Case study
Why your AI tool ignores the end of your description
A video team's AI-made presenter photos kept coming back in the wrong place and clothes, with no error on screen. The cause was a hidden limit that cut the end of longer descriptions, and the fix was live within a day.
Key takeaways
- A tool that accepts your text has not necessarily read all of it.
- Fixed house rules inside an AI tool can use up most of the room before your words get any.
- Put the place, the outfit and the mood at the start of a description, not the end.
- When a result misses a detail, first ask whether that detail was near the end of what you wrote.
- A limit on what a tool reads should be shown where you type, together with the words it will drop.
A video team at a client in the United Arab Emirates describes the presenter for each advert in a short box: who the person is, what they wear, where they stand. For longer descriptions the AI-made presenter photos came back wrong, with the person in an ordinary living room instead of the place written and in different clothes, while the tool itself showed no error. The fix was live within a day of the first wrong photos, and it puts that limit in front of the team as they type.
A tool that accepts your words and then ignores them
A description can be accepted without complaint and still be only partly read. The result is a picture that looks confident and is wrong.
The team's longer descriptions came back with the wrong place and the wrong clothes, and nothing on screen had said anything was missing.
That matters more than a wasted picture. The presenter photo is the base of the whole advert: a wrong place or outfit means another round of photos, or an advert that does not match its own script. A tool that drops words in silence also teaches a team to stop relying on it.
Why an AI tool ignores part of your prompt
An AI tool usually sends your words to the model, the part that actually makes the picture, inside a fixed amount of text, and that size is not shown to you.
The tool's own rules for a believable photo, covering lighting, framing and a realistic person, are written into that same block and take most of the room before your description gets any.
What is left is filled from the start of your text, so the end is what gets cut. The details people write last, the place and the outfit, are the first to disappear, and the tool gives no warning.
What our engineers found
Our engineer compared what the team typed with the instruction the platform actually sent to the image model, and for a typical longer description only about the first half arrived.
Most of the instruction was already taken by the platform's fixed rules for a good presenter photo. A further rule made it worse: when a description did not name a place the platform recognised, it added a default setting, an ordinary bright home, which used up even more room.
The kind of place the team wrote was not on the platform's list. So the photo got the default home, and the words after it were lost. When the place sat only in the part that was cut off, both went: no place from the team and no default either. The image model was never ignoring the detail. It had not been given it.
What we changed
We made four changes on the client's platform.
- The platform now recognises many more kinds of place: sports and fitness places, shops, salons, schools, workshops, outdoors and others. Ordinary phrases that only look like a place are not mistaken for one.
- The default setting is now decided on the words that are really sent. If the place the team wrote was cut off, the setting goes back in, so a photo is never left with no place at all.
- The photo dialog and the advert's brief form both show, as the team types, how much of the description the photo takes, how long this one is, and, when it is too long, which words will be left out.
- When the cut part holds the place the team wrote, the note names it first in the team's own words and says the advert's setting will be used instead.
Before going live, we compared the instructions sent before and after across well over a hundred thousand combinations of descriptions, looks and settings. Every description that fits is sent exactly as before; only the cut ones that had lost their place changed.
What the team has now
The change is live. The team can see, before making any photos, whether a description fits and what would be dropped, so they can shorten it or move the place and outfit to the front. A description that is cut no longer loses its setting along with it.
What you can check yourself
- Look at the box where you type. Is there a counter or a warning showing how much of your text the tool will use? A tool with a limit should show it there.
- Write the same description twice, once short and once long with the key detail at the end, and compare the two results.
- Put the place, the clothing and the mood at the start of any description you give an AI tool, and keep the rest for after.
- When a result misses a detail, read your own text again and see where that detail sat. If it was near the end, you have probably found the cause.
- Ask your tool's provider whether there is a limit on what reaches the model, and what happens to the words past it. Get the answer in writing if you can.
What to do next
Start with the description you use most and move its essential details to the front. Then look for a tool that tells you where your words stop counting; if it does not, ask the provider, and treat any long description as one whose end may not arrive.
Making a limit visible, and saying which words will be dropped, is the kind of small piece of work our engineers do on the tools clients use every day.
Questions people ask
Why did the AI tool accept my description but ignore the details at the end?
Because the tool forwards a fixed amount of text to the model, and its own rules for a good result take most of that room. Whatever is left is filled from the start of your text, so the end is cut without a warning. Nothing on screen looks wrong, which is why the picture is the first sign.
Is there a character limit on what I type into AI image and video tools?
Often there is a limit, but it is not usually on how much you type. It is on how much of your text is passed on to the model. The box can accept the whole description and send only part of it, so ask your provider what happens to the words past that point and whether the tool shows where the cut falls.
How can I tell whether my AI tool read all of my description?
Make the same image twice from the same text, with the key detail moved from the end to the beginning, and compare. If the detail appears only when it comes first, the tool is dropping the end of your text. A counter or warning near the box, where a tool has one, tells you before you press anything.
What should I put first when writing a description for an AI tool?
The details the picture cannot do without: the place, the clothing, the mood and the framing. Treat the opening line as the part that will certainly be read. Put background and refinements after it, and accept that some of them may not arrive.
Does a longer description give better AI results?
Not automatically. A longer description carries more detail, but it also carries more risk that the end is cut, and a cut description can lose the very setting you cared about. A short description with the essentials first is usually the safer choice, and it makes it easier to see what the tool actually used.