Skip to content
Get packs

Guide · Agent skill

How to use the Audio Cleanup Editor agent skill (AudioEditor)

Updated · written with AI from the project's own documentation and checked against our catalogue data

AudioEditor is an agent skill, a set of instructions and scripts that an AI agent such as Claude loads when it is relevant to a request. Once installed, it lets the agent clean a spoken recording: it transcribes the audio, decides which parts to cut, and makes the cuts with ffmpeg. This guide is for podcasters, narrators and anyone with a recording full of ums, stutters or clicks, and it covers installation, how to ask for the work in plain language, and the limits you should know before trusting the output.

Audio Cleanup Editordanielmiessler/LifeOS · MIT

Quick start

  1. 1Install the tools. Install ffmpeg and insanely-fast-whisper, and make sure Bun is available.
  2. 2Add the skill. Unzip the downloaded AudioEditor folder into ~/.claude/skills/ or a project's .claude/skills/, or upload the ZIP under Settings → Capabilities → Skills.
  3. 3Ask for a preview. Request a preview of the edits on your recording so you can see the proposed cuts without any change to the file.
  4. 4Review the edit list. Check the proposed cuts against what you remember saying, and ask for changes if a deletion looks wrong.
  5. 5Run the cleanup. Ask the agent to apply the cleanup. The edited file appears beside the original as <name>_edited.
  6. 6Listen to the result. Play the output against the original, focusing on each cut and any pauses you meant to keep.

What does the AudioEditor skill make the agent do?

The skill handles recorded speech. Whisper produces a word-level transcript with timestamps. Claude then labels each segment as KEEP or as one of several cut types: filler words such as 'um' and 'uh', false starts, stutters, spoken edit markers, and dead air. The skill is designed to tell rhetorical emphasis apart from accidental repetition.

Edits are executed with ffmpeg. Each edit point gets a 40ms qsin crossfade, and extracted room tone fills gaps so the audio does not drop to digital silence. The output is a cleaned MP3 or WAV file.

Four separate tools handle narrower problems: BreathTame softens loud breaths and mouth clicks, GateScan detects noise-gate ticking, GateRepair fixes it, and LoudnessLock measures or sets delivery loudness to -14 LUFS and -1 dBTP, the YouTube standard.

How do you install the AudioEditor skill?

Download the skill ZIP from this directory page and unzip it. Put the AudioEditor folder in one of these locations: ~/.claude/skills/ for all your projects, or a project's .claude/skills/ folder for one project only.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

Before the first run, install the local tools. The setup notes list Homebrew commands for ffmpeg and pipx for insanely-fast-whisper.

  • Put AudioEditor in ~/.claude/skills/ (global) or .claude/skills/ (per project)
  • Or upload the ZIP in the Claude apps under Settings → Capabilities → Skills
brew install ffmpeg pipx install insanely-fast-whisper

What do you need before it runs?

The skill needs the Bun runtime to run its TypeScript tools, ffmpeg and ffprobe for every cut and measurement, and Whisper. The Whisper step is set up as insanely-fast-whisper running on Apple MPS. If insanely-fast-whisper is not on your PATH, the transcription tool falls back to the standard whisper command (pip install openai-whisper).

Claude classification runs through the LifeOS Inference tool. The optional Cleanvoice polish step needs a CLEANVOICE_API_KEY from cleanvoice.ai, and you only need it if you ask for polishing.

  • Bun runtime
  • ffmpeg and ffprobe
  • insanely-fast-whisper (or openai-whisper as fallback)
  • LifeOS Inference tool for Claude classification
  • Optional: CLEANVOICE_API_KEY for the cloud polish step

How do you trigger the skill with natural requests?

You do not need to name the tools. Describe the problem in plain language and the agent should route the request to the right intent. Phrases that match the skill include 'clean audio', 'edit audio', 'remove filler words', 'clean podcast', 'remove ums', 'cut dead air' and 'polish audio'. Standalone problems such as 'ticking audio', 'clicking audio', 'gate artifacts', 'loud breaths', 'mouth clicks' and 'tame the breaths' also route to it.

