaviours above arrives dressed as an improvement - reliability, polish, a smarter model. Each one makes the demo better and the product worse, and the damage happens inside a component where no downstream check can catch it.

The test we hold ourselves to is arithmetic: the stored transcript plus the accepted corrections must equal the text that was sent. Any edit whose source term is not on your own list is a defect.

Setting up your dictionary: teach it your words. Using dictation day to day: the dictation tutorial.