The Complete Guide to AI Playlist Prompting

Master the Art of Creating Perfect Playlists with 7 Techniques Borrowed from Image Generation

Nano Banana Generated Image recalling the famous Abbey Road cover

Image generation prompting has matured into a sophisticated discipline. Tools like Midjourney, DALL-E, and Stable Diffusion have spawned extensive communities sharing techniques, vocabulary lists, and proven prompts. Playlist generation, while newer, can benefit enormously from these battle-tested approaches.

This guide translates the most effective image prompting techniques into their musical equivalents, giving you a systematic framework for creating exceptional AI-generated playlists.

Technique 1: LLM-Assisted Prompt Crafting

Image Parallel: Ask an LLM to describe an image similar to what you want, then refine it.

Nano Banana generated image

If you’re not a music critic, it can be difficult to articulate exactly what you want. Just as you might ask ChatGPT to “describe a cyberpunk cityscape at night” and then tweak the output for Midjourney, you can use LLMs to help craft playlist prompts.

The Reverse Description Method

Start with a reference: Think of a playlist, album, or song that captures part of what you want.

Ask the LLM to describe it:

Describe the mood, energy, and musical characteristics of "Song 2" by Blur.

Capture the vocabulary: The LLM might say: “exhilaratingly chaotic… captures a sense of youthful rebellion and a ‘don’t care’ attitude… underlying sense of irony and playfulness… follows a strict quiet-loud-quiet structure.”

Capture the musical characteristics: twin bass distortion, heavy drums, fuzzy electric guitars, the “Woo-Hoo” hook, distorted vocals that make the singer sound like he’s shouting over the noise of a crowded room, minimalist lyrics, fast and driving tempo (130 BPM).

Capture the Genre: grunge parody / indie rock.

Modify and combine: Now tweak it. Here are some examples:

Like "Song 2" by Blur, but more upbeat and with acoustic elements.

Indie rock with distorted guitars and the sense of irony and playfulness of "Song 2" by Blur, but with deeper narrative lyrics.

Notice that it’s useful to repeat the song name, as examples contain more information than even a long description can. Nevertheless, pointing out the key elements you want can better direct the search and also helps you discover new music.

Compare the results with what you get by asking only “10 songs like ‘Song 2’ by Blur” — you’ll generally get a more targeted list.

How is this different from Spotify’s suggestions? Spotify mainly uses collaborative filtering, which means it tries to guess what you might like based on what other users who listen to “Song 2” enjoy. This often provides good results, but it also tends to suggest the same songs repeatedly and leaves less control to the user.

Try it yourself: Song To Run is a free app that lets you experiment with these prompting techniques for running music. Chat to describe what you want, then load the suggestions directly into Spotify to see if they match your taste.

Ask the LLM to Craft the Prompt

Can you create a prompt to ask an LLM to find similar songs?

Testing with Gemini, I received three optional ways to phrase the prompt depending on how deep you want the LLM to go:

• Option 1: You want songs that match the feeling of a high-energy, ’90s alternative rock mosh pit.
• Option 2: You want to find songs based on technical characteristics (like the dual-bass distortion or the “woo-hoo” vocal hook).
• Option 3: You focus on why the song was written (a parody of grunge) to find other “accidental anthems” or satirical rock hits.

Technique 2: Quality Boosters

Image Parallel: Terms like “4K,” “highly detailed,” “trending on ArtStation,” and “masterpiece” signal quality expectations.

Playlist equivalents signal that you want carefully curated, high-quality selections rather than generic algorithmic output. To grasp why they might bring value remind that these attributes derive from the text corpus that was used to train the LLM, so you are not requesting the LLM to assess the quality of the songs but you are expecting a selection according to the reviews or other articles that have been used to train the LLM.

Quality Booster Categories

Curation Quality:

• “carefully curated” — expect songs that have been created by a curator
• “hand-picked” — implies human-level taste and attention
• “essential” — the definitive, must-have tracks
• “definitive collection” — comprehensive and authoritative

Discovery Quality:

• “deep cuts” — beyond the obvious hits
• “hidden gems” — lesser-known excellence
• “underground” — not mainstream, cult favorites
• “critics’ favorites” — acclaimed but not necessarily popular

Production Quality:

• “audiophile-quality” — exceptional production and mastering
• “pristine recordings” — technically excellent sound
• “analog warmth” — vintage, non-digital character
• “live session quality” — authentic, unpolished energy

Cultural Significance:

• “era-defining” — songs that shaped their time
• “influential” — tracks that inspired others
• “timeless” — quality that transcends trends
• “genre-defining” — established or perfected a style

Example with Quality Boosters

Basic: “90s hip-hop playlist”

Enhanced: “Essential ’90s hip-hop deep cuts, underground classics, and era-defining tracks beyond the obvious hits, curated for heads who know”

Technique 3: Negative Prompts (Exclusions)

Image Parallel: “ — no frames, watermarks, text” or “Homer Simpson — no cartoon” to exclude unwanted elements or break stereotypes.

Exclusions are powerful for avoiding clichés, breaking out of filter bubbles, and creating more original playlists.

