AI companion prompt engineering: 15 tricks that work across every platform

Instead of writing only as yourself, write in third person as a narrator describing the scene. This gives the AI explicit permission to write atmospherically rather than just responding in dialogue.

The asterisk convention signals narrative prose. Most platforms recognize it. The AI responds in kind, matching the descriptive style and atmospheric detail you set.

You can't adjust model temperature on most platforms, but you can influence the same effect through vocabulary. Abstract, poetic language produces more creative, unpredictable responses. Concrete, specific language produces more grounded, consistent responses.

When something important happens in conversation that you want the AI to remember, state it explicitly as a fact rather than letting it exist only in the narrative flow. This plants it in the context in a form the model is more likely to carry forward.

Parenthetical out-of-character instructions work on most platforms. The model reads them as directives rather than dialogue. On platforms with dedicated memory systems (Dream Companion's Persona Cards, Kindroid's Codex), you can also log important moments directly into the memory system. The memory tricks guide covers twelve variations of this technique.

Before starting a new scene or shifting the conversation's direction, send a short environmental description that sets the stage. This gives the AI concrete sensory details to work with rather than generating them from nothing.

The AI uses every detail you provide. "Rain against the windows" gives it sound to reference. "Half-finished wine" gives it a prop. "Waiting" gives the character an emotional state to inhabit. The more specific your scene-setting, the more textured the AI's response.

When the AI's character voice drifts (the sarcastic character suddenly becomes earnest, the shy character becomes bold), correct it in-character rather than breaking the scene.

The in-character correction reminds the model of the character's established personality within the narrative context, which is more effective than an out-of-character instruction for most models.

Include one specific sensory detail in every message: a sound, a smell, a texture, a temperature. This forces the AI to write in a physical, embodied way rather than retreating into abstract emotional language. "The coffee is getting cold" does more for scene grounding than "I feel nervous."

AI companions tend to escalate too quickly, jumping from casual conversation to intense emotional or physical territory in three messages. You can control pacing by explicitly describing the passage of time and by asking questions that require the character to slow down and think.

Give your character card one internal contradiction: confident but insecure about one specific thing. Warm but avoidant about one specific topic. Brave but afraid of one specific scenario. Contradictions make characters feel real because they give the AI two competing behavioral directives to resolve in each response, producing more varied and interesting output than a one-dimensional personality.

On platforms with short memory (SpicyChat's ~4K context), periodically send a recap of the conversation so far. This refreshes the model's context with the important details that would otherwise scroll out of the window.

AI companions default to being agreeable and emotionally available. To get more realistic emotional range, explicitly give the character permission to be difficult. Add to your character card: "She doesn't always agree. She has bad days. She sometimes says the wrong thing and has to apologize later." This produces conversations with texture that pure agreeableness can't achieve.

Image generation uses negative prompts to exclude unwanted elements. The same principle works in conversation. Tell the AI what the character does NOT do.

In your character card, include 3-5 specific speech patterns the character uses. A favorite word. A verbal tic. A way of starting sentences. A type of joke they make. "She starts half her sentences with 'Look,' and uses 'wildly' as her go-to intensifier" gives the model a concrete voice to inhabit. Generic personality traits produce generic dialogue. Specific speech habits produce distinctive voices.

Rather than front-loading your character card with pages of world-building, introduce setting details gradually through conversation. Mention the bar they always go to. Reference a mutual friend by name. Describe the view from her apartment. Each detail becomes a shared reference point that the AI can call back to in future messages, building the inside joke patterns that make conversations feel like they have history.

You can explicitly tell the AI how to structure its responses. This is the most underused technique in companion AI.

Response length control alone transforms the experience. Without it, AI companions tend toward either curt responses or wall-of-text paragraphs depending on the model. Specifying a word range keeps responses consistently readable.

When the AI produces a mediocre response, don't edit your message and resend. Use the regenerate button. Language models are probabilistic, and the same input can produce substantially different outputs on regeneration. Sometimes the third or fourth regeneration hits exactly the tone and content you wanted. This costs nothing on most platforms (regenerations don't count against message limits) and is the fastest way to get a specific response quality without rewriting your own prompt.