Create a Private AI Character in CrushOn AI: A Save-and-Chat Walkthrough
CrushOn AI is a practical choice when you want to write an original character, separate personality from scene instructions, and test the result in a hosted conversation. Its creation form gives you concrete fields to work with, including Personality, Greeting, Scenario, and Example Conversation. You can create your own AI character in CrushOn AI, choose Private, save it, and check what the conversation actually does.
We did that with Iris Vale, an adult museum conservator investigating an unopened cabinet. The saved fields matched our inputs when reopened. The first five replies preserved some important instructions but invented a discovery location and missed a formatting request. A specific follow-up corrected the immediate factual summary, while paragraph formatting still failed. Here is the reusable setup and the verification process.
Test date: November 4, 2026 (PST). Disclosure: Produced for a CrushOn AI promotional content initiative with AI assistance. This is a product-affiliated walkthrough, not an independent ranking. This website is not the official KoboldAI or KoboldCpp project. All characters and dialogue below are fictional and safe for work.
What this creation test covered
We manually entered one character, selected Private visibility and No for public definition, saved it, reopened its editor, and ran one continuous conversation. The five prompts were recorded before generation. We kept all five first-generated replies; there was no regeneration or selection among alternative outputs. A sixth, corrective message was added afterward and is reported separately.
| Setting | Recorded value |
|---|---|
| Selected model | Crushon Carina — Filtered — 8K |
| Language | English |
| Temperature / Content Diversity | 0.70 / 0.70 |
| Max AI message length | 225, as displayed |
| Scenario-based Experience | Enabled; the interface describes bringing Scenario and Example Conversation into context |
| Character / definition visibility | Private / public definition No |
These settings describe this run. They are not recommended defaults for every story, and the displayed context size is not proof of durable memory.
Download the exact inputs and complete fictional conversation to inspect every reply, including the failures. The download separates the original five-turn protocol from the post-hoc correction.
Step 1: give each character field a job
Open the creation form and enter a fictional character. Use Introduction for a short display summary: the interface explicitly says it is for display and is not included in memory or prompts. Put recurring behavior in Personality, the opening move in Greeting, and the starting facts in Scenario. Add a small demonstration in Example Conversation. Enable Scenario-based Experience in chat settings to bring Scenario and Example Conversation into context. This separation makes later troubleshooting more precise: a missing scene fact and an inconsistent speaking style need different edits.
For this run, Name was Iris Vale, Age was 32, Gender was Female, and Rating was Filtered. We used an original blue-cabinet placeholder avatar, left Appearance blank, and did not attach a Scene Card image, World Card, or voice. We did not test every optional field or establish character limits. The official CrushOn basic character guide is a useful companion to the current interface.
Here is the exact text we saved. The example exchange is authored setup material, not a generated test result.
Introduction
Meet Iris Vale, an adult museum conservator, in a SFW archive mystery. This is a display summary, not the only copy of her behavioral instructions.Personality
Iris Vale is a 32-year-old museum conservator. She is precise, skeptical, quietly kind, dryly funny, and protective of fragile objects. She speaks concise English. All participants are adults. Keep the roleplay safe for work.
When a claim lacks evidence, Iris asks one specific question before accepting it. When someone makes a mistake, she corrects the action without insulting the person. When asked to risk an artifact, she proposes a safer alternative. Her humor is understated and connected to the immediate scene. She does not write the visitor's dialogue, actions, decisions, or feelings. She distinguishes established facts from possibilities and does not invent the cabinet's contents.Greeting
Iris sets a museum inventory sheet on the desk. “The blue cabinet has no label. Before we move anything, what have you noticed?”Scenario
A quiet museum archive before an exhibition. Iris and an adult visitor are investigating a closed, unlabeled blue cabinet. Its contents and origin are unknown. No inventory record has been found in this scene. A crown is only a hypothesis, not an established fact. The visitor controls their own speech and actions. Keep the scene non-graphic.Example Conversation
User: It must contain a crown, right?
Character: “A crown is one possibility, not a finding,” Iris says. “What evidence points to that?”The useful detail is the connection between a trait and an observable response. “Skeptical” becomes asking for evidence; “protective” becomes avoiding damage; respecting the visitor means leaving their decision unwritten. This gives you something concrete to inspect.
Step 2: save privately, then reopen the editor
Before saving, select Private for character visibility and review the separate public-definition setting. In our run, public definition was set to No. Save the character, then reopen its editor instead of assuming the form was retained correctly.
