Debugging a Prompt
Locate the faulty layer before rewriting adjectives, whether input, constraint, model or randomness.
Related
Locate, then rewrite
When a generation misses, the reflex is to add adjectives. A faster move is to identify the layer that failed: the input itself, the way constraints are written, the model's capability, or plain randomness. Each layer needs a different repair; fixing the wrong one only lengthens the prompt.
Four layers
| Layer | Symptoms | Check first | Action |
|---|---|---|---|
| Input | Displaced lyrics, irrelevant reference | Text encoding, audio length, format | Use clean input and a shorter excerpt |
| Constraints | Tempo drift, wrong ensemble, missing sections | Conflicting rules, placement too late | Cut to three core rules, add back one by one |
| Model | Long form collapses, language unsupported, texture ceiling | Whether the task exceeds public capability | Generate in parts, change tool type or model |
| Randomness | Same prompt, very different results | Seed, version, candidate count | Fix the seed and sample several times |
Order of checks
1. Rerun the same prompt unchanged twice to see whether variance explains it. 2. Fix the seed and run again; does the fault persist. 3. Strip every modifier, keep tempo, ensemble and form, then add items back. 4. Swap in a shorter, cleaner input to rule out the material. 5. Try the same job in a different tool class to test the capability boundary. 6. Log each change and move one variable at a time.
The order controls cost. Changing models is expensive; adding adjectives is cheap and buries the cause deeper.
A log worth keeping
- Full prompt and version number
- Model name, version and date
- Seed and candidate count
- Input file and duration
- The single change made this round
- The fault and where it appears
- Whether it reproduces
Without a log, debugging becomes memory. Anything compared more than three times needs a sample you can replay, or the next round repeats a hypothesis already eliminated.
When to change the method
Long forms that keep collapsing, diction that never passes in the target language, exact harmony that will not land: these are not prompt problems. More words only crowd the constraints. Generate in parts, switch to a symbolic tool, or treat the output as raw material. Debugging ends in a change of method, not a longer prompt.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- Suno 与 Udio 官方帮助中心关于随机性、种子与版本行为的说明。
- 实验设计与对照测试的公开教材章节，关于单变量原则。
- 音乐制作质量控制惯例，关于分段验收与 A/B 对照。