AI Music Generation in Practice: A Complete Workflow from Prompt to Final Track

Making background music with AI has a lower barrier than you'd expect—but also more pitfalls. To score a demo video for his side project, one developer spent several weekends testing various text-to-music tools and ultimately distilled a reusable workflow. His key finding: the quality of AI music doesn't depend on how advanced the tool is, but on whether you can write a clear "music brief."

This article builds on his hands-on experience to break down the complete process from prompt writing to post-editing, and fills in often-overlooked pieces like tool selection and copyright compliance.

What text-to-music actually does

You simply describe a song in natural language, and the model turns that description into audio. Different tools have very different positioning: some excel at full songs with vocals, while others focus on instrumental tracks such as film scores or looping background audio.

According to the original author's observations, result quality depends heavily on how specific the description is. Vague prompts produce generic tracks; specific prompts yield material you can actually use. This held true across every tool he tested.

Write a prompt that actually works

The original article provides a prompt template you can use directly:

Lo-fi hip hop instrumental, 75 BPM, warm electric piano, soft vinyl crackle, mellow drums, relaxed mood, no vocals, about 90 seconds, simple loop-friendly ending

This prompt covers six dimensions: style (lo-fi hip hop), mood (relaxed), tempo (75 BPM), instrumentation (electric piano, vinyl crackle, drums), vocals (explicitly marked no vocals), and structure (about 90 seconds, loop-friendly ending).

The author particularly emphasizes that BPM and "no vocals" deliver the biggest improvements. Without them, the model repeatedly generates vague vocals he didn't ask for. Naming two or three instruments keeps the sound more focused and avoids piling on elements.

If the video needs vocals, write the lyrics first, segment them with tags like verse and chorus, then paste them into the tool, and the model will structure the track accordingly.

BPM (Beats Per Minute) is a quantitative unit describing musical tempo, directly determining a track's rhythm and energy level. 75 BPM roughly corresponds to a slow hip-hop or R&B groove, suitable as background for talk-style videos or relaxed scenes; around 120 BPM is close to the standard tempo of pop dance music. For AI music generation models, BPM is one of the few numerical parameters that can be executed precisely, so including it in the prompt significantly narrows the result range and prevents the model from randomly swinging between fast and slow.

In text-to-music tools, "vocals" is typically a switch that must be explicitly declared. Many models default to attempting full songs with singing, because such samples make up a high proportion of training data. If the use case is purely instrumental background, you must specify no vocals or instrumental in the prompt, otherwise you'll easily get vague humming that both interferes with narration and is hard to remove in post. Conversely, when vocals are needed, provide segmented lyrics and use structure tags like verse, chorus, and bridge to guide the model's arrangement, otherwise the sung content and song form will be disconnected.

The five-step process from prompt to final track

The workflow the original author got working includes the following stages:

Step one, define the use case. Is it video background music, a song demo, or a game loop? The use case determines all subsequent parameters.

Step two, draft the prompt. Write in style, mood, BPM, instrumentation, vocals, and duration.

Step three, batch generate. Run at least three to five versions. The same prompt varies noticeably between rounds, so generating multiple times is a necessary cost.

Step four, screen. Listen once on speakers and once on headphones, skip the first ten seconds of "showing off," and focus on judging overall usability.

Step five, edit and finish. Trim the opening silence, add a fade-out, and lower the volume so it sits beneath the voiceover. Free tools like Audacity are sufficient.

If a track is half good and half bad, just cut and splice. Some people generate stems or variations and keep only the parts they like. The original author's experience: a bit of post-editing makes AI music "feel more human-made."

Tool selection: three hard criteria

The original article doesn't recommend specific products but provides a screening framework:

• Free quota: most tools have daily or monthly credit limits, so check upfront whether it's enough
• Vocal support: if you need vocals, rule out instrumental-only tools
• Commercial terms: commercial rights are usually tied to paid plans, and free versions often don't include them

The author cautions that prices and rules change frequently, so you should check each tool's official website directly rather than relying on old blog posts—including this one.

Commercial terms refer to the provisions in a tool's terms of service regarding whether generated content can be used in for-profit scenarios. A common pattern: free plans grant only personal, non-commercial use, while paid plans come with commercial usage rights; some tools also require attribution or restrict distribution channels. Because training data sources and rights ownership are disputed, terms vary widely across providers and are revised frequently, so judging commercial usability shouldn't rely on impressions—read the current version of the terms of service clause by clause and keep records of your account, prompts, and payments at generation time as evidence for future rights claims.

Copyright and ethics: the most easily skipped step

The original article's tone is most serious in this section: commercial rules depend on the tool and paid plan, and industry disputes with rights holders are also driving changes to terms. His list of safe habits includes: read the current terms before publishing or selling, confirm whether your plan includes commercial rights, don't ask the model to imitate a living artist's voice or a specific song, label AI assistance as required by the platform, and keep prompts and payment records.

The author explicitly states he is not a lawyer, and this section can only serve as general reference.

Four common mistakes

The original article's list of pitfalls: prompts too short ("make a happy song" only yields generic results); stacking too many styles (five styles mixed together is usually mush—pick one main style plus one accent); skipping editing (the first version is rarely the best); ignoring licenses (a great track you can't use means a wasted afternoon).

The essence of this method is treating AI music generation as a combination of "writing a requirements doc + post-production editing," not one-click magic. For independent creators who need steady output of background music, this framework's reusability is higher than any single tool.