What this skill produces is intent, not parameters.
The distinction matters:
- "Vocals 3 kHz +2 dB" is aparameter— only a DAW can execute it, and it's useless with a different backend
- "The vocal must remain the clearest element even in the densest chorus; what gives way is the pad, not the guitar" isintent— a human engineer can execute it, a DAW can execute it, a text backend can partially execute it, and it can be checked item by item at acceptance
This library stops at intent. EQ frequencies, compression ratios, and the mastering chain are the business of mixing; this package only provides mix_intent this intent.
Which sections to read by task
| Symptom / task | Read |
|---|---|
Filling in ARR-SPEC's mix_intent |
§2–§4, then §8 |
| Technically clean mix but "wrong" | §3.1 Emotion |
| Not sure whether to pursue balance | §4（Depth is about consistency, not balance） |
| Parts covering each other | §5 Masking (it's mostly the pad's problem, not the drums'） |
| To hand off to a mixing engineer | §2 Five domains + §8's handoff format |
| Judging how much a given backend can handle | §7 |
| Want to find reference tracks | §6 |
Boundaries: what this skill does not cover
| Not here | Where it belongs |
|---|---|
| EQ frequencies, compression ratios, reverb parameters, bus chains, mastering | Mixing's business; this package only provides mix_intent this intent (the fields handed off and how to read them:mc-workflow §5） |
| Frequency-band conflicts'rewrite-typesolutions (change register, reduce notes, alternate) | mc-texture-layering（solve it there first; only come here if it can't be solved） |
| Structural planning of foreground/midground/background | mc-texture-layering §2 |
| How the vocals should be sung | mc-vocal-direction |
| Concrete panning plans for harmonies/doubles | mc-vocal-direction §6.2 |
| How the timbre itself should be designed | mc-sound-design |
| Drum kitinternalleft-right and high-low placement of each piece | mc-rhythm-section §6 (this skill's soundstage map only definesthe drum kit as a wholesits) |
★ The boundary with
mc-texture-layering(repeated because it's the easiest to confuse): texture-layering's solution is "rewrite"; mix-intent's solution is "remix". What can be solved by rewriting should not be left to mixing.
1. Must-fill before starting
| Required | Question |
|---|---|
| ① Emotion | Say in one sentence what this mix must convey.Write this one first; it vetoes all later decisions |
| ② Soundstage map | Each instrument's left-right position and front-back position (§2) |
| ③ Protagonist sequence | Write section by section: who is the clearest element in this section |
| ④ Yielding relationships | Who yields to whom, and when (ducking_map） |
| ⑤ Reference tracks | At least one, and be able to state thespecific aspect of it you are referencing |
2. Five domains and the sound stage
The mixing process must involve 5 main domains (p0089)：
| # | Domain |
|---|---|
| 1 | Time |
| 2 | Frequency |
| 3 | Level |
| 4 | Stereo |
| 5 | Depth |
Stereo and depth together make up "Space" (p0089)。
Sound stage：
The mix should be presented within animagined sound stage, with each instrument having a certain **left-right (stereo)** and **front-back (depth)** position (p0089).
This is the skeleton of the handoff format—one sound stage diagram + one sentence for each of the five domains, and the mixing engineer can start work.
What cues do left-right and front-back each rely on（Howard & Angus）：
- What the pan pot provides isinteraural level difference(p0112); interaural level difference mainly works for high-frequency localization, interaural time difference mainly works for low frequencies, 700 Hz–2.8 kHz is the transition zone between the two cues, where localization ability is weakest(p0108). → The pan of low-frequency elements is basically ineffective; just center them. If you want an instrument to have a "clearly defined position," it needs components above 3 kHz. Failure condition: headphone monitoring (in-head localization, p0109).
- Front-back relies on the ratio of direct sound to reverberant sound and the time interval (p0261, p0264); delayed sound arriving within 30 ms is not heard separately, localization still follows the leading sound (precedence effect, p0110); an initial delay of 10–20 ms is a natural front-back interval, and 50 ms is an echo (p0346). → When writing depth, give only three levels: "near / mid / far"; do not write milliseconds.
Macro vs micro (p0089)：
| Targets | Example | |
|---|---|---|
| Macro mixing | Elements that affect the entire mix | The frequency balance of the mix |
| Micro mixing | Processing each instrument part individually | Whether the vocals sound natural |
This skill mainly writes macro—micro is the mixing engineer's craft and should not be rigidly specified by the arrangement specification sheet.
