How to Mention a Character in an AI Prompt
Chris · · 9 min read
Introduction
Mentioning a character used to mean typing out a description and hoping the wording matched whatever made a previous render look right. nocensor.ai now supports a more direct way to do it: type @ in the prompt box, and a picker opens that lets a saved character be dropped straight into the prompt as a chip. No trigger words to remember, no copy-pasted phrase from an old generation.
The feature is easy to demonstrate and slightly harder to explain correctly, because what actually happens under a mentioned character's chip is different from what most people assume. It is also easy to confuse with a completely different feature — face swap — which looks similar from a distance but works at the opposite end of the pipeline. What follows covers the mechanics of the @ picker, how a mention actually reaches the model, what happens when two characters share a prompt, and why the visible chip never shows the text that does the real work.
What Happens When You Type @ in the Prompt Box
Typing @ anywhere in the prompt composer opens a searchable grid of saved characters rather than a plain text-completion dropdown. Typing after the @ filters that grid by name in real time, and selecting an entry closes the picker and inserts a chip at the cursor position — a small pill showing the character's name and thumbnail, not raw text.
The grid respects context automatically. If a character is filtered by gender in the workspace, only matching characters appear in the @ results. If a character is still mid-training, it shows up tagged as locked, and a filter toggle for all/locked/unlocked only appears once at least one character is actually in that state — there is no point showing a filter with nothing to filter.
This differs from typing a description in one important way: nothing about the chip is negotiable text. A typed phrase like "a woman with red hair" can be edited word by word, but a character chip is atomic — it represents a specific trained identity, and the only edit available is removing it and picking a different one (or the same one again, which toggles it off).
A Mentioned Character Shapes the Render, a Face Model Only Replaces a Face After
This is the distinction that trips people up most, because two different nocensor.ai features both start with "put a specific identity into an image" and only one of them does it during generation.
A mentioned character is a trained LoRA. Including it in a prompt influences the diffusion process itself — the pose, framing, lighting and the character's face are all generated together, informed by the character's training data from the first denoising step onward. This is why a mentioned character can appear in any scene, any pose, any style the rest of the prompt describes, and still look like the same person: the model was trained on that identity, not pasted onto a finished image.
A face swap, covered in a separate guide, works the opposite way — it takes an image that already exists and replaces a face in it afterward. It cannot invent a pose or reshape a scene, because the scene is already rendered by the time it runs. A saved face model used for swapping is also a different kind of object entirely from a saved character: one exists to be applied after a render, the other exists to be described during one.
The practical upshot is that a mentioned character is the right tool when the goal is "generate this specific person doing something new," and a face swap is the right tool when the goal is "take this exact image and put a specific face into it." Reaching for one when the other is what the situation calls for is the single most common source of "why doesn't this look right" for both features.
Combining a Mentioned Character with Scene and Style Details
A character mention replaces the identity description in a prompt, not the whole prompt. Everything else — setting, lighting, pose, outfit, camera angle, art style — still needs to be written out the normal way, exactly as it would for a prompt with no character in it.
The cleanest results tend to come from prompts that put the mention where a description of the person would naturally sit, then build the rest of the scene around it: @Character standing on a rain-lit street at night, neon reflections, cinematic framing reads the same way to the model as writing out a full physical description in that spot would, except the identity itself came from training data instead of adjectives.
A mentioned character does not automatically dictate outfit, expression or setting beyond identity — those still come from whatever else the prompt describes, which means the same saved character can appear in a hundred completely different scenes without needing a hundred slightly different descriptions of what they look like. That reusability, more than the convenience of not retyping a description, is the actual value of the feature.
Mentioning Two Characters in the Same Image
A single prompt can mention up to two characters at once. Trying to add a third does not silently fail — the picker disables every unselected option once two are active and explains why, rather than letting a third mention get inserted and quietly ignored.
Two-character prompts are also where phrasing matters most, because the model has to decide which described action or position belongs to which mentioned identity. Naming an action right after the relevant mention, rather than describing both characters and then both actions in separate clauses, gives the model less room to mix up which trait belongs to which person. This mirrors exactly how two-person prompts have always needed to be written even without character mentions — mentioning by name (or by chip) instead of by physical description just gives the model a cleaner anchor to attach the rest of the sentence to.
There is no reason to expect worse quality from a two-character prompt purely because two identities are involved; the more common failure mode is a prompt structure that leaves the model guessing about pairing, which is a writing problem rather than a model limitation.
Why the Trigger Word Never Shows Up in the Editor
Every trained character carries a specific trigger phrase behind the scenes — the string of text a LoRA actually needs present in a prompt to activate. Older workflows across the industry expose that phrase directly, which means a prompt full of "real" text is actually full of opaque tokens nobody chose to write and can't safely edit.
nocensor.ai's composer keeps that phrase out of the visible editor entirely. Selecting a character inserts a clean, human-readable chip; the trigger word is resolved and injected only when the prompt is compiled for dispatch, invisibly, and never round-trips back into what a user can see or accidentally delete a word of. Saving or reloading a prompt with a character mention in it restores the chip, not a decayed fragment of trigger text — there is nothing in the visible state to decay in the first place.
The result is a prompt that stays readable no matter how many characters have been mentioned in it. A prompt with three edits and two mentioned characters still reads like a sentence, not like a sentence interrupted twice by strings that mean nothing to the person who wrote them.
Building a Character Worth Mentioning
A character has to exist and finish training before it shows up in the @ picker at all — mentioning is the consumption step, not the creation step. Building one happens from the character library, where a new character is trained from a set of reference images the same way a reusable face model is built, except the result is meant to be described in prompts rather than swapped onto finished renders.
A character that will be mentioned often benefits from the same input discipline that improves any trained identity: a spread of angles and lighting conditions in the training set produces an identity that holds up across the very different scenes a mention will eventually be dropped into. A character trained on five nearly identical photos will render a consistent face in a narrow set of poses and drift everywhere else; one trained on a genuinely varied set carries that resilience into every future mention.
While a character is training, it appears in the @ picker tagged as locked and cannot be selected — the filter toggle exists specifically so a workspace with several characters in progress at once can still find the ones that are actually ready to use.
Conclusion
A character mention is a shortcut for describing a specific identity, not a shortcut around the rest of prompt writing. The @ picker replaces a wall of physical-description text with a chip, keeps a LoRA's trigger phrase out of sight permanently, and — because it acts during generation rather than after — produces a character that can appear in any scene the rest of the prompt describes, which is the one thing a post-render face swap can never do.
The two features that look similar from a distance, mentioned characters and face swap, solve different problems and are worth keeping straight: one shapes a render from the first step, the other corrects a finished one. Picking the right one for the job, and writing the rest of the prompt with the same care a fully-described identity would need, is what makes the feature worth the two extra keystrokes it takes to trigger it.