AI won’t write your song. It will tell you why your bridge doesn’t work. The prompts that actually help songwriters are diagnostic rather than generative — they interrogate a draft you’ve already written, locate the line you’re protecting because you like it rather than because it earns its place, and name the structural reason a section falls flat. Ask a model to write your chorus and you get something forgettable. Ask it to explain why your chorus is forgettable and you get something you can use.

Best for: songwriters with a folder of half-finished songs who can tell something is wrong but can’t name it, and who’ve tried asking an AI for help and got back generic praise or worse, generic lyrics. Skip if: you’re looking for a tool to write songs for you — this guide is explicitly not that, and the reason why is here.

Time cost, honestly: 10 minutes a day. The first prompt below will surface something useful about a song you’ve been stuck on within a single session — that’s not a promise about the tool, it’s a consequence of being asked structured questions about your own work. The compounding benefit takes about three weeks, when you start running the diagnostic questions in your head before you open anything.

---

Why won’t this guide give me prompts that write lyrics?

Because they don’t work, and because they cost you something. Generative lyric prompts return the statistical middle of everything ever written on a theme — which is precisely the opposite of what makes a song land. The model is optimising for plausible; songwriting rewards specific.

There’s a second reason. If a song is worth writing, the writing is the part you wanted. Outsourcing it removes the only step that had value. What you actually want outsourced is the diagnosis — the thing a good co-writer does, which is notice what you can’t see because you’re inside it.

What can’t AI do for songwriters?

It cannot hear. That single limitation defines everything else. A language model has no access to your melody, your voice, your timing or your mix — it has access to descriptions of them. So it cannot tell you whether a hook is catchy, whether a take has feeling, or whether a chord voicing sounds muddy.

What it can do is analyse structure, prosody, specificity and logic — all of which are textual properties. Keep it on that side of the line and it’s genuinely useful. Ask it to judge sound and it will confidently invent an answer.

Why does AI keep telling me my song is good?

Because it’s trained to be agreeable, and unconstrained it defaults to encouragement. Paste a verse and ask what it thinks, and you’ll get “evocative imagery” and “strong emotional resonance” regardless of quality. This is the single biggest reason songwriters try AI feedback once and dismiss it.

The fix is instructional, not technical. Every prompt below contains an explicit anti-flattery constraint — usually a line telling the model that praise is not useful and that it should argue the opposite case. Without it, you’re reading a compliment generator.

How do most songwriters use AI wrong?

They ask for output when they should ask for analysis. The difference determines whether you get something usable or something you delete.

Situation	What most people type	What actually works
Stuck on a verse	“Write me a second verse about losing someone”	“Here’s verse one. What question does it raise that verse two has to answer?”
Bridge falls flat	“Write a bridge for this song”	“What can a bridge say here that the verses and chorus structurally cannot?”
Wanting feedback	“What do you think of these lyrics?”	“Argue this song is generic. Use my own words as evidence. Do not soften it.”
Chord help	“Give me a sad chord progression”	“My progression is 1–5–6m–4. Give three substitutions and the emotional reason for each.”
Too many drafts	“Which version is better?”	“What is each version optimising for? Which one is closer to what I said the song was about?”
Can’t finish	“Finish this song for me”	“Interview me about this song. One question at a time. Don’t offer opinions until I say go.”

Related: [INTERNAL LINK 1 — your existing “10 Songwriting Prompts” post. Suggested anchor: “the non-AI prompts that start a song from nothing”]

How should I format a prompt so it actually works?

Four elements, and skipping any one of them degrades the result. Assign a role with a stated bias, so the model isn’t averaging every perspective at once. Constrain the output — one question at a time, no praise, no lyrics. Give it your material in a format it can parse, which for harmony means numbers rather than letters. State the failure condition — tell it what a bad answer looks like.

Every prompt below carries a stated success criterion and an honest note on where it breaks.

---

Prompt 1 — The Song Interview (start here)

The flagship. It refuses to give you an opinion until it has interviewed you properly, which is where the actual value sits — most songwriters discover what’s wrong while answering question nine.