3. Four goals
Mixing has 4 basic goals (p0089):
| # | Goal |
|---|---|
| 1 | Mood |
| 2 | Balance |
| 3 | Definition |
| 4 | Interest |
When evaluating a mix, first look at its coherence and appeal across the main domains, then analyze how well it achieves each goal (p0089)。
3.1 ★ Mood: it has veto power
Mood refers to whether the mix can express theemotional factors (p0090)。
Two counterexamples, both very telling (p0090):
- Heavy compression, extreme EQ, noisy distortion, overly strong snare impact → ruins agentle jazz piece
- Sweet vocal reverb, overly consonant drums, soft guitar → ruins ananger-filled heavy metal piece
Note the second example:each item taken individually is a hallmark of a "good mix", but together they are wrong. Technically correct does not equal emotionally correct.
A mixing engineer who is only good at handling a certain musical style often gets poor results when encountering other styles (p0090)。
Significance for this library: A generative model is also a "mixing engineer who is only good at a certain style"—
it will pull any input toward the balance it has seen most often.
Therefore mix_intent.mood must be hard-coded, andplaced early in the prompt。
3.2 Clarity
Clarity represents whether the sound is clear and recognizable (p0090)。
On the spec sheet, this is §1③'sprotagonist sequence: state clearly, section by section, who must be recognizable.
Clarity is essentially contained within separation: whether an instrument is clear depends on whether it can be heard and identified (intelligibility), and also on whether it sounds natural; it has little to do with loudness (p0260). The criterion can be tested with two counterexamples: a piano recorded as if coming from the bottom of the sea, or a hi-hat missing its fundamental high-frequency components—in both cases the level may be normal, yet clarity has collapsed (p0260).
Level and clarity are often treated as simply proportional: "the higher the instrument's level, the higher the clarity" (p0100). But this holds only when the instrument's own frequencies are healthy; for an instrument with defective frequencies, raising the level only makes it louder, not clearer (p0100). When diagnosing "this part is not clear enough," first determine whether it is insufficient loudness or a hole in the frequencies, then decide whether the level should be pushed.
3.3 Interest
Interest often comes fromviolating balance(see §4.3), but violating balance is only one extreme form of interest; the mechanism of interest is more fundamental: the ear adapts, and if the sound stays unchanged for a long time, attention drops; only change can recapture the ear (p0091). Verse-chorus repetition, track doubling, and instrumentation changes are all essentially about sustaining this attention by creating change (p0091).
Mixing can layer another level of change on top of the arrangement structure; the criterion is "which section has it and which section does not," without involving specific parameters (p0092):
- A certain part is automated up only in specific sections
- A certain instrument appears only in specific sections (e.g., trumpet enters only on verse repeats)
- The same instrument gets different processing in different sections, letting brightness and darkness change with the section
- The chorus compresses the whole more aggressively, allowing bass distortion
- The snare gets different reverb in different sections
Style determines how far this should be taken: ambient music should keep people interested but not too attention-grabbing; trance relies on repeating the same theme, so changes actually need to be more finely crafted to avoid becoming tiresome (p0092).
Writing method: Write these change points into focus_by_section[]by section, and write imbalance-type ones into deliberate_imbalance[](§4.3); do not write them as a uniform state for the whole song.
4. Balance: where to seek balance in three places, and where not to in one
4.1 Three areas where balance is sought (p0090)
-
Frequency balance: A lack of mid-high frequencies makes the mix sound muddy and blurred, and the stereo image also seems farther away (p0090). The criterion is not to fill in to some specific value; once the listening impression falls into the three words "muddy, blurred, farther away," it means the mid-highs are lacking.
-
Stereo image balance: Supported by the two pillars of level balance and frequency balance, try to keep the instruments overall left-right symmetrical (e.g., guitar, shaker, backing vocals) (p0252). The "balance" mentioned here refers tothe overallPanning distribution does not require a single instrument to be centered; in fact, center is the position with the worst clarity (see §2 localization cues).
-
Level balance: here it meansrelative level, not absolute level. Mix evaluation asks "how loud is this instrument relative to the other instruments," not "how loud is this instrument by itself"; turning up the monitoring volume overall cannot solve a relative level problem (p0097). Relative level is usually determined by the instrument's importance at this moment and must be adjusted as the song progresses; for example, the level of a lead guitar solo section should be higher than that of an accompaniment section (p0098). Cost: if, to make relative levels tidy, all hits are squashed to the same level, the drum kit will sound like a "drum machine" and lose the timbral variation brought by the original dynamics; level balance does not equal level uniformity (p0388).
