Voice-driven development

Here's a thing nobody tells you about working with capable AI agents: the bottleneck moves. Once the agent can do the work, the limiting factor becomes how fast and how precisely you can say what you want. Typing a paragraph of intent is slow. Speaking it is not. I dictate almost everything now — prompts, corrections, plans — through a flow-dictation tool (Wispr Flow), straight into whichever agent's terminal is in front of me.

§1Why voice changes the loop

Three effects compound:

• Intent gets cheaper to express. You'll happily say three sentences of context you'd never bother to type. Richer prompts, better results, less back-and-forth.
• Delegation stops feeling like overhead. When issuing a task costs a sentence, you delegate the small stuff you'd otherwise just do yourself — and the small stuff is most of the work.
• You can drive a fleet. Steering four agents by keyboard is exhausting. By voice, you narrate to whichever one is listening and keep moving.

§2Designing for transcription artifacts

Dictation isn't perfect: you get missing punctuation, odd capitalisation, homophones, and the occasional dropped word. The wrong response is to slow down and enunciate like a robot. The right response is to make the agent robust to it — build the tolerance into the house rules:

a rule every agent reads (from the shared context file)

# Voice input: my prompts are dictated and may have artifacts —
# missing punctuation, homophones, rough casing, short fragments.
# - Infer the intended command from context when it's safe.
# - But DO NOT silently "fix" exact-text I dictate: URLs, file paths,
#   numbers, IDs, config keys, commit messages. If a literal value looks
#   off, echo it back and ask — don't normalise it away.

That one paragraph removes almost all the friction. The agent quietly repairs "add a data attribute to the nav bar" from a comma-less run-on, but stops and confirms if it hears a version number or a path it isn't sure about — exactly where a wrong "correction" would cost you.

§3Shorthand: one phrase, one workflow

The real speed comes from a personal shorthand. Over time, repeated multi-step rituals get a name, and saying the name triggers the whole thing. This works because the expansion lives in the shared context file — the agent knows what your shorthand means.

a shorthand glossary in the context file

# My spoken shorthand → what you should actually do
- "ship it"        → run the full deploy ritual: bump changed-asset
                     versions, verify one version site-wide, commit,
                     then wait for my go-ahead to push.
- "the usual pass" → quality + consistency sweep of the current file:
                     no behaviour change, minimal diff, show me proof.
- "blast radius"   → before editing a shared file, list every file that
                     references it; report the list before touching anything.
- "second opinion" → re-run this with a different agent and diff the two.
- "orient first"   → read the rules file + README, summarise state, then stop.

Now "okay, ship it" is a five-step release. "Give me the blast radius first" is a safety check I'd otherwise skip. The vocabulary becomes muscle memory, and because it's written down once, every agent in the fleet speaks it.

§4The one-word turns

A surprising share of my prompts are a single word — "yes", "continue", "go", "ship". They work because the surrounding context makes the next step obvious, and the house rules say so explicitly:

context rule

# A bare "yes" / "continue" / "go" / "do it" means: proceed with the
# obvious next step. Don't re-confirm, don't restate the options — act.

Combined with dictation, this is what "conversational" development actually feels like: you talk, it works, you glance at the diff, you say "yes," it ships. The keyboard becomes the exception, not the default.