PROMPT — THE SONG INTERVIEW
Use when:  You have an unfinished song and can't name what's wrong
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]
Reasoning: On, if available

You are an experienced co-writer who has finished about 200 songs.
You are not a fan of mine and you are not here to encourage me.

Interview me about a song I haven't finished. Ask ONE question at
a time and wait for my answer. Do not give any opinion, suggestion,
lyric or chord until I type READY.

Adapt each question to what I just said. Ask follow-ups when an
answer is vague, evasive, or sounds like something I've said to
other people before.

Cover, in whatever order the conversation earns:
  1. What happens in this song? Not the theme — the events.
  2. Who is speaking, and to whom?
  3. What does the singer want, and what stops them getting it?
  4. What changes between the start and the end?
  5. Which section came first, and which came last?
  6. Which part do you play for people, and which do you skip?
  7. What is the song's best line, and why?
  8. What is in it only because you couldn't think of anything else?
  9. Where does it get boring? Be exact — which bar.
 10. What would have to be true for you to abandon it?

Rules:
  - One question per message. Short. Conversational.
  - Never summarise my answers back to me.
  - If I give a non-answer, say so and ask again.
  - Do not compliment anything. Praise is not useful to me.

When you have enough — usually 12 to 18 exchanges — say:
"I have what I need. Type READY when you want the diagnosis."

Then produce:
  - What the song appears to be about (in my words)
  - What it is actually about (what my answers imply)
  - The gap between those two, if there is one
  - The three weakest structural decisions, named specifically
  - The one thing to fix first, and why that one

Begin with your first question.

Done when:  You can state, in one sentence, what the song is
            about — and it's different from what you'd have
            said before the interview.
Fails when: You answer in generalities. The prompt is only as
            good as your willingness to say "I don't know."
            It also fails if you paste the whole lyric up front —
            that lets the model skip to opinions.

Related: [INTERNAL LINK 2 — “How to Structure a Song.” Suggested anchor: “what each section is supposed to be doing”]

Prompt 2 — The Darling Audit

Nine words of instruction, and the most reliably useful prompt in this guide. Every songwriter has a line they’re protecting. Most of us know exactly which one the moment we’re asked.

PROMPT — THE DARLING AUDIT
Use when:  A song is nearly done and something won't sit right
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

Here is a lyric: [PASTE LYRIC]
Here is what I said the song is about: [ONE SENTENCE]

Identify the line I am keeping because I like it rather than
because it earns its place. Name exactly one.

For that line, tell me:
  - What it is doing that the song needs
  - What it is doing that the song doesn't need
  - What the song would lose if I cut it
  - What the song would gain

Then do the same for the line that is working hardest and
getting the least attention.

Do not rewrite anything. Do not suggest alternatives.
Do not praise the lyric. If you cannot identify a darling,
say so plainly rather than inventing one.

Done when:  You feel defensive about the line it named.
            That reaction is the diagnostic.
Fails when: Your lyric is short — under two verses it has too
            little to compare against. It also fails if you
            don't give it the one-sentence intent, because
            "earning its place" is meaningless without a
            statement of what the place is for.

Prompt 3 — The Bridge Diagnostic

A bridge should do something the verse and chorus structurally cannot. Most failed bridges fail because they say the same thing in a different key. This prompt names which failure you’ve made.

PROMPT — THE BRIDGE DIAGNOSTIC
Use when:  The bridge is written and it lands flat
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

VERSE:  [PASTE]
CHORUS: [PASTE]
BRIDGE: [PASTE]

Progression (in numbers, not letters):
  verse [e.g. 1 5 6m 4] · chorus [   ] · bridge [   ]

Answer in this order:

1. What information does the listener have after the second
   chorus that they didn't have at the start?
2. What does the bridge add that neither the verse nor the
   chorus could have delivered?
3. If the answer to 2 is "nothing," say so directly.
4. Which of these failure modes am I in:
     a) The bridge restates the chorus in different words
     b) The bridge introduces something the song never uses
     c) The bridge is a key change with no new content
     d) The bridge answers a question the song never asked
     e) The bridge is fine and the problem is elsewhere
5. What question does the second chorus leave unanswered?
   That question is what the bridge is for.

Do not write me a bridge. Do not suggest lines.

