Guide · Production
A concept album with AI: from research to booklet
A concept album is the opposite of what music generators are built for: they deliver single songs, fast and unconnected. An album that tells a story therefore does not come out of the generator but is made before and after it. We have made two – one about the Viking age, one about Al-Andalus. This is the workflow that went much faster the second time.
1. Break the theme into characters
A theme like “537 years of Moorish Spain” is not a song. Twelve people who lived through that time are twelve songs: the general on the day of the crossing, the musician at court, the poet, the philosopher, the last king at his farewell. Each character gets a perspective, a year and a feeling. That is the tracklist – before a single line of lyrics exists.
Why characters and not events: events turn into history lessons. Characters turn into songs, because songs work in the first person.
2. A research sheet before the first prompt
One page per character: who they were, what is documented, what is legend, which places, objects and words belong to their world. The generators know astonishingly much, but they invent just as readily – without a research sheet an instrument ends up in the lyrics that only existed 300 years later. The sheet is also the source for the booklet, so no time is lost.
3. Style settings per track, not per album
An album needs a common sound and twelve different songs. You get both with two levels of settings:
- Album-wide: tonal material (for Al-Andalus Phrygian-dominant scales, for the Vikings Dorian and drone), instrument palette, language, vocal character.
- Per track: tempo, mood, one lead instrument, one form (ballad, march, dance, lament). The variation lives here – not in the basic sound.
These settings precede the lyrics in the prompt and change across the album only where the dramaturgy calls for it.
4. Lyrics: short, concrete, singable
Historical lyrics tempt you into long lines and foreign words. Both fail in the generator: syllables get swallowed, names mangled. What works: short lines, clear vowels at line endings, proper names at most once per verse and only those pronounceable in the target language. Several runs per song; the best one will not be the first.
5. The tracklist as dramaturgy
Chronological order is the starting point, not the result. An album needs an opener with pull, a middle section alternating loud and quiet, a climax around tracks eight to ten and a closer that fades out quietly. If the chronology doesn’t provide that, it may yield – a retrospective as the last track explains itself.
6. The booklet: the part listeners keep
For an AI album the booklet is what no generator can deliver: context. One page per song with the character, the year, three paragraphs of history and the full lyrics; plus map, timeline and a foreword that says honestly how the album came about. Bilingual if the music is heard bilingually. As a PDF on the website, as an image gallery for the phone. The booklet is the reason a listener perceives the album as a work and not as twelve files.
7. Artwork from the same world
Cover and illustrations per track come from the research sheet: the same places, objects and lighting moods that appear in the lyrics. One colour world for the album (terracotta and gold for Al-Andalus, iron and fog for the Vikings) that repeats in booklet, Canvas and clips.
Time required, honestly
Research and characters: two to three days. Lyrics and generation: one to two weeks, because every song needs several runs. Booklet and artwork: one week. So around four weeks for a twelve-track album – more than twelve single songs, but a result that lasts for years.
The generator turns a prompt into a song. Turning twelve songs into an album remains handwork – and that handwork is exactly what listeners hear.
All guides are deliberately kept general: which models and tools run on your machine depends on your hardware. Questions or corrections? Write to us.