Describe the job, not an imagined masterpiece
A brief such as “make a beautiful cinematic song” leaves almost every production decision open. Start instead with function: an eight-bar introduction that leaves room for spoken words, a sparse bass response to a chorus melody, or a percussion variation for the final refrain. Each describes a job inside a session. It also creates an audition test. If the generated introduction competes with the voice, it fails the brief even when its sound is impressive.
Write down what must not change. That might be the singer’s phrasing, the harmonic rhythm, a recorded drum feel or the silence before a downbeat. These protected elements prevent a tool’s most polished suggestion from quietly becoming a different song. For a small team, agree on those boundaries before generating. Otherwise one person may select for harmonic surprise while another selects for mix compatibility, and both will believe the tool is being inconsistent.
Separate musical instructions from file requirements
Musical instructions describe register, density, articulation, meter and development. File requirements describe duration, sample rate, separate parts and editability. A tool may follow one category better than the other. Requesting MIDI does not make an audio-only generator produce editable notes. Likewise, specifying a tempo does not guarantee a rigid beat grid. Check the product’s documented inputs and export options before building an entire plan around an instruction it cannot implement.
Format note: the MIDI Association’s overview explains MIDI as performance and control information, not recorded sound. This distinction matters when choosing between note-level revision and audio editing.
Use a compact specification you can read while listening. Include the part’s role, approximate length, tempo context, tonal centre if relevant, desired movement and exclusions. “Restrained, low-register two-bar bass figure; support the existing kick; leave beat four open; avoid fills” offers more actionable information than a stack of mood adjectives. Treat these instructions as requests rather than guarantees. Some systems interpret technical language loosely, so the audition remains the final check.
Choose inputs you can explain and use
A reference can communicate feel efficiently, but uploading a recording introduces separate questions about permission and the service’s handling of files. Prefer your own sketch or material you are authorised to upload. Do not assume access through a streaming subscription grants upload rights. Describe observable characteristics—dry room sound, a narrow melodic range, syncopated eighth notes—instead of asking for a living performer’s identity. That makes the brief more transferable and avoids unnecessary imitation.
Useful working references include the Ableton Live manual for clips and MIDI editing, the REAPER user guide for project and media handling, and the Audacity project for accessible audio inspection. These are workflow references, not endorsements of generation capabilities.
Run one comparison at a time
Generate a limited set, then change one important variable. If the first result is too busy, reduce density before changing the key, instrumentation and mood together. Keep the same audition passage and match playback levels. A louder candidate often feels more exciting even when it contributes less. Label versions by the question they answer: “less movement,” “higher register” or “shorter tail.” That is more useful than a folder full of files called “final.”
Our example brief is deliberately modest: “Create a short, sustained harmonic texture beneath this original vocal sketch. Keep attacks soft, avoid a lead melody, and leave the low register clear for bass.” Audition the result at the verse’s quietest line and the chorus entrance. Listen for harmonic clashes, distracting motion and reverb that masks consonants. Keep only the sections that help. A two-second texture can be a successful outcome; a complete generated song is not the required prize.
Leave a record for the next session
Save the written brief beside the source sketch, output and a brief decision note. Record the service, model label when available, generation date, settings and any seed it exposes. A seed is not a universal promise of reproduction: changes to a service or model can change behaviour. Preserve the actual audio or MIDI file rather than relying on a remote history. Note what you edited after generation so collaborators can distinguish the raw suggestion from your production work.
Stop when the question has been answered. If several attempts do not improve the track, play the part yourself, use a conventional instrument or leave the space empty. A useful brief is not a commitment to use AI. It is a way to keep a speculative tool inside a musical process whose standards you control.