Done when:  You can name what your bridge is for in one
            sentence, without using the word "contrast."
Fails when: You haven't written a bridge yet — this is a
            diagnostic, not a generator. Write a bad one
            first, then run this.

Prompt 4 — The Specificity Audit

Vagueness is the most common failure in amateur lyrics and the easiest to fix. This prompt finds every place you wrote a category where you could have written a thing.

PROMPT — THE SPECIFICITY AUDIT
Use when:  The lyric is finished but feels like anyone's song
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

Lyric: [PASTE]

Go through line by line. For each line containing an abstract
noun, a category word, or a stated emotion, flag it and tell me
what CONCRETE detail could occupy that space instead.

Do not write the replacement. Ask me the question that would
get it out of me. For example, if I wrote "we were happy then,"
ask what we were doing at the exact moment I'm thinking of.

Then rank the flagged lines by how much the song would gain
from replacing each.

Finally: name any line that is already doing this well, so I
can see the contrast in my own writing.

Do not rewrite. Do not praise. Do not suggest lyrics.

Done when:  You have at least three questions you can't answer
            without going back to the actual memory.
Fails when: The song is deliberately abstract — some songs are,
            and this prompt will flatten them. Use judgement.

Related: [INTERNAL LINK 3 — your Taylor Swift specificity/craft piece. Suggested anchor: “the small-detail technique, analysed”]

Prompt 5 — Chord Substitution by Function

Give the model numbers rather than letters. Roman-numeral or Nashville-number input produces substitutions that transpose to any key, and it dramatically reduces the rate at which the model invents chord tones that aren’t there.

PROMPT — SUBSTITUTION BY FUNCTION
Use when:  A progression works but feels predictable
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

My progression, in numbers: [e.g. | 1 | 5 | 6m | 4 |]
Key I play it in: [KEY]
What the section is doing emotionally: [ONE SENTENCE]

For each chord, tell me:
  - What job it is doing (home / away / pull-home)
  - Two chords that could do the same job
  - One chord that would do a DIFFERENT job there, and what
    that would change about the section

Then give me three whole-progression variants:
  a) One that is sadder without changing the melody
  b) One that delays resolution by one bar
  c) One that uses exactly one borrowed chord

Give every answer in numbers. Then, and only then, convert
to letters for my key.

Flag anything you are uncertain about rather than guessing.

Done when:  You can play one variant and hear the difference
            without reading the explanation.
Fails when: You give it letters instead of numbers — accuracy
            drops noticeably. It also fails on specific
            voicings and inversions; models routinely state
            chord tones that aren't in the chord. Verify
            anything it claims about individual notes.

Prompt 6 — The Prosody Check

Prosody is whether the natural stress of the words matches the stress of the music. When it’s wrong, listeners feel it without being able to name it — the line sounds slightly foreign in the singer’s mouth.

PROMPT — THE PROSODY CHECK
Use when:  A line is hard to sing and you don't know why
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

Line: [PASTE ONE LINE]
Where the strong beats fall: [MARK THEM, e.g. "beat 1 on 'never',
beat 3 on 'said'"]

1. Break the line into syllables and mark the natural spoken
   stress of each.
2. Show me where spoken stress and musical stress disagree.
3. For each disagreement, tell me whether it is a problem or a
   deliberate-sounding tension.
4. Suggest where the line could be re-broken across the bar to
   fix it — the same words, distributed differently.

Do not change my words. Do not write a new line.

Done when:  You can say the line out loud and hear which
            syllable was fighting the beat.
Fails when: You don't tell it where the beats are. Without
            that it is guessing at your melody, and it will
            guess confidently. This prompt is also unreliable
            on syncopated or heavily melismatic lines.

Prompt 7 — The Adversary

Run this against a song you think is finished. It’s the cheapest error-correction available and the one most people skip because it’s unpleasant.

PROMPT — THE ADVERSARY
Use when:  You think a song is done
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]

Song: [PASTE LYRIC + SECTION LABELS + PROGRESSIONS IN NUMBERS]

Argue that this song is generic. Build the strongest honest
case against it, using my own words as evidence.

