Prompts that work across AI companions

The internet is full of "perfect prompt" posts that promise magic results on a specific platform. Most of them fall apart the moment you switch to a different AI companion app. The system prompt is different. The model is different. The memory architecture is different. What worked on Character AI might do nothing on Candy AI, and vice versa.

What does work, reliably, are a handful of prompt patterns that hold up across platforms because they exploit something deeper than platform-specific quirks. These patterns work on the underlying language models themselves, regardless of which app is wrapping them. Once you know what they are, you can drop into any AI companion and immediately steer conversations more effectively than most users ever do.

The patterns that actually transfer

Most platform-specific tricks are working around a particular memory architecture, filter, or interface limitation. Cross-platform patterns work on the actual language model, which means they keep working as you move between apps and even survive when platforms update their underlying tech.

The five patterns that hold up most reliably:

Direct, declarative framing of important information. State things plainly rather than mentioning them in passing.

Position-aware emphasis. Put the most important information at the start or end of long prompts, where the model attends most strongly.

Positive instructions over negative ones. Tell the model what to do rather than what not to do.

Concrete specifics over abstract directives. "Speaks in short sentences" beats "uses concise language."

Tone setting through example, not description. Show the model what voice you want; don't tell it.

Each of these works because of how language models are trained, not because of how a specific platform implements them. Let me walk through each one with examples.

Direct framing for important information

Models pay more attention to information that's framed as important than to the same information mentioned offhandedly. This is partly because of training data patterns (statements flagged as facts get reinforced more than incidental mentions in the source material) and partly because of how attention mechanisms work in practice.

The difference between weak framing and strong framing:

Weak: "I had a long day at the firm today, my boss is being impossible."

Strong: "Please remember that I work as a lawyer and my boss's name is Patricia."

Both convey the same information. The strong version explicitly flags the facts as worth remembering, which makes the model more likely to incorporate them into ongoing context. As covered in the post on AI companion memory, most platforms run extraction processes that look for clearly-marked facts. Direct framing makes those processes much more likely to catch the information.

This applies to every kind of information. Character traits ("Mira is fundamentally distrustful but values loyalty above all else") work better when stated directly than when implied through behavior. Setting details ("This story takes place in 1920s Shanghai during the warlord era") work better than letting the AI infer the setting from incidental details.

Position-aware emphasis

The "Lost in the Middle" research finding has a practical implication for prompt writing: if you put the most important instruction in the middle of a long prompt, the model attends to it less than if you put it at the start or end. The U-shaped attention curve that researchers identified means the start and end of context get more processing weight than the middle.

In practice:

When writing a long character description, put the most important traits first and last, with secondary details in the middle. The character's core identity should anchor both ends of the description. Their middle-name preferences and favorite foods can live in the middle.

When writing prompts with multiple instructions, lead with the most critical and end with a recap of the most critical. Don't trust the middle to carry essential information.

When asking for something specific, restate it at the end. "Help me understand X. Walk me through it step by step. As you explain, focus on X specifically." The repetition might feel awkward but the output will be more focused than a single statement.

This pattern works whether you're prompting on a chat platform with a long context, writing a character card, or composing the kind of multi-paragraph initial message AI roleplay sometimes calls for.

Positive instructions over negative ones

Models follow positive instructions more reliably than negative ones. "Write in short sentences" produces better results than "Don't write long sentences." "Stay in character" produces better results than "Don't break character." This isn't a perfect rule but it holds up across platforms.

The underlying reason is that models are next-token predictors. They generate text by extending what's there, not by checking against a list of prohibited patterns. A positive instruction gives them a target to hit. A negative instruction gives them an absence to avoid, which is harder to operationalize.

Examples of conversion:

Don't be cheerful → Maintain a serious, measured tone Don't break character → Stay in character throughout Don't decide what I do → Let me drive my own actions Don't rush the scene → Let the scene unfold gradually Don't be repetitive → Vary your phrasing and structure each response

Some negative instructions are unavoidable (the SillyTavern documentation acknowledges that "Do not decide what {{user}} says or does" is so common in roleplay prompts that it's effectively standard). When you do use negative instructions, follow them with a positive alternative. "Don't be cheerful. Maintain a serious tone. Be measured." The positive instruction does most of the work; the negative one prevents one specific failure mode.

Concrete specifics over abstract directives

Abstract instructions get interpreted inconsistently. Concrete specifics get followed reliably. Compare:

Abstract: "Use sophisticated language." Concrete: "Use vocabulary at the level of a New Yorker article. Vary sentence structure. Include occasional well-placed rare words but don't be ostentatious."

Abstract: "Be more emotional." Concrete: "Show internal reactions before speaking. Hesitate before answering hard questions. Let your replies vary in length depending on what the moment calls for."

Abstract: "Better pacing." Concrete: "Slow down physical scenes. Spend at least three exchanges on emotional setup before any major event. Don't summarize, scene-set."

The concrete version gives the model something to actually do. The abstract version gives it a vibe to gesture at, which produces wildly variable results.

This is particularly important when you're trying to fix something that isn't working. "I want this to feel different" doesn't help the model understand what to change. "I want shorter replies, more dialogue and less narration, and I want her to be more skeptical" gives the model three specific levers to pull.