Modes change how strict the cleaning is. --preview lists proposed edits without changing the audio. --aggressive uses tighter thresholds. --polish adds the Cleanvoice pass for mouth sounds and normalization.

  • 'clean up the audio on this podcast file' runs the full pipeline and produces a cleaned file
  • 'show me what edits you'd make to this recording' runs a preview and changes nothing
  • 'aggressively clean this audio and polish it' uses tighter thresholds and then the Cleanvoice pass
  • 'my recording has ticking, fix the gate artifacts' runs GateScan and then GateRepair

What files does it produce and how do you check the result?

The edited file is written as <original-name>_edited.<ext> in the same folder as the input. The report gives the number of edits, the total time removed, the original and edited durations, and the paths to the transcript, edits JSON and audio.

Passing the tools does not mean the audio is good. Listen to the output against the original, especially at each cut. The skill itself says that a clean scan is not the same as an audibly clean result.

  • Output: <original-name>_edited.<ext> beside the input file
  • Check the preview list first, then run the edit
  • Listen to every cut before publishing

What are the safety notes for scripts and network use?

The skill runs scripts on your machine and uses the network. Claude classification goes through your Claude setup. If you enable the optional Cleanvoice polish, your audio is uploaded to an external service, so do not use it for sensitive recordings unless you accept that. Previews modify nothing, ambiguous file matches are questioned, and cloud polish runs only when you ask for it.

  • Cloud polish uploads audio to Cleanvoice; skip it for sensitive material
  • Recorder noise gates or macOS Voice Isolation can cause ticks; remove gate filters from the mic chain where you can

What people use it for

  • Removing ums, uhs and stutters from a podcast episode before a human edit
  • Cutting long dead air from narration recordings
  • Repairing clicks or ticks caused by a noise gate or Voice Isolation
  • Softening loud breaths and mouth clicks in voice-over
  • Setting a finished recording to -14 LUFS for YouTube-style delivery

Frequently asked questions

Is the AudioEditor skill free?

The source material in this listing does not state a price or licence terms beyond the MIT licence shown for the item, so check the licence file in the repository before you use it commercially. The local tools are installed on your machine, while Claude classification uses your Claude setup and the optional Cleanvoice step needs its own API key.

What do I need to install to use AudioEditor?

You need the Bun runtime, ffmpeg with ffprobe, and a Whisper tool such as insanely-fast-whisper. The setup notes give brew install ffmpeg and pipx install insanely-fast-whisper as the install commands. The cloud polish step also needs a CLEANVOICE_API_KEY, but it is optional.

Can AudioEditor remove video edits or edit video?

No. The skill explicitly excludes video composition and points that work to a separate video skill. It is for spoken audio recordings.

Is AudioEditor good enough for publishing a tight podcast cut?

The skill notes that a meter-clean result is not the same as ear-clean. For publication-ready tight cutting, a person should listen to the output and adjust it. Treat the skill as a fast first pass.

Why does my audio have clicks or ticks after cleaning?

The skill's notes say ticks are often caused by noise gates in the recorder's filter chain or macOS Voice Isolation, which truncate audio to digital zero. Ask the agent to run GateScan to confirm the defect, then GateRepair with --finalize to fix it. Removing the gate filter at capture is the lasting fix.

Does AudioEditor send my recordings to the internet?

The skill uses the network for Claude classification, and the optional Cleanvoice polish step uploads your audio to an external service. Without the polish step, your audio is not sent to Cleanvoice, but check the setup notes before using sensitive recordings.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

AudioEditor suits podcasters, narrators and LifeOS users who want a repeatable first pass at removing fillers, stutters, dead air and gate-related clicks. Pick something else if you need video editing or publication-ready tight cutting without a human listening, or if your recordings must never leave your machine.

Download the skill