Guide
Writing issues an agent can act on
A chat prompt is thrown away when the tab closes. An issue is the same instruction written down where an agent, a teammate or next-month-you will pick it up cold. That difference is what makes it worth speaking rather than typing.
Agents now read the tracker. A task with a title and one line of body produces the same thing from a coding agent that it produces from a new hire: a plausible attempt at the wrong problem, and a round trip to find that out.
The fix is not a template. It is the two paragraphs of context that live in your head at the moment you file the issue and nowhere else — and the reason they do not get written is that you are filing the issue while holding something else, and typing two paragraphs is the part you skip.
Say it while you still know it
The best moment to describe a bug is the thirty seconds after you hit it, when you still remember what you were doing when it happened. That is also the moment you least want to stop and type. Speaking gets the whole account down while it is still whole — what you did, what you expected, what happened instead, and the thing you noticed but were not sure mattered.
That last one is the sentence dictation actually buys you. Nobody types "it might be unrelated, but the window had just come back from sleep". Almost everybody says it. It is also, often, the bug.
What an agent needs that a human infers
A teammate fills gaps from shared context; an agent fills them from the average of its training data. Four things close most of that distance, and all four are easier to say than to type:
- Where. The file, the screen, the endpoint. Say the name — that is what the personal dictionary is for.
- What good looks like. One sentence of done, in your words, is worth more than a checklist someone pasted from another issue.
- What is out of scope. "Do not touch the migration" prevents more work than any amount of description invites.
- What you already tried. A rejected approach removes a whole branch of the answer, and it is the line most reliably lost to typing fatigue.
Dictating into the tracker
Issue trackers live in a browser, and browser text fields take Voicecape’s clipboard path — the text is placed on the clipboard, a paste is sent, the paste is observed to land, and whatever you had copied before is put back. Your previous clipboard survives, which matters here because the thing you had copied is usually the stack trace you are about to paste under the description.
Hold the key, speak a paragraph, let go, and it appears at the cursor. For a long issue, speak it in paragraphs rather than in one breath: the release is a natural place to think, and it leaves the text in a state you can read before continuing.
Markdown is the one thing to say out loud rather than expect: dictation gives you prose, so the dashes and backticks are still yours to type. In practice the shape that works is prose by voice, structure by hand — which is the same split that works for code.
Verbatim for the parts that must survive exactly
Reading an error message aloud into an issue is a real use, and it is the case where cleanup would be a defect rather than a feature. Verbatim mode leaves the words exactly as spoken; Exact Clean, the default, would still not rewrite them, but Verbatim is the mode to be unambiguous in. Switch for the quoted line, switch back for the explanation.
A version number, a flag, a config key: same rule. Anything a reader might copy back out has to survive character for character, and the difference between the two modes is exactly this.
Review comments are the same job
The comment that says why is the one nobody writes, because it is three sentences and the approve button is right there. Speaking it is the difference between "nit: rename" and the paragraph explaining the failure mode you are actually worried about — which is also the paragraph an agent can act on, and the one the author does not have to guess at.
The same hotkey works in the review box, the commit message and the release note. None of those are agent prompts by name, and all of them end up in front of one.
Where this sits
When the prompt is a conversation rather than a document, the technique changes: dictating prompts to ChatGPT and Claude. Inside a coding agent, the terminal and the editor take a different insertion path and Claude Code has its own voice command — voice input for Claude Code covers that. Which apps take which path at all is mapped in typing with your voice in any app.