Post

When I change my mind halfway through a sentence, I'd want a quick readback of the final decision before the code changes. 'Date archives yes, tag pages no.' Did you use that, or catch misunderstandings in the preview?

Did you find yourself explaining the why more out loud than you would in a typed prompt?

Nice So both you and the agent cooked 👏👏🤷‍♂️ You’re welcome

cooked dinner, shipped a feature, never touched a keyboard. my version of voice coding is dictating one bug into my phone and getting three back. real question, how many spoken turns did the newsletter page take.

I ran a voice-vs-typed test with a coding agent last week: ~40% of voice utterances needed a correction pass, vs ~10% typed. Bottleneck wasn't transcription - it was catching my own vague instructions mid-sentence. Curious if you found a correction flow that keeps pace with the cooking.

I tried a smaller version for writing: dictation, transcription cleanup, then my own edits. Your mid-sentence changes of mind are the interesting part here. They keep the thinking in the conversation, instead of handing over the topic and getting a finished opinion back.

Working but wrong for the project is the hard catch. The import it picked: was that something earlier work in the repo had already settled, or something only a review could have caught?

Me mumbling to the AI has been a daily ritual for a while now. Also while sitting behind the computer it's just a quicker way to get info into the model. They understand STT artifacts surprisingly well.

Conversation as Specification.

The split between voice-driven building and typed PR review is useful context. The private-repo import is exactly the sort of choice that a working page won't reveal.

The trickiest part here seems to be content visibility, not the voice interface: Substack duplicates stay out of search, while sponsor posts become searchable only when public. Did you add tests for those rules, or mainly verify them through the preview and PR review?

Reverting a fix from twenty minutes ago is the regression that stings most, because you already paid for it once. When the read-back catches one, does the fix get written down anywhere the next session will see?

A before/after diff plus a readback you didn't write yourself is a sharp bar. When you sign off, which matters more: seeing exactly what changed, or knowing you can undo it?

It’s so fun isn’t it? I’d much rather code like that and the voice session can coordinate other sessions.

Several rounds inside one session makes sense for review. When the same comment comes up again in a later session, do you carry it over by hand, say into a notes file, or just flag it again?

Cooking dinner while shipping a feature is a good stress test for voice input. The rambling "um, I don't want it on the tag pages" prompt still working says a lot about how low the bar has got. Did you have to fix much afterwards? Nico@whatsticks

Coding by voice while the pasta boils is peak agent workflow. Did you have to review much of the diff afterwards, or did it mostly land first try?

Why does your repo have a commit entitled “where’s the paprika?”