Voice for coding that doesn’t suck: brainstorm vs dictation workflows for Claude Code, Codex, and OpenCode

A practical voice workflow for phone-sized moments: when to brainstorm vs dictate, how to keep context, and how to keep voice providers on a privacy leash.

Based on
CLI
0.2.17
Preview ref
a2c45c8

Voice is not about “coding by talking.”

It helps in the moments when typing is the bottleneck:

  • you’re on your phone
  • you’re walking
  • you have 30 seconds

And you still want to keep a Claude Code, Codex or OpenCode session moving.

The trick is that “voice” is two completely different workflows that get confused:

  • brainstorming (talk first, then turn it into a clean instruction)
  • dictation (talk to produce exact text/code)

If you mix them up, voice feels like a gimmick.

If you separate them, it becomes genuinely useful.

TL;DR

  1. Pick the right mode:
    • Brainstorm when you want to explore and then summarize into a crisp next step.
    • Dictate when you need precise text (but expect more friction).
  2. For brainstorm mode, the key habit is: ask for a summary and a 3-step plan before you send anything into the real session.
  3. For dictation mode, the key habit is: confirm formatting before you send.
  4. If you care about privacy, treat voice providers as external services and default to not sharing file paths/tool arguments.
  5. Happier is one option that helps because voice can target a real session (not just a floating mic), and it has explicit privacy controls.

The problem

Voice fails in coding workflows for predictable reasons:

  • you ramble
  • you lose decisions
  • you send half-formed instructions into the session
  • you accidentally expose too much context to a voice provider

So the goal isn’t “voice everywhere.”

It’s:

make voice useful in the few moments where it’s better than typing.

Common approaches and their trade-offs

  • Phone dictation straight into the message box: fast, but error-prone.
  • Voice notes in another app: safer, but now you have a copy/paste workflow.
  • Hands-free ‘assistant’ voice apps: fun demos, but hard to keep anchored to a real session.

Happier, an open-source app for running and orchestrating coding agents cross-device, gives voice an explicit target session and a clear privacy model.

Brainstorm mode (what actually works)

Brainstorm mode is not “send everything you say into the session.”

It’s a two-step loop:

  1. talk freely
  2. ask for a summary + the exact next message to send

Useful prompts:

  • “Summarize what I just said into a 3-step plan.”
  • “Propose the exact message I should send next.”
  • “Ask me one clarifying question before you act.”

Dictation mode (when you really need precision)

Dictation mode is for:

  • short instructions
  • commit messages
  • tight, formatted text

The failure mode is formatting.

So always add:

  • “Repeat back what you heard as text before sending.”
  • “Use bullet points.”
  • “Don’t invent file paths.”

Privacy: treat voice providers as an external boundary

Voice providers are powerful, but they are another trust boundary.

So a good default is:

  • do not share file paths
  • do not share tool arguments

…and only relax that when you understand what’s being sent.

The Happier workflow (voice modes + target session + privacy toggles)

Happier supports:

  • Happier Voice (managed realtime)
  • Use my ElevenLabs
  • Local voice

Voice keeps a hidden voice conversation transcript, but actions target a real session.

Privacy toggles include:

  • share recent messages
  • share file paths (off by default)
  • share tool arguments (off by default)
mobilevoiceclaudecodexopencodeworkflows
Last updated: 2026-10-26