Specifically:
  1. Which images or phrases have been used many times before?
  2. Where am I saying something the listener already assumed?
  3. What is the most predictable moment, and why?
  4. If a hundred songwriters were given my first line, how
     many would arrive at my second?
  5. What am I avoiding saying?

Then, separately: what is genuinely unusual here? Name it
precisely, or say there is nothing.

Do not soften your conclusions to be encouraging. If the song
is strong, say so plainly and stop — do not invent objections
for balance.

Done when:  You disagree with at least one point and can say
            exactly why. That's the song defending itself.
Fails when: The model finds nothing and says the song is great.
            That usually means you pasted too little context.
            Include structure and progressions, not just words.

Worked example: a bridge that wasn’t a bridge

Composite from teaching sessions — structure and progressions unchanged, lyric details altered.

Input. A songwriter brings a nearly-finished song. Two verses, chorus, bridge, final chorus. Their stated problem: “the bridge is boring but I don’t know what to replace it with.”

STRUCTURE AND HARMONY (as submitted)

VERSE   | 1  | 5  | 6m | 4  |    ×2
CHORUS  | 4  | 1  | 5  | 6m |    ×2
BRIDGE  | 4  | 1  | 5  | 6m |    ×2   ← identical to chorus
CHORUS  | 4  | 1  | 5  | 6m |    ×2

Stated intent: "It's about moving out of a house."

Step 1 — run Prompt 3, the Bridge Diagnostic. The model’s answer to question 3 is immediate and unambiguous: the bridge adds nothing the chorus couldn’t, because it is harmonically identical to the chorus. Failure mode (c), with elements of (a).

That’s not a subtle diagnosis. But it’s one the writer couldn’t see, because they’d been listening to the melody — which is different — and hadn’t noticed the harmony underneath was the same four bars they’d already played eight times.

Step 2 — the model’s question 5 is the useful one. “What question does the second chorus leave unanswered?” The writer’s answer: “Whether they’re glad they left.”

That sentence is the bridge. It didn’t exist before the question was asked.

Step 3 — run Prompt 5 on the bridge harmony. Input: the chorus progression, with the note that the section needs to feel unresolved rather than settled.

OUTPUT — three variants, in numbers

a) sadder, melody unchanged
   | 4  | 1  | 5  | 6m |  →  | 4m | 1  | 5  | 6m |
   the borrowed minor 4 darkens bar 1 without moving
   the vocal at all

b) delays resolution
   | 6m | 4  | 1  | 5  |
   same four chords, rotated — the section now ENDS on
   the 5, so it cannot sit still. It has to go somewhere.

c) one borrowed chord
   | 6m | 4  | ♭7 | 5  |
   the ♭7 removes the pull home entirely for one bar,
   which reads as hesitation

CHOSEN: (b), because "whether they're glad they left"
is a question, and a section that ends unresolved is
the harmonic form of a question.

Output — the fix, in one line of harmony. Same four chords the song already used, rotated to end on the 5 instead of the 6m:

BRIDGE (revised)   | 6m | 4  | 1  | 5  |   ×2

In G:   | Em | C  | G  | D  |
In C:   | Am | F  | C  | G  |
In D:   | Bm | G  | D  | A  |

The final D (the 5) does not resolve until the last
chorus starts. That unresolved bar is the whole fix.

No AI wrote a lyric, a melody or a chord. It asked a question the writer hadn’t asked themselves, and named a structural fact they’d stopped being able to see. That’s the entire proposition of this guide.

Level-up: the Three-Voice Panel

The advanced version, and the part competing guides won’t have. A single model asked for feedback averages every perspective at once and produces mush. Assigning three conflicting roles with explicit biases forces the disagreements into the open — and the disagreements are where the useful information is.

PROMPT — THE THREE-VOICE PANEL
Use when:  A song is finished and you have a decision to make
Model:     [TESTED AGAINST — MODEL + VERSION + DATE]
Reasoning: On

Convene three people to review my song. They have different
jobs and different blind spots. They must genuinely disagree.

1. THE PRODUCER — cares whether it holds attention. Thinks in
   seconds and sections. Blind spot: will sacrifice meaning
   for momentum.
2. THE LISTENER — has never heard it, isn't a musician, has
   no reason to be generous. Cares only whether they'd play
   it twice. Blind spot: can't articulate why.
3. THE RIVAL SONGWRITER — technically excellent, slightly
   competitive, wants to find the weakness. Blind spot:
   overvalues cleverness.

