Songwriting Craft Loop
A simpler, more universal writing loop — find a hook, pick a voice, structure it, and write a solid song.
Anyone who wants a strong, performable song without a deep theme-study pass — a lighter, general-purpose writing method for everyday songs, and the generic alternative to the deeper theme-first loop.
This is the craft-first writing loop. Where a theme-first loop starts from "what is this song about," this one starts from the songwriter's actual job: find the one thing worth repeating, build a hook a stranger can sing back after one listen, and let melody, performance, and production carry the rest. It is the antidote to the most common failure mode in AI songwriting — the "concept developer" who writes thoughtful, meaningful, completely forgettable mini-sermons with choruses. Run this when you want lyrics that actually stick.
The loop is deliberately divergent before it converges. Your agent does a research-lite pass on the seed (the real emotion, the power dynamic, the one image), then brainstorms wide: six to ten hook candidates, multiple chorus directions, lines pulled straight from the source that can be sung as-is or lightly reshaped to be more musical. Only then does it run the tests and pick. Keep those brainstorm artifacts attached to the song as documents, so the rejected hooks and alternate angles are never lost — they become the well you draw from when you write the next track.
Voice is a deliberate decision, not a default. Before writing a single verse, the agent names the emotional register and energy that fits the song — and may quietly use a well-known artist as a private craft reference for cadence, vocabulary, and attitude. Across an album it rotates that register track to track so the record doesn't go monolithic. Crucially, any artist reference lives only in the agent's thinking. The final platform style prompt describes the voice with adjectives and craft language — "intimate melodic rap, soft piano, breathy vocals" — never a real artist's name, and the goal is an original voice, not an impersonation.
What comes out is a complete, performable spec: a tested hook, a structure where every section has a job, lyrics with embedded performance and production cues, and a platform-ready style prompt shaped to length. You hand that spec to a create loop to render it. This loop's reference brain is the suno-songwriting skill — the source of the mindset, the hook tests, and the structure jobs below.
Skills & actions it uses
The concrete, reusable skills this loop calls to actually do the work.
suno-songwritingThe craft reference brain for this loop — supplies the songwriter mindset, hook tests, structure jobs, bridge/outro craft, and the style/voice translation library the loop runs on.suno-addThe create-loop handoff — takes the finished spec (title, style prompt, tagged lyrics) and renders it into audio. This loop writes the spec; suno-add produces the track.
The loop
- 1
Research-lite the seed
Before any lyrics, have your agent interrogate the seed and write its thinking down as an attached document. The goal is to find the one emotion and the one image worth a whole song — not to summarize everything the seed contains.
- Name the real feeling underneath the seed, in plain words (not "be strong" but "stand still"; not "feel good" but "midnight comin'").
- Identify the power dynamic and its modern analog — the boss, the algorithm, the diagnosis, the addiction.
- Pull the single strongest image, the one you can almost see.
- If working from source text, mine quotable lines now: phrases that can be sung as-is or lightly reshaped to scan and rhyme.
- Write a one-sentence summary of the song. If you can't, the concept isn't ready — keep digging.
- 2
Brainstorm hooks wide, then test to converge
Generate six to ten hook candidates before committing to any. A great hook is short (4-10 syllables), repeatable, image-loaded, emotionally specific, and lands on a stranger with zero context. Then run the tests and keep the one or two strongest.
- Hook patterns to mine: command + image ("Mark the door"), confession + image ("We got tired of waiting"), reframe, counter-thesis ("I thought I was burning out / but I was burning holy"), negation, question-that-becomes-answer.
- The stranger test: can someone sing it back after one listen? If not, simplify.
- The chant test: does it still hit chanted six-plus times?
- The see-it test: is it loaded with a concrete image, or is it abstract "meaning"?
- Keep the rejected hooks as an attached document — they seed future tracks and alternate versions.
- 3
Decide genre, sub-genre(s), and the target voice
Choose the sound and the voice on purpose to kill AI slop. Pick a primary genre and one or two sub-genres for texture, then lock the emotional register and energy you'll write to — using any artist as a private craft reference only, never in the final prompt.
- Match the emotional register to a voice: confessional 3am vulnerability, hard prophetic-conscious, smooth West Coast warmth, dark cinematic trap, mournful soul — each implies different cadence and instrumentation.
- Write the verses to that register's rhythm, vocabulary, and attitude. This is what makes lines sound original instead of generic "modern hip-hop."
- Across an album, rotate the register track to track so the record gets variety you didn't have to manufacture.
- Translate the voice into a sonic signature — instruments, vocal mode, energy — ready for the style prompt. Any artist reference stays in your head.
- 4
Give every section a job (structure)
Outline the song so each section does distinct work, then write to those jobs. Most replay-worthy songs run 2:30-3:30; anthems and finales can stretch to 4:00. Resist the 5-minute "audiobook with excellent drums" trap.
- Intro: set tone in 4-8 lines or just chant fragments.
- Verse: 8-12 lines. Verse 1 is the listener's actual life with the seed as texture — never a recap of the source.
- Pre-chorus (optional but powerful): 4 lines of lift into the hook.
- Chorus: 6-10 lines. The thesis, the payoff, the quotable line.
