Most AI-generated lyrics share a single defect that is hard to name and impossible to unsee once you’ve named it. The lyric describes a feeling, sometimes evocatively, sometimes for several stanzas, and never once tells you what the feeling cost the person having it. Cost is the missing ingredient that separates a lyric from mood-board copy. This post is about the question Pat Pattison taught a generation of Berklee songwriters to ask, why it works, and how we built it into our forge pipeline at Build 2293.

The defect, in one sentence

Here is a typical AI-generated chorus, picked from the first dozen forges I ran on a generic LLM (not ours) with the prompt “write a country song about loss”:

The empty road stretches before me
The rain falls on my weary soul
I’m lost in the silence of memory
A heart that the years cannot console

This is not bad writing. It’s grammatically clean. It rhymes. It uses concrete nouns (road, rain, heart). It has a clear emotional posture (sadness). If you sang it over a sparse acoustic accompaniment in a major key with a minor IV, listeners would feel something.

And yet it is not a song. It is a mood-board. The narrator has lost something. The narrator is sad about it. The narrator’s sadness is ongoing. None of these is news. Every listener who has lost something has felt this. The lyric makes no claim about what specifically was lost, what the loss cost the narrator to bear, or what the narrator gave up by choosing to bear it rather than do something else.

The defect is that the lyric describes a state without explaining what entered the state from outside. Cost is what enters. Cost is what makes the state news.

What “cost” means in songwriting craft

Pat Pattison, the Berklee songwriting instructor who shaped two generations of working songwriters — Gillian Welch, Susan Cattaneo, John Mayer’s phrasing teacher, dozens of Nashville staff writers — teaches a principle he calls the “object writing” discipline. The principle says: don’t write about the feeling. Write about the object that produced the feeling. The object will carry the feeling without describing it, and the listener will reverse-engineer the feeling more deeply than any direct description could deliver.

Pattison’s object is the surface application of a deeper rule. The deeper rule is: every line should carry evidence of a decision the narrator made under cost. The narrator chose to keep one thing and lose another. The narrator chose to remember instead of forget, to forgive instead of leave, to leave instead of forgive, to drink instead of call, to call instead of drink. The lyric is the record of those choices.

Andrea Stolpe, in her textbook Popular Lyric Writing, frames this as the lyric’s “underlying message of trade.” What the narrator gave up is the load-bearing element. The feeling is the byproduct.

Townes Van Zandt, who never read a songwriting textbook in his life and who Pattison cites constantly, said it in an interview with No Depression in 1997: “If a song says you lost her, the song has to also say what you spent to lose her.” Van Zandt rarely talked craft; this line was uncharacteristically explicit. It’s the same idea Pattison and Stolpe formalized later: cost is the load-bearing element.

The question

Pattison teaches songwriters to ask, before writing any verse: “What did the narrator give up to be in this lyric?”

If the lyric is a breakup song, the question is not “how sad is the narrator about the breakup?” — that’s mood. The question is “what did the narrator give up by ending it, and what would they have had to give up by staying?” The lyric is the difference between those two givings-up.

If the lyric is a parent-and-child song, the question is not “how much does the parent love the child?” — that’s mood. The question is “what did the parent give up by becoming a parent, and what would they have given up by not?”

If the lyric is a leaving-town song, the question is not “is the narrator excited or scared?” — that’s mood. The question is “what specifically does the narrator stop being able to do by leaving, and what specifically becomes possible only by leaving?”

The cost question is asymmetric. There is always a cost on both sides — staying costs something, leaving costs something. The lyric is the documentation of the cost the narrator chose to pay. The other cost, the one the narrator refused, is the lyric’s shadow.

Why AI-generated lyrics fail this test

The default behavior of large language models in lyric-generation mode is to interpolate the average of all the lyrics in their training corpus matching the prompt. The average lyric describes the average feeling using the average vocabulary about the average loss. The cost question is, by definition, a question about specificity — what this specific narrator gave up — and specificity is the first thing averaging destroys.

An LLM asked to write a country song about loss will output the median of every country-loss song it was trained on. The median country-loss song describes loss without naming what was traded for it. The output inherits this.

The fix is not to make the model more “creative” in any general sense. The fix is to constrain the model toward cost specifically. The model needs to be told: before writing the next line, decide what the narrator gave up to be able to say this line.

