I Started Keeping the Prompts That Failed

I used to save only the prompts that worked.

If a generation looked good, I'd copy the prompt into a text file and give it a name like:

camera-good.txt
night-scene-final.txt
character-motion-v2.txt

The failed ones usually disappeared.

At the time that felt sensible. Why keep something that didn't work?

A few weeks later, though, I'd open one of those "good" prompts and have no idea why half the lines were there.

For example:

Keep the subject centered.
Maintain facial identity.
Avoid sudden camera movement.

I knew those lines had probably fixed something at some point.

I just couldn't remember what.

So I changed one small habit: I stopped deleting the bad versions.

My folders are messier now.

A recent one looked like this:

01.txt
02.txt
03-face-changed.txt
04-camera-static.txt
05-better.txt
06-too-fast.txt
07.txt

Not elegant, but at least I can follow what happened.

I usually leave a short note for each version:

03
face changes during turn

04
face is better, camera barely moves

05
camera works, walking too fast

That's normally enough.

I don't score every result. I tried doing that once and stopped after a few days because I was spending too much time deciding whether something deserved a 7 or an 8.

One camera test is a good example of why I keep the old versions now.

I started with:

Camera slowly moves closer while following the subject.

The movement wasn't what I wanted, so I made the instruction more specific:

Smooth cinematic tracking shot, gradually moving
toward the subject while maintaining stable framing.

It got worse.

Then I tried:

Slow forward tracking shot.

That version behaved better.

I wouldn't turn this into a rule about short prompts. I've had other scenes where adding detail helped.

The only thing I wrote in my notes was:

longer camera description didn't help this scene

A month later, that note was more useful than trying to remember why the final prompt was short.

Old prompts became more useful when I changed models

I ran into the same thing again while going back through old prompts during a MiniMax H3 test.

My normal habit had been to take the best prompt from the previous model and use that as the starting point.

It sounds reasonable, but there's a problem with it.

A successful prompt often contains fixes for problems that belonged to the model I was using at the time.

I found a very simple example in an older folder:

v1
Woman turns toward the window.

result
body turns, face changes

Then:

v2
Woman turns toward the window.
Keep the same facial features during the turn.

result
face is more stable

The second line stayed in my prompts for months.

I copied it into other scenes, then into other projects.

At some point I couldn't remember which model had originally needed it.

When I tested the old v1 again with a different workflow, the face stayed reasonably consistent without the extra instruction.

So the sentence I'd been treating as part of my normal prompt wasn't really a rule. It was an old patch.

I started noticing more of these.

Keep camera distance consistent.

Preserve clothing throughout the scene.

Maintain natural body proportions.

Do not change the background layout.

Some are still useful.

Others are probably there because something went wrong in a generation months ago.

If I only keep the successful prompt, I lose that history.

Keeping the version immediately before it tells me why a line was added.

That's also why I don't clean up my test folders as aggressively anymore.

I did try replacing all of this with a spreadsheet once.

It had columns for:

model
prompt
settings
duration
failure
score
notes

I abandoned it fairly quickly.

For short experiments, it felt like more work than the testing itself.

Now I normally keep a folder and a small notes.txt file:

01 - baseline
camera static

02 - added tracking
camera moves, face okay

03 - added cinematic lighting
can't see much difference

04 - changed walking speed
too fast

05 - reduced walking speed
usable

This isn't particularly organized, but I can open the folder a few weeks later and understand what I was doing.

It also makes me a little less likely to rewrite everything after one bad result.

If 04 doesn't work, I usually want 05 to differ in one obvious way. Otherwise I end up with two completely different prompts and no idea which change mattered.

Not every test works like that. Sometimes the result is bad enough that I throw out the prompt and start again.

I just don't delete the old one anymore.

I've also stopped naming files like:

final.txt
final2.txt
final-new.txt
final-really-final.txt

I had enough versions of those that "final" stopped meaning anything.

Numbering them is boring, but it works.

These days I usually keep the whole experiment folder. Text files take almost no space, and three weeks later I rarely remember why 04.txt was bad without the note sitting next to it.