4.2 ★ Depth should be "consistent," not "balanced"
Depth issues should be made consistent, not balanced (p0090)。
Example: ifthe kick drum is panned so close it is in your face, while the snare is as far away as the neighbor's house, it will make the entire drum kit produce an extremely distorted panning layout in depth (p0090)。
This can easily be misread as "depth should also be centered." It is not— depth can be overall very close or overall very far, but within the same source group it must be consistent.
Self-check action: list the depth values of instruments by group (drum group, keyboard group, string group), within-group differences greater than between-group differences → wrong.
Physically, "the same group" means sharing the same set of early reflections and reverb tail: the early reflections at each panning position can differ, but the decay rate of the tail must be the same (Howard & Angus p0348–0350).
4.3 Two situations where imbalance is acceptable (p0090)
- Create a creative sound effect
- Make the song interesting
Example: in Delgados' "The Past That Suits You Best," mixing engineer Dave Fridmann putthe panning of the drum kit and vocals entirely into one channel, creating an effect that is imbalanced in stereo but very charming (p0090).
Interesting imbalance is usually a momentary treatment—for example, doubling some low-frequency signals in a dance track's highlight section (p0090).
The phrase "usually momentary" is key: imbalance isan event, nota state. Imbalance across the whole song is an error; imbalance at one spot is a technique.
Writing method: imbalance must be written into mix_intent.deliberate_imbalance[],
withstart and end measuresandreason. Any imbalance without a written measure range is treated as a mistake.
5. Masking: why the problem is mostly in the pad rather than the drums
The core principle of masking: louder sounds mask quieter sounds; the frequency range of one instrument masks the frequency range of another instrument (p0038).
But what is truly useful is the followingtime dimensioncorollary:
- Percussive instruments (such as the kick drum) sound for a short time, so their masking effect is relatively small
- continuously sounding instruments (such as a sawtooth pad) fill the frequency spectrum and stereo space, so their masking effect is large (p0038)
- ★ Raising the level of a dense pad part causes more masking than raising the kick level; the pad masking the kick is a persistent, major problem (p0038)
Howard & Angus laid these out as mechanisms:
- Masking extends upward: low frequencies easily mask high frequencies, and the reverse is much weaker (p0239) → the pad's low frequencies don't just cover the kick, they cover everything above it, so "leave a gap in the pad's low end" is the first cut.
- Within the critical band where the masker lies, the masking threshold is only about 20 dB below the masker (p0240) → make way properly: being more than 20 dB quieter in the same frequency range counts as "making way"; being 3 dB quieter just makes it muddy.
- Temporal masking: within 100–200 ms after a sound stops, a quiet onset is still inaudible; the reverse (backward masking) is only about 10 ms and almost doesn't exist (p0241). The loudness of a short percussive sound itself is also discounted by its duration—the shorter a sound under 200 ms is, the quieter it seems (p0095). → When a vocal or hook enters right after a crash/fill, either leave ≥ 200 ms, or make its first note loud enough on its own. Self-check: find each lead element's entry point and see whether something louder has just ended within the previous 200 ms.
Corollary (usage in this library): When diagnosing "muddiness",check sustained sounds first, not percussion first. Intuition always points to "the drums are too loud", but by this principle the culprit is more likely that ever-sounding pad or underlying guitar.
The preferred solution is still upstream: let the pad leave a gap in the low-frequency range (rewrite,mc-texture-layering),
rather than using sidechain to push it down (remix).Sidechain is only used when there's no other solution.
