Voice First Planning

• During active implementation (typing precise code instructions is faster)
• For short, well-defined commands ("fix the typo on line 42")
• When you are in a shared/quiet workspace without a private space to speak
• For editing or refining an already-written spec (just edit the text directly)

Common Mistakes

The Workflow

Step 1: Set up a speech-to-text tool

Pick one and install it:

Any tool that converts speech to editable text works. The key requirement is that you can paste the output into a terminal or editor.

Step 2: Speak your idea freely

Open your transcription tool and start talking. Cover:

• What you want to build and why it matters
• Who will use it and what their workflow looks like
• The constraints — what you cannot change, what is already decided
• Edge cases you are already thinking about
• Anything you are unsure about — name the uncertainty out loud

Do not organize your thoughts first. Do not outline. Just talk. Aim for 1-3 minutes of continuous speech. This typically produces 200-500 words of transcript, which is far more context than most people would type.

Step 3: Quick-scan the transcript for errors

Speech-to-text tools mangle technical terms. Before pasting, scan for:

• Variable/function names — "getUserById" might become "get user by ID" or worse
• Library names — "Playwright" might become "play right"
• Homophones — "route" vs "root", "cache" vs "cash"
• Jargon — domain-specific terms the model may not have in its vocabulary

Fix only the obviously wrong terms. Do not rewrite — the raw, spoken style is the point.

Tip: Use TMUX with VIM keybindings to quickly jump through the transcript and fix transcription errors before pasting. VIM's word-motion keys (w, b, cw) make surgical fixes fast.

Step 4: Paste into Claude Code with a framing prompt

Prepend a one-line instruction that tells Claude Code what to produce:

Structure this spoken transcript into a feature spec with requirements,
constraints, and open questions:

[paste transcript here]

Other useful framing prompts:

• "Extract the architectural decisions from this transcript and list the trade-offs for each."
• "Turn this into a task breakdown with estimated complexity for each item."
• "Identify the requirements, assumptions, and risks in this transcript."
• "Summarize the core idea in 2 sentences, then list all the details I mentioned."

Step 5: Iterate on the structured output

Claude Code will return a well-organized version of your rambling thoughts. Now you can:

• Remove sections that were tangents
• Add precision where the spoken version was vague
• Ask follow-up questions about parts the LLM flagged as ambiguous
• Feed the structured spec back into your planning workflow

This is where the real leverage appears: you went from a blank prompt to a structured spec in under 5 minutes, and it contains context you never would have typed.

Quick Reference

Key Principles

1. Typing makes you self-edit. Speaking does not. When you type, you delete and rephrase to sound "right." When you speak, you explain what you actually mean, including the messy context that matters most.
2. LLMs are excellent at extracting structure from rambling speech. You do not need to organize your thoughts before speaking. That is the LLM's job. Your job is to provide maximum raw signal.
3. Voice is for planning, not implementation. This technique is most effective in the ideation and spec-writing phase. Once you are deep in code, typed instructions give you the precision you need.
4. Fix transcription errors, but do not rewrite. A quick scan for mangled technical terms is essential. Rewriting the transcript into polished prose is counterproductive — you are just re-introducing the self-editing problem.
5. Always frame the paste. A raw transcript dump without a prompt like "structure this into a spec" forces the LLM to guess what you want. One sentence of framing turns a wall of text into useful output.

Attribution

Based on Josh's technique from the Coding Agents: AI Driven Dev Conference. Josh uses WhisperFlow to speak initial specs, noting that speaking gives more context than typing because people naturally self-edit when they type. He also recommends TMUX with VIM mode for quickly fixing transcription errors before feeding the text to an LLM.