SONG: [LYRIC + SECTION LABELS + PROGRESSIONS IN NUMBERS
       + APPROXIMATE SECTION LENGTHS IN SECONDS]

ROUND 1 — Each gives their reaction in under 120 words,
including the one thing they'd change.

ROUND 2 — Each responds to the panellist they most disagree
with, by name. I need at least two real disagreements. If
all three agree, say so explicitly and explain why the song
made it easy.

ROUND 3 — Each names the one condition under which they'd
change their own position.

ROUND 4 — A joint note covering:
   · what all three agree on
   · what they couldn't resolve, and what would resolve it
   · what they suspect I'm not telling them
   · the single change with the highest ratio of impact
     to effort

Then, in your own voice as facilitator: which panellist is
most right, and why? Name them. Do not hedge.

No praise in any round. No lyrics written by anyone.

Done when:  Two panellists disagree about something you also
            feel two ways about. That's the real decision.
Fails when: You give it lyrics alone. Without section lengths
            and harmony the Producer has nothing to work with
            and the panel collapses into one voice.
            It also degrades over long conversations — start
            a fresh session for this one.

Related: [INTERNAL LINK 4 — “Music Theory for Songwriters” pillar. Suggested anchor: “how to give a model your progression in numbers”]

The 30-day plan: 10 minutes a day

Every session uses a song you’ve already written. No new material required.

═══════════════════════════════════════════════
WEEK 1 — DIAGNOSIS
10 min/day
  Run PROMPT 1 (Song Interview) on one unfinished
  song per session. Do not skip to READY early.
  Songs : [YOUR 5 MOST STUCK DRAFTS]
  Log   : one sentence per song — what it's
          actually about, in your words.
  Goal  : 5 songs correctly diagnosed. At least
          two turn out to be about something
          different from what you assumed.

═══════════════════════════════════════════════
WEEK 2 — LINE LEVEL
10 min/day
  Day 8–10  : PROMPT 2 (Darling Audit) on three
              near-finished lyrics.
  Day 11–14 : PROMPT 4 (Specificity Audit) on the
              same three. Answer the questions it
              asks — in a notebook, not in the chat.
  Goal  : one line replaced with a concrete detail
          you had to go and remember.

═══════════════════════════════════════════════
WEEK 3 — STRUCTURE AND SOUND
10 min/day
  Day 15–17 : PROMPT 3 (Bridge Diagnostic) on any
              song with a bridge you don't like.
  Day 18–19 : PROMPT 5 (Substitution) — feed it
              NUMBERS. Play every variant.
  Day 20–21 : PROMPT 6 (Prosody) on the three
              hardest lines you've ever had to sing.
  Goal  : one section rewritten harmonically, and
          you can hear why the new one is better.

═══════════════════════════════════════════════
WEEK 4 — PRESSURE AND PANEL
10 min/day
  Day 22–24 : PROMPT 7 (Adversary) on your best
              finished song. Sit with it.
  Day 25–27 : THE THREE-VOICE PANEL on the same song.
  Day 28–30 : Write four new bars using nothing but
              the questions. No AI open. This is
              the actual goal — the prompts become
              questions you ask yourself.
  Goal  : you run the diagnostic in your head
          before you open anything.
═══════════════════════════════════════════════

Frequently asked questions

What are the best AI prompts for songwriters?

The best prompts are diagnostic rather than generative. The single most useful is an interview prompt that makes the model ask you fifteen questions about an unfinished song before offering any opinion — most of the value arrives while you’re answering. After that, a “darling audit” that identifies the line you’re keeping because you like it, and an adversarial prompt that argues your song is generic using your own words as evidence.

Can AI write a good song?

Not one worth releasing. A language model optimises for plausible, and songs work by being specific — the exact opposite objective. Generative lyric prompts return the statistical middle of everything written on a theme. What AI does well is analyse structure, prosody, specificity and logic, which are textual properties. It cannot judge melody, feel or sound, because it cannot hear.

Why does ChatGPT always say my song is good?

