WavRead
Your agent cannot hear your mix. This is how it reads it.
WavRead measures your track on your own machine and writes what it measured as text. The WavRead skill is the instruction file that teaches any AI agent how to read that text — what was measured, what was only estimated, and how to connect the evidence to your DAW, setup and creative intent.
npx skills add EZmannBuilds/wavread-skill
No CLI? Download the file, or read it on the skill repo.
What a reading looks like
The skill holds an agent to the same three layers WavRead uses itself. The measurement is stated flatly. What it might mean is offered as a possibility. What to listen for is handed back to you, because you are the one who can hear it.
Low-mid crowding where the arrangement is densest.
The arrangement losing definition in the loudest stretch, 0:56–1:36.
Why a skill at all
An agent handed a WavRead analysis with no instructions will do predictable things wrong. It will treat an estimated key as a fact. It will read a section that was never analyzed as a section where nothing was found. It will compare your track to a loudness target nobody asked for. It will read a machine lyric transcription back to you as though you wrote it.
The skill fixes each of those with a rule, and each rule came from something the tool actually did.
Four things the skill provides
Read your analysis correctly
Provenance first, so it knows which measurements exist before it discusses any of them. Every claim labelled by where it came from: measured, estimated, user-entered, compared, or the agent's own inference. An estimate never gets quietly promoted to a fact.
Tell you what to fix first
Priorities adapt to your intended sound, experience, production stage and delivery needs. Each item names the measurement, where in the track it is, and the listening check that confirms it.
Help you run WavRead
Analyzing a bounce, bringing your own stems, keeping versions comparable, and working out whether a surprising result is your mix or the app. From 1.5.6 WavRead carries its own MCP helper inside the app. The app does not show it; add it to Claude Desktop, Claude Code or Codex by its path, given in MCP.md.
Understand your intent
It asks which DAW and setup you use, what the lyrics mean to you, which imagery or feeling you want, and what should stay intact. It compares versions against those goals.
Three steps
| Analyse | Drop your bounce or your stems into WavRead. Everything runs on your machine. |
|---|---|
| Export | WavRead writes a Track Document of pure measurements and a Feedback Report of its own findings. Both are plain text. |
| Hand it over | Give the export to your agent. Ask your actual question. |
What it will not do
WavRead observes, measures, explains, compares and documents. Your DAW edits the audio. Your assistant adapts its guidance to the tools and experience you bring.
| No score or grade | There is no number at the end that says how good your mix is, because there isn't one. |
|---|---|
| No universal targets | There is no correct loudness, no correct spectral tilt, no correct width. A reference track is a comparison, never a destination. |
| Useful starting points | Your assistant can suggest DAW steps and processor experiments grounded in your setup, intent and listening checks. |
| No invented numbers | If a measurement isn't in the analysis, the answer is that it isn't there. |
| No reading your lyrics back to you | Machine transcription is the least reliable thing in the pipeline, and the skill forbids interpreting it. Supply your own words and it uses those. |
| No diagnosing you from your song | Measurements describe a recording. They do not describe the person who made it. |
Your audio still never leaves your machine
WavRead analyzes locally. No sign-in inside the app, no upload, no cloud processing.
The agent handoff does not change that. You send text, not audio — the analysis WavRead wrote, and only if you choose to. The skill tells the agent to keep analysis local and share only the context you choose.
Get the skill
npx skills add EZmannBuilds/wavread-skill
Installs the skill where your assistant will find it. Update it later with
npx skills update.
Works with any assistant that reads a skill or instruction file. If yours does not, download SKILL.md and paste it at the top of the conversation with your analysis underneath. The main workflow is in that file; keep the linked references alongside it for app operation and field details.
Two reference files go deeper, and an agent loads them only when it needs them: reading-sections.md, the trap hiding in each section of a document, and operating-wavread.md, how to run WavRead itself. The install command brings all three.
Questions
Do I need an AI subscription to use WavRead?
No. WavRead generates its own Feedback Report locally, with no agent involved. The agent handoff is an optional deeper layer.
Which assistants does it work with?
Any of them. The skill is a Markdown file and the analysis is plain text.
Will the agent be able to hear my song?
A text export supplies measurements, not listening access. If your assistant has audio tools and you choose to use them, it can add listening observations while keeping them distinct from measured facts.
Does this replace a mix engineer?
No. It reads measurements and explains them. Every judgment call is still yours.