Keep a Prompt Library, or Keep Rewriting the Same Sentence

For about two months I wrote every prompt from scratch. I would get a genuinely good result, feel pleased, close the tab, and a week later reconstruct roughly the same sentence badly from memory. I knew this was stupid at the time and did it anyway, because writing it down felt like overhead for a thing that takes forty seconds.

The file that fixed it is a plain note with no structure to speak of, and it improved my results more than any technique I have read about. Here is what is actually in it and why.

Why history is not a library

The obvious objection is that the app already keeps every generation, and it does. That covers the wrong problem.

History tells you what you asked for. It does not tell you which of eleven near-identical attempts was the one that worked, what you had been trying to fix, or which specific phrase was doing the work. Scrolling back through it three weeks later, everything looks the same and the reasoning is gone. It is a record. A library is a record plus the conclusions.

The four lines

Every entry in mine is four lines. I tried more elaborate schemes first, all of which I stopped maintaining within a week, which is the real test of any system.

• The prompt, exactly as sent. Not cleaned up, not shortened. The version that produced the thing, typos included.
• One line on what came out.“Good, slightly too warm.” “Face fine, hands wrong.” Enough to know whether to reach for it again without regenerating to find out.
• The fragment that did the work. This is the important one and I will come back to it.
• What I would change. Written at the moment I still remembered, which is the only moment I ever know.

No tags, no categories, no dates. I search the file with the browser’s find function and that has never once failed me. Every categorisation scheme I invented required a decision at save time, and a decision at save time is how a system dies.

The fragment line is the whole point

This is the part I would have skipped if someone had described this to me, and it is the reason the file works.

A long prompt is mostly filler. Somewhere in it are five or six words that actually changed the output, and everything else is describing what would have happened anyway. Isolating that fragment turns a prompt you can reuse for one image into a phrase you can reuse everywhere.

Mine are things like “hard directional light from a low angle”, “shot on a long lens, background compressed”, “matte skin, visible pores”. None of those belong to a particular image. They are the ones that keep reappearing, and after a while you are not writing prompts so much as assembling them. That is roughly what the twelve-rewrites piece was circling around without naming it.

What not to bother saving

The file gets useless if you put everything in it. Three things I stopped saving:

Prompts that merely worked. A perfectly good result from an obvious prompt teaches nothing. If I could rewrite it in thirty seconds from the idea alone, it does not go in.

Long lists of style words. I had a section of these, copied from various guides, and I never once used it. Style words without a scene attached are not reusable, they are just vocabulary I already had.

Most failures. With one exception below, a file of failures is unreadable and nobody reads it, including the person who wrote it.

The exception: failures that repeat

Worth its own short section because it has saved me more time than the successes.

Some phrasings backfire reliably, and you will rediscover each one three or four times before the pattern registers. Mine include the general principle that naming something you do not want tends to produce it, and that stacking three style references in one prompt produces something that commits to none of them.

Three lines recording “this does not work and here is what happens instead” is worth more than fifty saved good prompts, because the good ones you would eventually reconstruct and the bad ones you would eventually repeat.

Where the file lives

Somewhere you can paste from on a phone, because most of my generating happens there now and retyping a long prompt with a thumb is exactly the friction that stops people iterating. A synced note app is fine. The requirement is that it opens in two seconds, not that it is well designed.

One caution worth stating: a prompt file is a record of what you have been making, in plain text, wherever you put it. If any of it is private, put it somewhere you would be comfortable having that record, which for most people means not a shared workspace note. The privacy piece covers the platform side; this part is entirely on you.

What actually changed

Two things, and neither is the one I expected. My prompts got shorter, because I could see which parts had never earned their place. And I started iterating properly, because opening a note and changing one word is a completely different act from writing a paragraph again.

Twenty entries is roughly where it starts paying, which is a week or two of ordinary use. If you are generating in Fellowi Images without keeping one, that is the single highest return change available to you, and the prompt examples are a reasonable set of first entries to steal.