Because it’s trained to be agreeable and defaults to encouragement when unconstrained. Paste a verse and ask for thoughts and you’ll get “evocative imagery” regardless of quality. The fix is instructional: include an explicit clause telling the model that praise is not useful and instructing it to build the strongest case against your song. Without that constraint you’re reading a compliment generator.

How do I get useful feedback on my song from AI?

Give it more than lyrics. Include section labels, approximate section lengths in seconds, and your chord progressions written as numbers rather than letters. Then assign a role with a stated bias and forbid praise outright. Feedback quality is almost entirely a function of context supplied and constraints imposed — vague input reliably produces flattery.

Should I give AI my chords as letters or numbers?

Numbers. Writing a progression as 1–5–6m–4 rather than C–G–Am–F produces substitutions that transpose to any key, and it measurably reduces the rate at which models invent chord tones that aren’t present. Language models are unreliable about specific voicings and inversions in any format, so verify anything a model claims about individual notes inside a chord.

Will using AI make my songwriting worse?

It depends entirely on which half you use. Outsourcing the writing removes the step that had value and tends to flatten a writer’s voice toward convention. Using it to interrogate work you’ve already done does the opposite — it surfaces the structural decisions you stopped being able to see. The distinction is between asking for output and asking for analysis.

Can AI help me finish a song I’m stuck on?

Yes, but not by finishing it. The mechanism is that structured questions surface what you already know but haven’t articulated. Being asked “what question does your second chorus leave unanswered?” frequently produces the bridge, because the answer is a sentence you could always have written but had never been prompted to say out loud.

Is it cheating to use AI for songwriting?

Using it to write your lyrics is substituting for the work. Using it to diagnose your draft is closer to what a co-writer, producer or trusted friend does — noticing what you can’t see because you’re inside it. Songwriters have always used outside ears. The relevant question is whether the words and music are yours, and with diagnostic prompts they remain so.

---

About the author

Publisher’s note: on an AI-related topic the author block carries more weight than usual, because the default assumption is that the page is machine-generated. Prioritise, in order: a named human with a linkable profile; audible proof of the skill; a quantified experience claim; and one specific admitted failure with AI. Do not publish this page under an unattributed byline or a generic “our team” description.

Sources and tools

• musictheory.net — free interactive lessons on scales, intervals and chord construction; useful for verifying anything a model tells you about harmony. musictheory.net/lessons
• teoria.com — free harmony and ear-training exercises, including chord function and progression recognition. teoria.com/en/exercises
• tonedear.com — browser ear trainer. Relevant here because it does the one thing AI cannot: test whether you can actually hear the difference. tonedear.com
• JustinGuitar — free structured curriculum; theory modules cover the number system these prompts require. justinguitar.com
• Hooktheory / TheoryTab — songs analysed in Roman numerals; the fastest way to check a model’s structural claims against documented analysis. hooktheory.com/theorytab
• Premier Guitar — lesson archives covering applied harmony and modal interchange. premierguitar.com

A caveat that applies to every prompt on this page: language models state incorrect musical facts with complete confidence, particularly about chord tones, inversions and key relationships. Treat every specific claim about notes as a hypothesis to verify on your instrument. The prompts here are structured to minimise that exposure — which is why they ask for function and structure rather than voicings — but the risk is not eliminated.

Last updated

Version notes

• Model testing — every prompt block carries a Model: line that must name the model, version and test date. Do not publish with those left blank; an untested prompt on an AI page is the exact failure the audience is watching for.
• Re-test triggers — any major release from a frontier lab. Priority order for re-testing: Prompt 1 (longest, most likely to break character), the Three-Voice Panel (multi-role, most fragile), then the rest.
• Known degradation — the interview and panel prompts both lose discipline in long sessions. Note this on the page if it worsens; currently mitigated by the instruction to start a fresh session.
• Tool links — six resources listed above, all long-established and current at publication. Re-verify at each review and log the date here.
• Musical claims — the worked example progressions were derived by hand, not generated. If the example is replaced, derive the replacement by hand rather than accepting model output.

More from this series: [INTERNAL LINK 5 — “Practise Theory While Writing.” Suggested anchor: “using these questions without a screen”]