What we built

At Build 2293 we added a gate to our forge pipeline. Before any lyric leaves the deliver phase, the model is required to produce a Cost Audit. The Cost Audit asks four questions:

What did the narrator give up to be in this moment? (Must be a specific concrete thing or action, not an abstraction.)
What did they refuse to give up? (The other side of the trade.)
Is the cost visible in the lyric, or does the listener have to be told? (Visible costs are weight-bearing; told costs collapse to mood.)
What line in the lyric carries the cost? (The model must identify a specific line. If no line carries it, the lyric fails the gate.)

The Cost Audit is internal to the pipeline. You don’t see it as a user. What you see is the result: lyrics that pass the gate, and a regenerate pass triggered automatically for lyrics that don’t.

The gate is part of the broader Coherence Audit we run between the forge phase and the gauntlet refinement. The Cost Audit is one of nine checks the audit performs; it joined Step 1 (the Detail-Of-Care marker), Step 2 (the Banned-Term Scanner), and Step 3 (the Antagonist Ceiling) at Build 2293.

Before and after

Same prompt, same brief, two forges — one before the Cost Audit gate, one after.

Prompt: “Write a country-Americana song about a woman calling her sister after a long estrangement. The phone call is the song.”

Before (B2292, pre-gate):

I picked up the phone and I dialed your number
After all these years, my hands were shaking
The line connected and I held my breath
Wondering if the silence between us was breaking

This is mood. We know the narrator is nervous, that time has passed, that there has been silence. We do not know what the narrator gave up to make this call, what the narrator refused to give up, or which line carries the cost. The lyric describes the situation and stops.

After (B2293, post-gate):

I dialed the seven numbers I have not dialed since Mama’s funeral
Your husband answered and I asked for you by the name our father used
You came to the phone and I said the thing I drove three states to not say
The kettle on your stove was screaming and you let it scream

This passes the Cost Audit. What the narrator gave up: the right to keep the seven numbers unsaid; the right to keep the long-form ending intact. What the narrator refused: anonymity (used the father’s name for the sister, which is a deep concession). The cost-bearing line: “You came to the phone and I said the thing I drove three states to not say.” The kettle line carries the sister’s parallel cost without describing her emotion (she let it scream).

It is the same brief, the same forge pipeline, the same model. The gate changed what the model was required to deliver.

Why the gate works

The Cost Audit gate works for the same reason the broader Pattison discipline works: it shifts the model’s attention from the feeling to the trade. When the model is forced to enumerate the trade before delivering the lyric, the resulting lyric carries the trade in its surface. When the model isn’t forced to enumerate, it averages.

This is the smallest possible engineering change with the largest possible craft outcome we’ve found in 18 months of forge-pipeline work. The Cost Audit gate is roughly 40 tokens of additional prompt instruction. The composite-score delta on matched briefs is +6.2 points (measured across 240 matched-brief forge pairs, B2292 vs. B2293).

It works because cost is the structural foundation of every song that anyone remembers a year after first hearing it. The lyric that doesn’t carry cost doesn’t survive the year. The lyric that does, does.

How to use this without our tool

If you write with any AI tool, add this to your prompts:

Before writing the lyric, decide:
(1) What did the narrator give up to be in this moment?
(2) What did they refuse to give up?
(3) Which specific line in the lyric will carry that trade?
Make the trade visible in the lyric. Do not tell the listener about the trade. Show the trade in the surface of the lines.

This is approximately what our gate does internally, formulated as a user-facing prompt addition. It will work with ChatGPT, Claude, Gemini, or any tool that accepts free-form input. The output quality will improve on the next forge.

If you write without an AI tool, the cost question is even more central. The blank page does not know what the narrator should give up. You do. Spend the first five minutes of every writing session deciding what the narrator gave up to be in this lyric, before writing a single line. Then write the lyric. The lines will carry the trade because you decided what the trade was.

The cost question is the single most reliable craft heuristic we’ve found. It separates the lyrics that get remembered from the lyrics that get heard once.

Want to know whether your draft passes the Cost Audit? Run it through the Crucible — the 8-voice adversarial panel will flag a mood-only lyric inside the first two voices. If the panel doesn’t flag it, your trade is visible. If it does, you know which line to rewrite.

Why your AI lyric sounds like mood, not song — and the one question that fixes it