---

name: voice-refine description: Transform verbose voice input into optimized Claude prompts allowed-tools: Read context: inherit agent: specialist

Voice Refine Skill

Transform verbose, stream-of-consciousness voice dictation into structured, token-efficient prompts for Claude Code.

When to Use

• Input from voice dictation (Wispr Flow, Superwhisper, macOS Dictation)
• Verbose text >150 words
• Contains filler words, repetitions, or tangents
• Natural speech patterns that need structure

Transformation Pipeline

Copy1. DEDUPE    → Remove repetitions and filler words
2. EXTRACT   → Identify core requirements and constraints
3. STRUCTURE → Organize into standard sections
4. COMPRESS  → Reduce to ~30% of original while preserving intent

Output Format

Copy## Contexte
[Project context, existing stack, relevant files]

## Objectif
[Single sentence: what needs to be built/changed]

## Contraintes
- [Constraint 1]
- [Constraint 2]
- [etc.]

## Output attendu
[Expected deliverables: files, format, tests]

Flags

Usage Examples

Basic Usage

Copy/voice-refine

Alors euh j'aimerais que tu m'aides à faire un truc, en fait j'ai une API
qui renvoie des données utilisateurs et je voudrais les afficher dans un
tableau React, mais attention il faut que ça soit paginé parce que y'a
beaucoup de données, genre des milliers d'utilisateurs, et aussi faudrait
pouvoir trier par nom ou par date d'inscription, ah et on utilise Tailwind
dans le projet donc faut que ça matche avec ça...

With Flags

Copy/voice-refine --direct --en

[voice input in any language → sends English prompt directly]

Compression Metrics

Integration with Voice Tools

Wispr Flow

1. Dictate with Cmd+Shift+Space
2. Paste into Claude Code
3. Run /voice-refine

Superwhisper

1. Record with hotkey
2. Text appears in active window
3. Run /voice-refine to structure

macOS Dictation

1. Fn Fn to start
2. Speak naturally
3. Run /voice-refine to clean up

What Gets Removed

• Filler words: "euh", "um", "like", "you know", "basically"
• Repetitions: same concept stated multiple ways
• Tangents: off-topic thoughts
• Hedging: "maybe", "I think", "probably" (unless relevant)
• Politeness padding: "please", "could you", "I'd like"

What Gets Preserved

• Technical requirements
• Constraints and limitations
• Context about existing code
• Expected output format
• Edge cases mentioned
• Business logic rules

See Also

• guide/ai-ecosystem.md - Voice-to-Text Tools section
• examples/before-after.md - Full transformation examples