AI easily interprets information in simple requests, but if input is very long and complex, model may misunderstand.

To avoid this, try adding structure to prompt and make response of AI more predictable and clear.

How to structure a prompt?

The creators of neural networks suggest using special markup that the AI understands. These can be:

☞ Markdown, a text formatting language. For prompts, you can use bulleted and numbered lists, as well as the # sign, which in Markdown signifies different levels of headings and, in the prompt, defines the hierarchy of tasks.

Task
Plan a birthday party for a company of 8 people.

Restrictions
- Budget: 10,000 rubles 
- Location: at home 
- There are vegetarians among the guests

What should be in the plan?
1. Menu
- Main dishes 
- Snacks 
- Drinks 

2. Entertainment 
- Games 
- Music 
- Activities 

3. Timing of the event

☞ XML tags that indicate the boundaries of any text element. The beginning and end of the element are marked with <tag> and </tag>, and the tags themselves can be any.

<goal>Create a weekly menu for a family of 3 people</goal>

    <restrictions>
        <budget>10,000 rubles</budget>
        <preferences>More vegetables, minimum fried food, soup every day</preferences>
        <exclude>Mushrooms, nuts, seafood, honey</exclude>
    </restrictions>

    <format>
        <meals>breakfast, lunch, dinner, snack</meals>
        <description>A detailed recipe for each dish with a list of ingredients</description>
    </format>

☞ JSON, a data structuring standard that allows you to mark up any information in the prompt with simple syntax.

{
  "task": "Make a shopping list for the week",
  "parameters": {
    "number_of_people": 2,
    "preferences": ["vegetarian", "minimum sugar"],
    "budget": "up to 10,000 rubles"
  },
  "categories": [
    "vegetables and fruits",
    "cereals and pasta",
    "dairy products",
    "drinks",
    "other"
  ],
  "format_of_answer": {
    "type": "list",
    "group_by_categories": true
  }
>

It seems that markup is complicated so you can show your prompt to the AI and ask it to add markup itself without changing the essence.

Kandinsky 5.0 Video Lite and Kandinsky 5.0 Video Pro generative models on the global text-to-video landscape

🔘Pro is currently the #1 open-source model worldwide
🔘Lite (2B parameters) outperforms Sora v1.
🔘Only Google (Veo 3.1, Veo 3), OpenAI (Sora 2), Alibaba (Wan 2.5), and KlingAI (Kling 2.5, 2.6) outperform Pro — these are objectively the strongest video generation models in production today. We are on par with Luma AI (Ray 3) and MiniMax (Hailuo 2.3): the maximum ELO gap is 3 points, with a 95% CI of ±21.

Useful links
🔘Full leaderboard: LM Arena
🔘Kandinsky 5.0 details: technical report
🔘Open-source Kandinsky 5.0: GitHub and Hugging Face