- Bridge: 4-8 lines that SHIFT something — strip the production, flip the register (whisper or scream), add call-and-response, or land one loaded line. Never another verse.
- Outro: 2-4 lines max. Repeat the hook's most loaded phrase and get out. No new content, no summary.
- 5
Write listener-first, image-driven lyrics
Write the lyric to the structure, obeying two rules that separate amateur from craft: don't explain it, feel it; and make Verse 1 the listener's life, not the source narrative.
- Don't-say-it-feel-it: "Mark the door" beats "paint the blood on your doorpost so the angel passes over." The action carries the meaning — trust the listener.
- The Verse-1 test: can someone who doesn't know your source still feel this song? If no, rewrite V1 using source imagery as a single phrase or image, not a story.
- Decide a source/modern/message ratio per track — most should lean modern-heavy (roughly 15-25% source, 65-75% the listener's life, 5-15% explicit message). Pure-message tracks fall flat; pure-source feels like a lecture.
- 6
Embed performance and production cues
Writing is more than words — bake the performance into the lyric with tags the platform respects. Use brackets for non-sung direction (always on their own line) and parentheses for sung ad-libs only.
- Section + performance tags: [Verse 1 - confessional], [Chorus - bigger], [Bridge - stripped down], [Final Chorus - climactic], [Outro - fading].
- Vocal direction: [Lead vocal - whispered], [Lead vocal - half-sung], [Choir - call and response], [Background vocals - chopped sample].
- Production direction: [drums fade], [strings swell], [vinyl crackle], [heavy reverb], [808s drop], [hand claps].
- Parentheses are sung background only — hook echoes, call-response answers, atmospheric ad-libs. Never put instructions inside parentheses.
- Choose tag density deliberately — both valid: HEAVY (force instrumental breaks, solos, no-vocal moments) or LIGHT (few tags, let the engine's style defaults decide). Don't leave it accidental.
- Singability: respell hard proper nouns phonetically (Ephraim → 'Ee-fraim', Moroni → 'Moe-roh-nye'), expand archaic contractions (e'en → even, o'er → over), and never write a real artist's name in the lyric.
- 7
Shape the lyric to length and how the style sings
Different styles sing things at different speeds, so the same idea needs more or fewer words depending on the genre. Adjust the lyric to fit the target duration before you finalize.
- A slow soul ballad sings far fewer words per bar than a drill or boom-bap verse — trim or expand lines so the lyric actually fits ~2:45 or ~3:30.
- Vary length and intensity across an album: 2:00-2:30 for interludes and sketches, 2:30-3:00 for standard cuts, 3:00-3:45 for anchors and singles, 3:45-4:30 for the finale.
- Encode the target length in the style prompt (e.g. ~2:45); most platforms hold within about plus or minus 30 seconds.
- If the song feels long after a draft, cut a verse rather than padding the chorus repeats.
- 8
Build the platform style prompt (adjectives, not names)
Compose the final style prompt from the chosen sub-genre(s) and the voice's sonic signature, shaped to the platform's limits. Describe the voice with craft language and adjectives — never a real artist's name.
- Anatomy that works: [Genre]. [3-5 instruments]. [Vocal type]. [One emotional descriptor]. [~Duration].
- Front-load the genre (the first words carry the most weight) and keep it tight — many platforms cap the style field around 200 characters.
- Specify the vocal mode (rap vs sung vs spoken vs mix) and exactly one emotional anchor.
- Don't list 3+ genres, don't repeat what's already in the lyric tags, don't use real artist names, and use at most one negation ("not churchy") — negating more just makes the model weight the thing you're negating.
- 9
QC, then hand the spec to a create loop
Run a final checklist, then pass the finished spec to a create loop to render it. Don't be precious — if the first renders don't land, the lyric or prompt is the problem, so rewrite the verses or change the style rather than rerolling the same seed.
- Checklist: hook is stranger-friendly; V1 is the listener's life; brackets only for non-sung direction on their own lines; parens only for sung ad-libs; chorus is the emotional headline; bridge does something different; outro is 2-4 lines; style prompt is front-loaded and within the limit.
- Keep the hook constant when iterating; vary the verses, the bridge production, or the voice's register.
- Different angle = different song: if the seed has multiple strong themes, write separate songs rather than cramming everything into one.
- Hand off the title, style prompt, and tagged lyrics to the create loop; the brainstorm and research documents stay attached to the song for the next pass.
Then run it again for the next song, chapter, or track.
You end with
- A tested hook (plus a documented set of rejected hook candidates and alternate angles) attached to the song
- A section-by-section structure where each part has an explicit job, sized to a target length
- Complete lyrics written listener-first and image-driven, with embedded performance and production cues
- A platform-ready style prompt built from the chosen sub-genre(s) and voice, described in adjectives and craft language — no artist names
- A finished, performable song spec ready to hand to a create loop for rendering
Skip this loop if…
- You want to start from a theme or message and work outward — reach for the theme-first writing loop instead
- You just want a quick sketch and are happy letting the platform invent its own lyrics and style
- You're past writing and only need rendering, mastering, or distribution
How it connects
Alternatives
A competing loop for the same job — different tool, taste, or depth. Pick whichever fits you.
Pairs with
Commonly run just before or after this one in the pipeline.