Tone through example

The single most powerful way to set tone in an AI conversation isn't to describe the tone you want, it's to model it. The model picks up the style of what's already in the context and continues in that direction. If your messages are short and clipped, the AI tends to respond in short and clipped ways. If your messages are flowery and verbose, the AI tends to match.

This is why the first message in a roleplay matters more than people realize. The first message sets the template the AI tries to match throughout the conversation. A first message written in noir prose produces an entire conversation in noir prose. A first message written in light romance produces an entire conversation in light romance. The model is essentially learning your voice from the example you provide.

Practical applications:

When starting a new conversation where tone matters, write your first 2-3 messages deliberately in the voice you want. Don't worry about content as much as register. The AI will calibrate to it.

When you want to shift tone mid-conversation, demonstrate the new tone in your messages for several turns rather than instructing the AI to change. "Write in a more melancholy register" works less well than just writing your own messages in a more melancholy register and letting the AI catch up.

When character drift starts, the recovery move is often to write a few exemplary messages in the original voice, not to issue a corrective instruction. The model recalibrates to recent input more reliably than to abstract directives.

How these patterns combine

Real prompting in practice usually combines several patterns. Here's an example of building a strong message using multiple techniques:

Weak version: "Hey Mira, what are you doing tonight? Hope you're not too busy."

Strong version, combining direct framing, positive instruction, concrete tone: "I open the door to her room. The light is on but she's not at the desk. Mira? Where are you? I keep my voice casual but the tension is there if she listens for it."

The strong version doesn't just contain more content. It demonstrates the tone (third-person narration mixed with internal observation), models the kind of writing the AI should match, and uses present-tense action that establishes the scene before any dialogue happens. The AI's response is likely to follow the same template.

This is the deeper principle behind cross-platform prompting. You're not gaming the platform; you're shaping what the model has to work with. Whatever app you're on, however the platform implements memory or filtering, the model is still extending the patterns it sees in front of it. The patterns you put in front of it determine the patterns it produces back.

What about platform-specific features

Cross-platform patterns work everywhere, but most platforms have specific features that amplify the patterns when used right. Pinned memories on Character AI, lorebooks in SillyTavern, persistent persona fields on Kindroid, character cards everywhere. These features are platform-specific. The patterns are universal.

The best use of cross-platform patterns is to write content that works regardless of platform features, then layer the platform-specific tools on top to extend what the patterns already do. A well-written character card works on every platform. A well-written first message sets tone on every platform. A well-placed author's note (or its informal equivalent in apps that don't have explicit author's note features) influences every platform.

Don't optimize so heavily for one platform's features that your prompts stop working when you migrate. Build the underlying patterns first, then customize.

A workflow for transferring conversations between platforms

If you're moving a character or roleplay from one platform to another (which the memory architecture post discusses isn't directly possible but can be partially recreated), the cross-platform patterns are what survives the move.

The character card written in concrete, declarative language on Character AI will work on Kindroid. The first message that established tone through example will set the same tone on Joyland. The voice rules captured in an author's note style instruction will translate.

What won't transfer: anything that depends on platform-specific memory features, automatic fact extraction, or particular UI conventions. Those have to be rebuilt on the new platform.

The fewer platform-specific elements you depend on, the more portable your characters and roleplays become.

Frequently asked

Why don't most "perfect prompt" posts work on every platform?

Because most of them are tuned to a specific platform's quirks: a particular system prompt the platform uses, a particular filter behavior, a particular memory architecture. When you switch platforms, those quirks change, and the prompt loses its effect. Cross-platform patterns work on the underlying model, which is more consistent across platforms.

Are these patterns specific to AI companions or do they work everywhere?

They work everywhere. The same patterns that work on Character AI work on ChatGPT, Claude, and any other LLM-based system. AI companion apps just expose them more visibly because the use case (long conversations with character consistency) puts more pressure on prompt quality.

Should I memorize these patterns?

Not exactly. Internalize the principles: direct over indirect, positive over negative, concrete over abstract, examples over descriptions. The specific phrasings emerge from those principles in any given situation.

What if a platform has terrible filtering or memory? Will these patterns still help?

They'll help, but they won't fully compensate for poor underlying architecture. Good prompting can produce noticeably better results on any platform, but a platform with weak memory is still going to lose long-term continuity faster than a platform with strong memory regardless of how well you prompt.

Do these patterns work for serious work, not just roleplay?

Yes. The same principles work for AI-assisted writing, research conversations, technical Q&A, and any other LLM use. The examples here lean toward roleplay because that's where prompt quality matters most visibly, but the underlying logic transfers.

How do I know when a prompt isn't working?

Inconsistent output is the main signal. If the same prompt produces wildly different results on different attempts, it's probably ambiguous. Tighten the framing, make instructions more specific, and use positive directives.

What's the single most important pattern for new AI companion users?

Lead with concrete tone-setting in your first messages. Whatever conversation you're trying to have, the first 2-3 messages set the template for everything that follows. If you get those right, a lot of subsequent prompting becomes unnecessary because the model has already calibrated.