I wrote a 400 word prompt once. Beautiful thing. Lighting, camera, wardrobe, pose, mood, film stock. Hit generate. Got back something that looked like it read the first two sentences and went to lunch.
That is exactly what happened. The tool ate the front of my prompt and dropped the rest on the floor. No warning. No error. Just a quiet amputation.
This is the stuff nobody explains, so here it is. Pure mechanics.
Character caps are real and they are silent
Almost every NSFW image tool has a hard cap on how much text it will accept. Sometimes it is enforced in the text box, so you literally cannot type past it. More often the box takes everything you paste and the backend trims it.
Practical ranges I have seen across the tools I test, and these move constantly, so treat them as shape and not gospel: short persona and caption fields around 200 to 500 characters. Standard image prompt boxes around 500 to 1,000. Generous ones in the 1,500 to 2,000 range. Free tiers are frequently capped tighter than paid tiers on the same platform, which is why the tail of your prompt vanishes on a free account and survives after you upgrade.
The important part is what happens at the cap. Very few tools reject an over-long prompt. They truncate it. So the last third of your prompt, the part where you put all the good detail because you were building up to it, never reaches the model at all.
Front-load. Always. Subject, body, action, then setting, then style, then camera and lighting garnish at the end where it is safe to lose.
The first words carry the weight
Even inside the cap, position matters. Text-to-image models do not treat your prompt as a checklist of equal items. Early tokens get more influence over composition. Later tokens fight over scraps of attention.
You can prove this to yourself in ten minutes. Take one prompt on CreatePorn and generate it twice, once with “redhead” first and once with “redhead” last. The first version gives you a redhead. The second gives you a coin flip.
So ordering is a lever, not a formality. If a detail is not landing, move it up. That single move fixes more failed generations than any magic keyword ever will.
Word filters are dumb on purpose
Every tool ships a blocklist. Promptchan, PornWorks, AiNudez, all of them. And the filters are almost always naive substring matchers, not clever language models.
Which means an innocent word gets caught because a banned word lives inside it. The classic is body-size vocabulary. Words describing a small or slight frame get blocked outright because those same words are used to describe minors, so the platform bans the whole string rather than trying to judge intent. Same with anything schoolish, anything youthful, anything in the age-adjacent family.
To be blunt: those filters are correct and I am glad they are aggressive. Anything involving minors is not a gray area, it is not a prompting problem to solve, and MadePrompt refuses that content too. If your prompt is tripping a filter for that reason, the answer is to write about adults, not to find a workaround.
The credit question is where tools differ and where it stings. Some refund the credit on a refused generation. Some log the refusal and charge you anyway. Some just burn it and show a generic error. Check the fine print before you buy a big credit pack, and check the per-tool notes on the rankings page.
Three prompt dialects, one prompt
This is the biggest reason a prompt that sings on one tool flops on another. They do not speak the same language.
- Booru tag engines. Comma separated tags, no grammar, order is weight. Anime and hentai models live here, and they want quality and score tags at the front. Leaving those off is why your hentai output looks flat and cheap.
- Natural sentence engines. Full descriptive sentences, and tag soup reads as noise. Photoreal platforms lean this way.
- Short persona fields. Chat and companion tools like CandyAI, SpicyChat and GPTGirlfriend want a compact character sheet, not a scene. Traits, voice, boundaries. Stuff a lighting setup in there and it gets ignored or leaks into dialogue.
- Companion memory is a published spec, not a mystery. Nectar AI sells 8K of context on its entry plan and 16K one tier up, which is somewhere around six thousand words of conversation before the earliest lines start falling off the back. If your character keeps forgetting, check what you bought before you rewrite the prompt.
Feed a sentence prompt to a tag engine and you get mush. Feed tag soup to a sentence engine and you get a mood board. Know which one you are standing in front of. The NSFW prompt dictionary lists the working vocabulary for each dialect, which saves a lot of guessing.
Negative prompts, and what to do without one
Some tools expose a proper negative box. Some bury it behind an advanced toggle. Some, especially the one-click ones, have none at all.
No negative box? Then say the positive instead. Do not write “no extra fingers,” write “hands behind back.” Do not fight a bad background, replace it with a specific one. Composition beats exclusion when exclusion is not available.
Caps you will hit that are not text
Resolution is usually locked by tier. Free accounts get a smaller base render with upscaling behind the paywall. Aspect ratios are often a short preset list rather than free numbers, and picking a wild ratio can wreck anatomy because the model was trained near square.
Batch size is tier-gated too. And queues are real. Evenings and weekends, US and Europe overlapping, is when a two second render becomes a ninety second render on tools like DeepMode or OurDream. Character consistency also drifts more than the marketing suggests. Locking a seed or a saved character helps, when the tool offers one.
The checklist
- Front-load: subject and body first, garnish last.
- Keep it under roughly 500 characters if you want it to survive anywhere.
- Strip filter bait. Write adults, clearly.
- Match the dialect: tags, sentences or persona sheet.
- Turn negatives into positives when there is no negative box.
- Pick a preset aspect ratio, not a weird one.
- Test off-peak before you judge a tool.
If ordering it by hand sounds tedious, that is what our free prompt generator is for. It builds the prompt in the right order, keeps it tight enough to survive a cap, and skips the vocabulary that gets you refused. Steal from the community library too, or read a reference image back into words with the image to prompt tool and clean up what comes out.
Your prompt was probably fine. The tool just never read all of it.
