Turn a music prompt into conditions you can hear

What this chapter helps you decide

This chapter combines official technical material with a proposed production exercise. Generation, performance, pricing and listening comparisons have not been run. It is a supplement to the creative production course and can also be read independently.

• Translate an abstract preference into audible conditions.
• Compare structure separately from timbre.
• Keep unexecuted evaluation fields empty rather than inventing measurements.

1. 1Use and three requirements
2. 2Baseline prompt
3. 3Change one condition

1. 1Authorized generation
2. 2Full-duration listening
3. 3Timestamped reason

1. 1No run
2. 2null result fields
3. 3No winner or score

This is an original conceptual workflow. The arrows show review order, not a product’s hidden architecture; a failure sends the work back to the responsible stage.

Follow an adjective with an audible clue

“Calm” covers too many possible results. Convert it to conditions such as a slow steady beat, soft keys, restrained bass and no singing. Describe instruments, pulse, texture and density instead of using an artist’s name or a real singer’s voice as a shortcut. An acceptance criterion should be observable: for example, narration remains intelligible when the bass enters.

Separate prose from supported settings

Writing “30 seconds” or “90 BPM” in prose does not guarantee exact duration or tempo. Use a supported length or instrumental setting when the chosen model exposes it, then inspect the output. Keep lyrics separate from a musical description in services with a lyrics field. Record an unsupported setting as unsupported; adding its name to the prompt cannot create the missing control.

Change only one condition from the baseline

If A includes light percussion and B excludes percussion, the proposed difference is clear. Changing instruments, speed and structure together hides the reason for a result. Generation can vary under the same conditions, so one output pair cannot establish a general model ranking. A seed alone also does not fix the model version and execution environment. This chapter defines an unexecuted comparison plan, not a completed experiment.

Listen to the ending as well as the beginning

A promising opening can be followed by unwanted silence or a cut-off ending. Review fit to purpose, development, defects and closure; write “00:12: bass overlaps narration” instead of “bad mix.” The MusicGen model card’s limitations include unrealistic vocals and occasional collapse into silence. This is a provider disclosure, not a defect observed in a run for this course. The exact official source remains in the source provenance; its public link is omitted under the site’s publication policy.

Separate comparison level from the finished work

Use listening copies with comparable playback levels to reduce the influence of loudness on preference. Preserve each selected source file unchanged. Check whether the main subject remains clear on small speakers and in mono as well as on headphones. This is a proposed review procedure; no listening panel, score, win rate or measured loudness has been collected here.

Record the work needed to obtain one usable cue

Keep generation time, editing time, failures, credits and accepted candidates in separate fields. A monthly subscription price is not the generation cost of one accepted song. Before execution, samples, latency, cost and loudness values are null. Use not-run when no run occurred and not-collected when a run occurred but a value was not acquired. Neither means zero. Hypothetical scores belong in a separately labelled example, never in the results ledger.

Use and revise the production record

The input is a brief naming the use, duration and destination, plus materials whose usage conditions can be checked. The outputs are the production note, review sheet and the exercise deliverable. Keep the selected original separate from the editing and release copies.

1. Write the use and three requirements.
2. Separate instruments, tempo feel, density, structure and excluded elements.
3. Save a baseline and select just one item to change.
4. After any authorized generation, listen to the whole piece and annotate time positions.
5. Record the accepted version, rejection reasons and edits.

Symptom	Likely cause to investigate	Revision
It is unclear which instruction caused a change	Several conditions changed together.	Return to the baseline and change only one condition.
The opening sounded good but the cue does not fit the video	The full duration and connections were not reviewed.	Listen to the whole cue and its transitions.

Prepare a listening plan with one changed condition

Write A/B prompts for the same explainer, differing only in the presence of percussion. Do not generate audio. Prepare four review criteria and empty measurement fields.

Deliverable: The two instructions, scoring criteria and an empty measurement sheet.

Completion criterion: The prompt difference is one condition and the unexecuted fields contain no fabricated numbers.

Execution status: not-run. This chapter contains no generated audio, timing log, listening result or completed publication.

Review checklist

• Prompt wishes and output settings are distinct.
• The instructions do not request voice imitation or reuse of existing lyrics.
• Measurements and predictions occupy separate fields.

Two unexecuted prompt templates

These model-agnostic audio templates are source material for the exercise. Their output is an instruction, not a generated recording. Keep their IDs when reusing them. Text placeholders do not guarantee support for duration, instrumental controls or seamless looping.

music-narration-bed

Create an original instrumental background cue for {{use}}. Use {{instruments}}, a steady {{tempo}} feel, and sparse arrangement. Keep the melody understated so spoken narration remains clear. Start gently, add only a small change in texture midway, and end with a short natural decay. No vocals, spoken words, sudden impacts, or dramatic drops.

Variable	Required	Default example
use	true	a calm learning-app introduction
instruments	true	soft electric piano and restrained warm bass
tempo	true	slow

Create an original instrumental background cue for a calm learning-app introduction. Use soft electric piano and restrained warm bass, a steady slow feel, and sparse arrangement. Keep the melody understated so spoken narration remains clear. Start gently, add only a small change in texture midway, and end with a short natural decay. No vocals, spoken words, sudden impacts, or dramatic drops.

music-loop-sketch

Create an original instrumental loop candidate for {{scene}}. Use {{instruments}} with a steady pulse, restrained dynamics, and a repeating short motif. Keep the texture consistent and avoid an introductory buildup or a final cadence. No vocals or sudden fills. This is a draft to be edited and checked for a seamless loop.

Variable	Required	Default example
scene	true	a quiet puzzle-game menu
instruments	true	muted mallet tones and a soft synth pad

Create an original instrumental loop candidate for a quiet puzzle-game menu. Use muted mallet tones and a soft synth pad with a steady pulse, restrained dynamics, and a repeating short motif. Keep the texture consistent and avoid an introductory buildup or a final cadence. No vocals or sudden fills. This is a draft to be edited and checked for a seamless loop.

For music-narration-bed, verify the service’s actual length and instrumental controls and avoid requesting imitation of a song or artist. For music-loop-sketch, check the end points by listening and by waveform; choose the loop length, bar alignment and treatment of tails during editing. A seamless connection is not guaranteed by the prompt. For A/B planning, append a single percussion condition to a saved baseline and keep every other instruction identical.

{
  "status": "not-run",
  "model": null,
  "modelVersion": null,
  "generationLatencySeconds": null,
  "editingMinutes": null,
  "costAmount": null,
  "costCurrency": null,
  "creditUse": null,
  "integratedLufs": null,
  "truePeakDbtp": null,
  "sampleUrl": null
}