No new tools in this release. No new CLI commands. All 19 tools work exactly as they did in v1.2.4.
What changed is the five domain expert skills, and they changed a lot.
The problem with generic advice
The old skills gave reasonable advice. That was the problem. Reasonable advice is what you get from a search result, and it doesn’t survive contact with an actual mix.
“The vocal is buried” has maybe six causes. Masking from a guitar sitting in the same octave. A compressor pulling the vocal down every time the snare hits. A reverb send smearing the consonants. Level automation that was right for the verse and wrong for the chorus. The old skills would pick one, suggest it, and move on. If it was wrong, you were back where you started.
The rewritten skills diagnose first. They check masking, check gain reduction, check whether the problem moves with the section, and only then commit to a fix. If the first fix doesn’t measure better, there’s a fallback and a reason to prefer it.
Diagnose, try, verify, fall back. That loop is the same in all five skills now.
Mix Engineer
The buried vocal sequence above is the headline, but the more useful change is three-way frequency conflict resolution. Two elements fighting is easy. Three is where people start cutting holes in everything and end up with a thin mix. The skill now assigns each element a home band and resolves outward from there instead of treating every pair independently.
It also knows that EQ and compression interact. If you boost 3 kHz on a track that’s already hitting a compressor, you’ve changed how hard that compressor works. The skill re-checks the threshold after tonal moves instead of treating them as separate steps.
Per-section processing is new too. A verse with four elements and a chorus with ten are not the same mixing problem, and one static setting won’t serve both.
Mastering Engineer
The send-back rule is now specific: 6 dB of corrective work is the threshold for handing a mix back, and that applies to corrective moves only. Enhancement and album-level consistency don’t count toward it. Vague “this needs more work” feedback helps nobody.
There’s a proper decision tree for minimum-phase versus linear-phase EQ, guidance on compressor and limiter attack interaction when you’re trying to keep transient punch, and a rule to skip parallel compression entirely on material that’s already squashed.
Final delivery now has pass/fail QC thresholds rather than a general sense that it’s probably fine.
Effects Engineer
The big idea here is that context beats preset. Before adding reverb, the skill assesses what ambiance the recording already has. A dry studio vocal is a blank canvas. A room-mic’d take already has space in it, and piling more on top trades clarity for nothing.
Extreme width now goes through a dual-bus approach: a wide bus for stereo playback and a mono-safe path alongside it, tagged explicitly so nothing collapses badly on a club system.
Chain order got real reasoning attached. Distortion into reverb and reverb into distortion are different effects, not stylistic preferences. Transients hitting distortion first lose definition as harmonics fuse together.
Audio Diagnostician
This one needed to stop crying wolf. It now tries to tell technical faults from deliberate choices before flagging anything. Tape hiss on a lo-fi record is not a defect. A guitar tone with deliberate grit is not clipping.
Noise is calculated per section now rather than across the whole file, with the worst case reported, because a quiet intro and a loud chorus have very different noise floors and averaging them tells you nothing.
Every finding carries a confidence prefix — Definite, Likely, or Possible — so you know which ones to trust and which to go listen to.
Session Architect
Prep triage for when you’re short on time: what has to happen before you can start, and what can wait until you’re already mixing.
Session scaling now keys off instrument density rather than raw stem count. Twenty stems of a sparse folk arrangement and twenty stems of a wall-of-guitars record need different routing.
There’s also a procedure for adding tracks mid-session without the bus structure falling apart, which is what actually happens on real sessions.
Getting it
uv tool install "phantom-audio[all]" --python 3.13 --forceThe skills ship with the Claude Code plugin, so if you’re using Phantom through Claude Code you get the new behaviour by updating.
Ask it something hard and see if the answer is more specific than it used to be. That’s the whole test.