6. Reference tracks: this library's calibration method
6.1 Six uses (p0051–0052)
| Use | Description |
|---|---|
| Imitate as a template | An excellent training method, but **"imitation is the greatest enemy of innovation"**; and differences in genre can make imitation ineffective |
| As a source of inspiration | Play itbefore starting themix, to inspire direction |
| A breakthrough when stuck | Study thedifferences |
| As a finished reference | After completion, compare frequency response and levels |
| Calibrate your ears | In an unfamiliar environment, play a familiar reference track first to get a feel for the speakers/room/seat |
| Evaluate speakers | Play works you're truly familiar with; it works better than the pop music the seller plays |
6.2 Selection and pitfalls
- Reference tracks shouldmatch the mix type(if you're mixing strings, find something with strings; if you're doing chill-out, find the same kind) (p0051)
- Save them inlossless format, not MP3 (p0051)
- It's best to includepre-selected sectionsfor quick comparison; you can addunmixed tracks (p0051)
- ★ Watch out for the effects of mastering (p0052): mastered reference tracks usually have denser dynamics (consistent relative levels + more compression), and their frequency characteristics may have been altered according tothe least ideal trackon the album
- ★ Align loudness before comparing: equal-loudness curves deform with sound pressure level; once the volume drops, low and high frequencies are suppressed relative to the mids, and the sound becomes thinner and darker (Howard & Angus p0091); a +10 dB sound pressure level roughly doubles loudness (p0094). The louder one almost always "sounds better"—without aligning loudness, you are comparing volume, not the mix.
This last point is especially important for this library: when reverse-engineering metrics from finished audio, the measured dynamic range ispost-mastering. So taking a reference track's dynamic value as a target will be systematically too tight— likewise, before measuring a reference track, first check its bitrate and high-frequency cutoff: lossy encoding shaves off high frequencies, and files of different bitrates cannot be compared crosswise.
7. What each backend can handle
| Backend | Can handle | Cannot handle |
|---|---|---|
| Suno / MiniMax | Descriptions of emotion and overall texture (warm、dry、wide chorus） |
soundstage map, ducking_map, per-section depth |
| YuE2 | Same as above, with finer section granularity | Precise panning and depth |
| MIDI / DAW | All of it—treat it directly as a direction sheet | — |
★ Rule: items marked in the backend profile as honors: none ,
when calculating the compliance ratemust be excluded and cannot be given 0 points.
Otherwise a backend that honestly declares "I cannot do panning" would score lower than one that pretends it can—
that would lead the scoring system in the wrong direction.
8. Correspondence with ARR-SPEC fields
| Field | Which section of this skill |
|---|---|
mix_intent.mood |
§3.1（One sentence, and placed near the front of the generation prompt） |
mix_intent.sound_stage[] |
§2 (pan + depth for each instrument) |
mix_intent.focus_by_section[] |
§3.2 protagonist sequence |
mix_intent.ducking_map[] |
§5 (who yields to whom, and when) |
mix_intent.deliberate_imbalance[] |
§4.3（Must include start and end measures + reason） |
mix_intent.reference_tracks[] |
§6（Must state which aspect of it is referenced） |
9. Must-pass checklist after writing
-
moodwritten,one sentence, and does not conflict with later decisions - Five domainseach has one sentence written, not just frequency
- Every instrument in the soundstage maphas left-right and front-backtwo values
- Instruments in the same group have consistent depth(within the drum kit, within the string section)
- Written section by sectionwho is the clearest element
- Every imbalance includesstart and end measures and reason, not a whole-song state
- When diagnosing "muddiness"checked sustained notes first, not blaming the drums first
- Within 200 ms before each protagonist's entrance point, no louder event has just ended (§5)
- Low-frequency elements are panned center; loudness is aligned when comparing reference tracks (§2, §6.2)
- Conflicts that can be solved at the composition layerhave already been
mc-texture-layeringresolved, not piled up here - At least one reference track,stating which aspect of it is referenced
- No EQ frequencies or compression ratios are written in this document (those are matters for mixing)
Appendix: Sources
- Roey Izhaki, 《Mixing Audio》 (The Mixing Guide)—§2, §3, §4, §5, §6 main content (6 the domains and goals involved in mixing, 6.1 goals of mixing, 2.x masking, 3.4 using reference tracks). Page numbers are those of the book. This library only takes the two parts "domains and goals" and "reference tracks"; the practical chapters on EQ/compression/reverb/mastering are matters of mixing and are not included in this package.
- Howard & Angus, Music Acoustics and Psychoacoustics(Acoustics and Psychoacoustics)— §2 localization cues and the precedence effect, §4.2 reverb tail, §5 masking mechanisms and temporal masking, §6.2 equal-loudness contours. Page numbers are those of the Chinese translation.
- **Kazuhiro Takeuchi, EDM Production**'s frequency balance and Lo-Mid management, volume balance of each part—likewise only the intent layer is taken, parameter organization is in reference.md。
- backend profile and
honorsdefinitions seemc-render-compile§6 andbackends.yaml。