Prompt Writing & Design Guide
[Demo promos] We recommend trying all three demo promos first — it'll be much easier to follow this guide once you have!
Demo 1: Fun interaction (greeting card) | Make a New Year's Card — Count Down with Little Fever!
Demo 2: Expert consultation (hairstyle) | Go Viral — Your Destined New Year Hairstyle Guide!
Demo 3: Logic-based curation (itinerary) | Which Kind of Year-End Survivor Are You? A Holiday Itinerary Guide
This guide will help you turn AI into a powerful marketing tool through precise "quiz design" and "prompt writing" — generating a one-of-a-kind, personalized result for every single participant.
This guide covers three core parts:
Quiz design principles: how do you embed insight into your questions?
Prompt structure formula: the essential elements every instruction needs.
Optimization tips: secrets to getting more precise results.
Part 1: Designing quiz questions — how do you capture participant insight?
For AI to write copy that feels like it "gets you," the key is what data we feed it. Quiz questions aren't just for fun — they exist to "tag" participants and "gather material."
1. Design principle: every option should represent a "persona"
Don't just design surface-level questions — design questions that reveal "psychological state" or "needs."
Example breakdown (from Demo 3's itinerary guide):
Question: "In what situation are you most likely to 'quietly explode'?"
Surface-level options: A. Your boss drops a last-minute brief on you / B. Replying to messages on your day off...
Underlying insight (the signal sent to AI):
Choosing A → means they're a "work addict" (needs mandatory rest)
Choosing C → means they're a "social deserter" (needs to get away from people)
Guidance:
When designing your questions, first sketch out 4–5 "personas," then design questions that sort participants into those personas.
[Strongly recommended] Use our AI Planning Agent — tell it your desired theme, and let AI help write the questions and answer options.
2. Gather "parameters" to feed the AI
Beyond psychological questions, also include concrete "parameter questions" so the AI-generated content includes personalized details.
Example breakdown (from Demo 1's New Year's card):
Question: Choose the New Year's scene you'd like to be in, and your favorite mascot pose.
Purpose: These become the "designated assets" the AI uses when compositing the image, so the resulting image directly reflects the user's choices, adding a personal touch.
Guidance:
If you want specific elements in the result image (like a specific product, background, or style), let users choose it through a question, or gather it through a photo upload.
Part 2: The golden framework for writing prompts
A great prompt is like a brief you'd hand to an outside vendor. Whether you're generating text or images, follow the CRIS framework:
1. C — Context & Role
Tell the AI "who you are" and "what you're doing right now."
Format:
You are [role definition]. Based on the user's responses, generate [target output].Examples:
Text: You are a senior domestic travel guide...
Image: You are an expert in image compositing... / You are a professional hairstylist...
2. R — Rule of Input
Clearly tell the AI what data the user will provide, and what that data means (this is the "insight" from your quiz design above).
Format:
The user will provide [number] elements:Element 1: [question content] (corresponding trait: if A, this means...)Element 2: [uploaded photo]
Important tip: Building a "strong-mapping" reference table like Demo 3 does (e.g., choosing A = "down-to-the-wire survivor") makes the AI's reasoning much more reliable.
3. I — Instruction & Logic
This is the most important part — tell the AI how to process this data and produce the result.
Key points for text-generation instructions:
Reasoning logic: "Based on the user choosing option A, determine that they're suited to..."
Tone and style: "playfully sarcastic," "warm and comforting," "professional advice."
Structure requirements: "Please include: 1. Result title 2. Reasoning 3. Recommendation."
Key points for image-generation instructions:
Compositing rules: "Remove the background from the user's photo and composite it with the mascot."
A sense of interaction (critical): As Demo 1 emphasizes, "make sure the mascot and the person feel warm and interactive," with examples (like both looking up together, or raising a toast together).
Preserving features: As Demo 2 emphasizes, "keep 90% facial similarity" and "adjust based on face shape."
4. S — Spec & Constraint
The final acceptance criteria, to make sure the output is correctly formatted and safe.
Format:
Character limit: under 100 characters.Language: Traditional Chinese.Image ratio: 9:16 portrait.Style keywords: American comic style / photoreal / pixel RPG style.Negative prompts (things to avoid): don't obstruct the main composition; if the subject is incomplete, fill it in naturally.
Part 3: Prompt optimization and common pitfalls
To make your prompts clearer, watch out for these details:
1. Text generation: be specific with your adjectives
AI may interpret abstract adjectives differently than a person would.
❌ Vague: Write a fun caption.
✅ Specific: Write a caption with a cute, sarcastic, humorous tone, similar to something like "may your bad luck take a direct hit."
Tip: Providing few-shot examples works really well! As in Demo 1, which supplied 5 sample captions so the AI could precisely match that tone.
2. Image generation: special instructions for "faces"
Compositing that involves a user's uploaded photo (like a hairstyle swap or scene composite) is the hardest part to get right.
Instructions you should always include:
Face consistency: strictly preserve the original facial features.Inpainting / blending: if the person's photo is cropped incompletely, fill in the body naturally.Lighting match: keep the lighting consistent between the composited subject and the background.
Style control: if you're not generating a realistic photo, always specify a style (like American comic or pixel art) — otherwise the AI tends to produce uncanny-valley results.
3. Fail-safe logic: when a user picks something unexpected
Set a default (fallback): for cases where the user selects "I don't know" or uploads a photo that isn't of a person.
Example: Demo 1 includes an option like "Not sure — surprise me (let the AI pick a scene outside these 5 options)." This is a great fail-safe that gives the AI room to improvise.
Summary: the structure of a great AI personality quiz prompt
Section | Example content | Purpose |
Role setup | You are a senior travel guide, a hairstylist... | Establishes expertise and tone |
Input parsing | Choosing A on Q1 means workaholic; input includes a personal photo | Helps the AI understand the user's data |
Processing logic | Recommend an itinerary based on Q1; recommend a hairstyle based on face shape | The core computation rules |
Style / examples | Reference: "Surviving another year"; style: pixel art | Keeps the output on-brand |
Hard constraints | Traditional Chinese, 9:16, under 100 characters | Ensures the output is correctly formatted and usable |
With this framework, you can turn your marketing ideas into precise AI instructions, and create personalized experiences that genuinely delight your participants!