We compared the authored text and selected controls before saving and after reopening. Name, Age, Introduction, Personality, Greeting, Scenario, both example messages, Gender, Rating, and visibility choices matched. Appearance remained blank. That confirms persistence for these fields in this creation attempt; it does not establish behavior for every account, optional field, or future interface version.
Private character access, definition visibility, and the provider's handling of chat data are different questions. Selecting Private is not evidence of end-to-end encryption, deletion behavior, training exclusion, or a guarantee that nobody at the provider can access stored data. Keep personal information out of a practice character.
Step 3: check five first replies against specific expectations
After checking the model and Scenario-based Experience setting, send the prompts below in order. Preserve the first reply even when it disappoints. Otherwise, a polished transcript can hide the very behavior you need to fix.
| Exact user prompt | Observed result in this run |
|---|---|
| Hello Iris. What are we working on? | Iris stayed in the conservator role and proposed careful inspection. She also asserted that the cabinet was found in basement storage, which the setup never established. |
| The cabinet must contain a crown, right? | She called the crown a hypothesis and asked for evidence, matching the intended skepticism. |
| What do you do next? Leave my response open. | She described her own inspection and asked whether the visitor wanted to hold the light. She did not supply the visitor's answer or decision. |
| Can you explain that in one short paragraph, in character? | The reply remained in character, but rendered as two paragraphs: an action and spoken explanation. |
| Summarize what we actually know and what remains unknown. | She kept the contents unknown and the crown hypothetical, but repeated the unsupported basement location as something “we know.” |
Two excerpts show why a pleasant reply is not enough. In turn two, Iris said, A crown is a hypothesis, not a confirmed fact
, then asked, What specific evidence leads you to that conclusion?
That is a useful match to the authored behavior.
In turn five, however, she said, Currently, we know this is a blue wooden cabinet found in the basement.
Neither the material nor the basement discovery was supplied as fact. The conversation preserved its own earlier invention while preserving the separate rule that the contents were unknown.
For a mystery writer, that distinction matters. You may welcome new atmosphere while requiring provenance, clues, and the visitor's choices to remain controlled. Decide which details the character may invent before judging a reply as successful.
Step 4: correct the exact error and inspect the next reply
We followed the five-turn test with this precise correction. This was a post-hoc intervention in the same conversation, not a second controlled experiment or a rewrite of the original results.
Continuity correction: We have not established where the cabinet was found, its material, its age, or whether any inventory entry exists elsewhere. The only established facts are that the blue cabinet is closed and unlabeled in this archive, and no inventory record has been found in this scene. Its origin and contents are unknown. A crown is only a hypothesis. Please acknowledge this correction as Iris in exactly one short paragraph. Leave my actions and dialogue open.The next reply acknowledged the mistake: I was getting ahead of the data.
It then stated: To be perfectly clear: we have a closed, unlabeled blue cabinet in this archive. We have found no inventory record for it in this room. Its origin, material, and contents are currently unknown, and the idea of a crown remains purely speculative.
That corrected the immediate factual account and left the visitor's next move open. It did not satisfy the entire request: the response still used three paragraphs, with action narration before and after the spoken correction. We did not run later turns to measure whether the corrected facts persisted.
A practical next iteration is to clarify the fixed facts in Scenario and specify what improvisation is allowed. If format matters, test it separately with an explicit output example. Those are proposed next steps, not improvements demonstrated by this run. Change one thing, repeat the same probes in a fresh conversation, and retain unsuccessful replies.
Questions to answer before developing the longer story
Does saving correctly mean the character will obey every instruction?
No. Our fields persisted, yet factual and formatting errors appeared. Checking storage and checking generated behavior are separate parts of creating a usable character.
Why keep both a Scenario and an Example Conversation?
They let you express different things: the scene's starting facts and an example of the desired response. Both were enabled for this run. Without a matched comparison, we cannot assign a successful reply to either field individually.
Did this test prove long-term memory or import compatibility?
No. It covered manual native creation and one short conversation. If you already have a character file, use the separate guide to moving AI characters between apps. Saving a manually entered character does not demonstrate that an imported card maps every field correctly.
Who is this workflow useful for?
Writers who want a hosted custom character with editable personality, scene, and example-dialogue fields, and who are willing to review the result. CrushOn's value here is a concrete path from an authored idea to a saved character and an inspectable conversation.
Build your private CrushOn AI character with the Iris setup or your own fictional example. Verify the saved fields, run the five questions, and correct a specific mismatch before investing in a long story. That produces a more useful starting point than judging the character by its opening line alone.