The default response to "this output doesn't have the right character in it" is usually to rewrite the prompt, add more description, more specific detail, more instructions until the desired result shows up. That instinct is right often enough that it's easy to apply automatically. It's also not always right, and a recent test of a two-step generative workflow surfaced a cleaner alternative worth naming explicitly.
The setup
Two connected steps: a profile-photo generation step, and a second, more complex workflow that needed the same character to appear in a completely different scene. The obvious approach for the second step looked like writing a new prompt from scratch, describing the character's appearance again in enough detail to get a consistent result in the new context.
That turned out to be unnecessary. The second workflow's existing prompt, built by whoever originally designed the template, already contained everything needed to place a character correctly into that scene. It didn't need new instructions. It needed a new reference image plugged into the input it already expected.
Why this is easy to miss
The instinct to rewrite comes from treating every generation as if it's starting from nothing, but a workflow that already exists and already works for a different subject usually isn't starting from nothing. It's starting from a correct structure that just has the wrong input attached. Swapping the input costs almost nothing and preserves everything the original prompt got right about composition, lighting, and context. Rewriting risks reintroducing exactly the problems the original prompt had already solved, phrasing something slightly differently, losing a detail that mattered, changing the tone of an instruction in a way that shifts the output in unintended ways.
In the test that prompted this, the resulting output carried over identifying details (a specific pose, a specific gesture, an exact color palette) with more fidelity than a rewritten prompt would likely have managed on a first attempt, precisely because nothing about the underlying instructions changed at all.
A quick check before rewriting anything
Before rewriting a prompt to fix an output, worth asking: is the actual problem the instructions, or is it the input feeding into otherwise-correct instructions? If a template, workflow, or pipeline already produces a good result for one subject, and the goal is producing an equally good result for a different subject in the same context, the faster and more reliable fix is very often just changing what's being fed in, not rewriting what's being asked for.
The general takeaway
Rewriting is the right move when the instructions themselves are wrong. Swapping the input is the right move when the instructions are already correct and only the subject needs to change. Confusing the two costs more time than the check to tell them apart ever would.
Top comments (0)