This lesson's interactive features are locked, please to keep using them

Speech to Text

Often, you already have the context your agent needs in your head, but typing it out is a huge pain in the butt.

Using speech-to-text (STT) to do brain dumps is probably the highest-impact/lowest-effort improvement you can make to your prompting.

Agents Summarize Well

You don't need to polish a spoken prompt as you would a PR description meant for a human. Agents have no problem understanding messy wording and corrections like "Um, actually, forget that, do this other thing," but what helps is the extra context you provide, not the additional words themselves.

A typed prompt might look like this:

Fix the checkout bug where discounts don't apply correctly.

Whereas a spoken prompt will look more like this:

Fix the checkout bug where discounts don't apply correctly. I think it only happens when the user has both a promo code and store credit. The old checkout flow handled this in applyCredits, but the new flow moved discount logic into the payment step. Please compare those paths before changing anything, and don't rewrite the whole checkout pipeline unless you find a real reason. Oh, and I think Dan made a PR about this 5 or 6 days ago; you can probably find that commit on the main branch.

That second prompt is messier, but it's much more useful to an agent.

For agent prompts, voice should be your default input method.

Which Tool to Use?

Some agent harnesses have STT built in, but honestly, we prefer to use a dedicated tool that will work the same way across your entire machine. After all, you might want to use it for tasks other than prompting in OpenCode.

Handy is a good default speech-to-text tool that works on Linux, macOS, and Windows. I've configured it so that:

I move my cursor into the text input I'm working with (like OpenCode's prompt field).
I hold down Ctrl+Alt+Cmd to start recording (it's a psychotic key combo, I know... but hey, it doesn't conflict with anything).
I speak my prompt.
I release Ctrl+Alt+Cmd, and the words are ~instantly pasted into the focused text input (it's wild how fast it is).

If you're on Apple silicon, Hex is another solid option. It's narrower in platform support, but good to be aware of.

My favorite thing about these speech-to-text tools is that they use local models. The recordings of your voice never leave your machine.

Assignment

Use your voice to give OpenCode a real task.

Set up Handy or another dictation tool if you don't already have one.
Dictate a prompt telling OpenCode to create a new examples/countdown.slop file. It should contain SlopLang code that prints 3, 2, and 1 on separate lines, then prints liftoff. Review the transcription before sending it.
Have OpenCode run the new example code and check its output.

Tell OpenCode to run and submit the CLI tests from the project root.