Eight Things I Stopped Doing With These Tools

Almost everything I learned about using these tools well turned out to be subtraction. Not new techniques, just noticing that some habit I had picked up in week one was doing nothing, or worse than nothing.

Eight of them, roughly in the order I dropped them.

1. Stacking quality words

My early prompts ended in a comma-separated tail of masterpiece, highly detailed, ultra realistic, 8k, award winning. It felt like insurance.

It does very little. Those words describe how good I would like the result to be, and the model has no way to act on that beyond a nudge you would struggle to see in a blind comparison. The words that actually move the output describe the thing: the light, the lens, the surface, the moment. I now spend those same twenty words on the scene and the results are better, which is the entire argument.

2. Rerolling instead of editing

The bad loop: result is not right, press generate again, still not right, press again. Three attempts, same prompt, no new information.

Rerolling tests randomness. Editing tests your idea. There is a narrow case for a second roll, when the prompt is good and one specific attempt was unlucky, but I was using it as the default and it is a way of avoiding the harder question of what is actually wrong with the description. The post about rewriting one prompt twelve times is the long version of this lesson.

3. Saying what I did not want

This one is counterintuitive enough that I had to hit it several times before it stuck. Adding a clause to exclude something tends to make that thing more likely, not less. The description contains the word, and the word is in the frame.

The fix is to describe what should be there. If a hand keeps ending up in shot, say where the hands are, do not ask for them to be absent. Positive descriptions of the intended state consistently beat negative descriptions of the unwanted one.

4. Generating at maximum quality by default

This is the one that cost money. For weeks I used the higher-quality setting for everything, on the reasoning that better is better.

Higher settings mostly change fine texture. On a draft, on a mock-up, on anything that will be looked at on a phone, that difference is invisible and you have paid extra for it. I now draft at Standard, 40 coins, and only step up when a specific image is going somewhere the detail matters. Over a month that was a substantial share of my spend on nothing. The costs piece has the arithmetic.

5. Trying to fix a bad idea with a better prompt

Some images are not going to work and the prompt is not the problem. A composition that requires three specific things happening in a precise spatial relationship is difficult in a way that no amount of description fixes.

The signal is a run of attempts that fail in different ways each time. Consistent failure means the prompt is wrong; varied failure usually means the request is too complicated. Simplify the idea or split it, do not keep rewording it.

6. Judging results on a phone screen

A small screen flatters everything. Images I was pleased with had visible problems the moment I opened them on a laptop, and I had already moved on by then.

I still do most of my generating on a phone, which I think is right. But anything that is actually going to be used gets one look at a proper size before I call it done.

7. Starting from scratch every time

I would write a genuinely good prompt, get a good result, close the tab, and a week later rebuild the same sentence badly from memory. Every time.

A note file fixed this and it was the biggest single improvement in my results, which is a slightly humbling thing to report given that it is not a skill. The prompt library post is the whole method, and it takes about a minute to start.

8. Treating a failed generation as my fault

Small one, but it changed how I work. Early on, every unusable result sent me back to redraft the prompt, as though the tool were a lock and I had produced the wrong key.

Some proportion of generations simply come out wrong from a good prompt. Knowing which is which is the actual skill, and the giveaway is whether the failure is in the thing I described or somewhere I never mentioned. Wrong mood means my description; three-fingered hand means regenerate and move on.

What replaced all of it

Shorter prompts, more specific about fewer things. Editing rather than rolling. Drafting cheap and stepping up deliberately. And a note file, which of everything here is the one I would tell someone to do on day one.

None of this is technique in the sense the guides mean. It is mostly noticing what you are doing out of habit and asking whether it earned its place. If you are somewhere in your own first month in Fellowi Images, the first-week account is where most of these habits started for me, and the prompt examples are what I wish I had read instead of inventing my own quality-word tail.