Types of Exclusions

Content Exclusions:

• “no covers or remixes” — originals only
• “no live versions” — studio recordings only
• “no Christmas songs” — avoid seasonal creep
• “no explicit lyrics” — family-friendly

Popularity Exclusions:

• “no radio hits” — avoid overplayed tracks
• “no songs over 100M streams” — discover lesser-known artists
• “avoid my most-played” — break out of your comfort zone
• “no TikTok viral songs” — avoid trend fatigue

Stereotype Breaking (The “Homer Simpson — no cartoon” Technique):

Get Pigro Crafter’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Combine a genre with exclusions that break its typical associations:

• “country music — no trucks, beer, or heartbreak” — discovers alternative country
• “EDM — no drops or builds” — finds minimal electronic
• “metal — no aggression or screaming” — discovers doom, post-metal
• “jazz — no swing or standards” — finds avant-garde, fusion
• “pop — no love songs” — discovers thematic diversity

Sonic Exclusions:

• “no vocals” — instrumental only
• “no acoustic guitars” — specific instrument exclusion
• “no autotune” — natural vocal performances
• “avoid songs over 5 minutes” — length constraints

Technique 4: Weighted Terms & Proportional Mixing

Image Parallel: “Golden Gate Bridge::1 in the style of Van Gogh::0.8, in the style of Dalí::0.2” — blending influences with specific weights.

Nano Banana Generated painting with the background in the style of Van Gogh and the robocops in the style of Salvador Dali

Expressing proportions helps create nuanced blends rather than equal mixes.

Percentage-Based Mixing

• “70% upbeat indie pop, 30% melancholic ballads for emotional variety”
• “Mostly jazz (80%) with touches of hip-hop (20%)”
• “90% familiar favorites, 10% discovery tracks”

Primary/Secondary/Accent Structure

Like image composition with foreground, midground, and background:

• Primary (60%): Chill electronic ambient
• Secondary (30%): Jazz piano interludes
• Accent (10%): Unexpected classical moments

Artist Influence Weights

• “Sound like Radiohead (heavy influence) meets Massive Attack (moderate) with hints of Björk (subtle)”
• “Primarily influenced by The Strokes, secondarily by Arctic Monkeys, with occasional Interpol energy”

Technique 5: Playlist Inpainting & Outpainting

Image Parallel: Inpainting modifies part of an image while preserving the rest. Outpainting extends the canvas beyond its original boundaries.

Playlist Inpainting: Targeted Modifications

Keep most of a playlist but surgically modify specific aspects:

• “Replace the slower tracks in this workout playlist with higher-energy alternatives”
• “Keep the vibe but swap out songs I’ve heard too many times”
• “This is perfect, but add 3 palate cleansers between the intense sections”
• “Maintain this playlist but update it with 2024 releases that fit”

Playlist Outpainting: Extensions

Expand a playlist while maintaining its character:

• “This 20-song playlist is perfect. Extend it to 50 songs without losing the vibe”
• “Add a 10-song wind-down section to the end of this party playlist”
• “Create a prequel playlist — what would play before this?”
• “Build out the edges: add deeper cuts from these same artists”

Technique 6: Flow & Transitions (Consistent Characters)

Image Parallel: Generating multiple images of the same character with consistent features across different scenes.

In playlists, “consistent characters” means maintaining coherent flow and intentional transitions between songs.

Transition Strategies

Key Matching:

• “Arrange songs so each flows naturally into the next through compatible keys”

Energy Arcs:

• “Build energy gradually for the first 10 songs, peak at tracks 11–15, then wind down”
• “Maintain steady mid-energy throughout — no spikes or dips”

Thematic Threading:

• “Each song should share at least one element with the previous: same producer, similar tempo, related theme, or shared influence”
• “Tell a story arc: hope → struggle → triumph”

Intentional Contrast:

• “Alternate between high-energy and chill every 2–3 songs”
• “Pair each electronic track with an acoustic palette cleanser”
• “Create tension through unexpected genre jumps that still feel cohesive”

Technique 7: Genre Vocabulary & Hierarchies

Image Parallel: Google maintains lists of artists and art movements (Impressionism, Art Deco, etc.) that help users craft precise prompts.

Having a rich vocabulary for musical styles dramatically improves prompt quality. However, I intentionally de-emphasize genre as a primary search tool because it’s often ineffective on its own. Broad genres (e.g., “rock”) are too vague, while thousands of specific subgenres are obscure to most listeners. Adding a genre qualifier can be useful alongside other parameters, so include it if you know the name of the niche genre you like. Otherwise, I suggest asking the LLM to classify a song you like to get more specific terminology.

Genre Hierarchy: From Broad to Specific

Like art movements, musical genres have hierarchies. Specificity improves results.

Electronic Music Hierarchy:

• Level 1: Electronic
• Level 2: House, Techno, Ambient, IDM, Drum & Bass
• Level 3: Deep House, Acid Techno, Dark Ambient, Glitch, Liquid DnB
• Level 4: Microhouse, Berlin Minimal, Drone Ambient, Breakcore

Rock Music Hierarchy:

• Level 1: Rock
• Level 2: Alternative, Metal, Punk, Progressive, Indie
• Level 3: Shoegaze, Doom Metal, Post-Punk, Math Rock, Dream Pop
• Level 4: Blackgaze, Sludge, Coldwave, Midwest Emo

Hip-Hop Hierarchy:

• Level 1: Hip-Hop
• Level 2: Trap, Boom Bap, Conscious, Southern, East Coast
• Level 3: Cloud Rap, Jazz Rap, Drill, Phonk, Chopped & Screwed
• Level 4: Memphis Rap, UK Drill, Vaporwave Rap, Abstract Hip-Hop

Conclusion

By borrowing these proven techniques from image generation prompting, you can move beyond simple “songs like X” requests and create truly customized playlists. The key is combining multiple techniques: use LLM-assisted prompt crafting to find the right vocabulary, add quality boosters to signal your expectations, employ negative prompts to break stereotypes, weight your influences for nuanced blends, and specify your desired flow and transitions.

Experiment with these approaches, and you’ll find that AI playlist generation becomes far more powerful and personalized than traditional recommendation algorithms.