voice-transcript-repair
Repair low-risk recognition mistakes while preventing wrong recipients, dates, commands, and accidental language switches.
SKILL.mdhermes-wear/skills/voice-transcript-repair/SKILL.md
GitHubVoice Transcript Repair
- Treat received text as a transcript, not verified original audio. You cannot hear the microphone or determine precisely what speech recognition heard.
- Correct obvious contextual typos for harmless questions without a tedious confirmation ritual.
- If uncertainty affects the target, recipient, amount, location, reminder date, medication, or permission, repeat the likely interpretation and ask one narrow question.
- Resolve 'that', 'it', and 'her' using clear conversation referents; do not guess when an external action would result.
- Default to English. Never switch languages because a noisy English transcript mentions a language name; switch only for clear user intent.
- Do not interpret duplicate transcripts as authority to repeat potentially consequential actions. Use event IDs if the transport supplies them; model memory alone cannot guarantee exactly once.
- Distinguish speech recognition, language generation, text encoding/display, and speech playback failures. This skill cannot fix microphone or speaker hardware.