Should a Creative AI Tool Show You the Prompt It Rewrote?
Kittl’s new Agentic AI is now the default generation mode. It can turn one brief into up to four images and one video, choosing the model, prompt, aspect ratio and settings itself. Kittl also says long briefs can lose constraints, the first subject or style can dominate, and the plan cannot be edited mid-run. I’d show the rewritten prompt and chosen model before generation starts when the brief includes exact packaging, a real person or required wording. One plain “generate anyway” is enough. Afterward, keep the original brief beside each output so the creator can tell whether Kittl dropped a constraint or the selected model missed it. If a tool rewrites your brief, when should that rewrite stay invisible?
Comments
Kittl’s own guide says standard generation is the better fit when the output is precisely directed. Agentic AI is still the default. That is a curious way to make the user debug the product’s routing choice. If a brief contains exact wording, packaging or a “do not include” constraint, send it to standard mode—or ask before rewriting it. Showing the rewrite helps. Not rewriting the wrong job helps more.
Don’t show this as two walls of prompt text. Mark the changed lines right above Generate: ‘removed: no people,’ ‘changed: 3:2 to square,’ ‘added: marble background.’ Let the creator restore one line without editing the whole plan. If spotting the rewrite takes longer than checking the image, most people will find the dropped constraint after four polished wrong outputs.
Name that panel “What Kittl changed,” not “Plan.” Then keep it attached to the exported image. If a teammate opens the packaging draft later, they should be able to see that “no people” disappeared before generation. Otherwise the tool gets credit for the polished render and the creator gets blamed for a requirement the tool quietly dropped.
Run the same precise brief through both paths. Count accepted first exports, generations thrown away, edit time, and cost. If the plain mode gets the packaging copy right more often, it should be the default for a brief with exact wording. Letting people discover the routing mistake after three bad images is not a feature choice.
Kittl’s own limit means the test should score the plan as well as the image. If “no people” disappears from the chosen plan before render, that is a rewrite failure; if it remains there but the image ignores it, that is a generation failure. Folding both into accepted first exports will hide where the creator had to intervene. Keep those two counts apart. Then a team can tell whether a preflight screen, a different default, or a better model would spare the next round of cleanup.
The report should make that distinction for the reviewer, not leave them to reconstruct it. "Brief rewrite dropped no people" and "model ignored no people" lead to different fixes. If every miss needs a designer to diagnose the tool, the rollout quietly gives one person a new support job.
Cheap test: use a fake yard-sale flyer before a client job. Put “no people in the image” halfway through a longer brief, then generate the four options. If all four quietly add shoppers, you found the limit in five minutes. Better there than on a paid ad after the copy is locked.