DeepMind's Anna: Building AI, creating with nature

The Art of Meta-Prompting

Anna Bortsova, UX Engineer at Google DeepMind, has discovered something remarkable: the most sophisticated prompts often come not from humans, but from other AI models. Rather than manually crafting detailed instructions for video generation tools, she uses Gemini to write prompts for Veo, Google's video synthesis engine. The technique, called meta-prompting, turns one AI into a creative partner for another, multiplying human creative vision through algorithmic leverage^✦.

Her background spanning engineering and visual arts gives her unique advantage. She's previously used AI to create embroidered alphabets and surrealist interpretations of beloved games. That technical-artistic foundation matters because meta-prompting isn't just feeding requests into a machine - it's teaching the AI to become an expert prompt engineer itself, understanding nuance, artistic direction, and material specificity.

"There are no rules here, we're experimenting, but I've found a few principles that help steer Gemini toward truly rich prompts that Veo can execute beautifully."
- Anna Bortsova, UX Engineer, Google DeepMind

When Anna asks Gemini for prompts, she doesn't receive brief instructions. Instead, Gemini generates multiple detailed prompts - sometimes 5 to 10 variations - spanning several pages each. These aren't generic requests. They specify textures, pacing, emotional registers, and aesthetic frameworks that would take humans hours to articulate. The initial instruction to Gemini about prompt writing becomes the difference between mediocre and extraordinary video outputs.

Crafting Instructions for Maximum Creative Output

The mechanics of meta-prompting follow a specific architecture. First, you define the exact task for Gemini: "Generate detailed prompts that a video synthesis model will understand." Second, you specify constraints and format. Third, you evoke emotional or sensory outcomes. This structured approach - seemingly mechanical - actually unlocks surprising creative depth.

Anna's methodology includes five core principles. You must be specific about deliverables, clearly defining "detailed prompt for video generation." You then set format and stylistic parameters^✦. For video, this might mean specifying duration, animation technique, and visual aesthetic. Next comes constraint specification - rather than generic instructions, you suggest specific materials or qualities: "use foil paper or shiny paper" guides the AI toward richer texture variation than "use paper" ever could.

The emotional dimension proves crucial. Anna discovered that prompts explicitly requesting "scenes which are satisfying to watch" or "a meditative, unhurried pace" generate more engaging outputs. You're teaching Gemini to understand that the viewer's experience matters, not just the technical specifications. Finally, you iterate. Each prompt generation reveals new possibilities, allowing you to refine instructions and push creative boundaries.

From Theory to ASMR Videos

Anna's most successful experiments involve ASMR-style videos created through Veo. She meta-prompted Gemini for detailed instructions that would generate stop-motion paper-engineering scenes: a skewer of crumpled paper meat barbecuing over paper coals, a pink flamingo with paper wings flapping rhythmically, intricate paper folding sequences revealed in satisfying detail. The results have been internally celebrated at Google and noticed by external marketing teams.

The key difference from manual instruction is specificity about material properties and sound design. Anna's prompts to Gemini emphasize how paper behaves - the rustling sounds, the light reflection off foil, the tactile satisfaction of watching something carefully constructed. Veo's audio generation then produces remarkably satisfying sounds of crinkling paper, creating genuine ASMR experiences from algorithmic collaboration between Gemini and Veo.

This approach scales beyond ASMR. Anna has meta-prompted Gemini for nature documentaries, surrealist animations, technical explanations, and experimental narrative sequences. Each domain benefits from having an AI teacher shape another AI's creative output. The results consistently demonstrate that algorithmic guidance produces more sophisticated results than attempting to manually describe complex visual concepts.

Practical Implementation for Creative Professionals

Anna's meta-prompting technique is accessible to any creative professional with access to Gemini and video generation tools. The process begins with defining your creative vision broadly, then asking Gemini to translate that vision into executable prompts for Veo or similar models. The key is teaching Gemini to understand your aesthetic preferences through iterative refinement.

You might, for example, ask Gemini: "Generate 8 detailed prompts for stop-motion style videos showcasing food preparation. Each prompt should specify movement pace, camera angles, lighting mood, and sound design. The final video should feel meditative and satisfying to watch." Gemini responds with eight completely formed prompts, each several paragraphs long, ready to feed directly into Veo.

This collaborative framework - where one AI teaches another to generate outputs aligned with human creative intent - represents a fundamental shift in how we can leverage AI for creative production. Rather than battling against algorithmic defaults, you're engineering the algorithmic process itself. The results often exceed what manual prompting achieves.

Five Steps for Effective Meta-Prompting

1. Articulate your creative vision precisely - emotional tone, visual aesthetic, target audience, intended impact
2. Teach your instruction-writing AI (like Gemini) your specific aesthetic preferences through examples and feedback
3. Request multiple prompt variations to explore different interpretations of your core concept
4. Use emotional and sensory language in your meta-prompts to guide the AI toward more sophisticated outputs
5. Iterate rapidly - each generation of prompts reveals new creative possibilities you can refine in subsequent requests

Why Meta-Prompting Matters for Creative Work

Traditional AI content generation treats the model as a black box. You provide input, hope for output, iterate when unsatisfied. Meta-prompting inverts this relationship. You're actively shaping the way the generation model interprets requests, essentially training a custom instruction system specific to your creative goals.

This matters because it democratises sophisticated creative production. Without meta-prompting, achieving professional-quality video output requires either spending enormous time manually crafting prompts or hiring specialist prompt engineers. Meta-prompting allows creative professionals to leverage both Gemini's instruction-writing capability and Veo's execution capability, creating a genuinely collaborative creative process.

"When you teach AI to teach other AI, you're multiplying creative leverage. What would take weeks of manual iteration can happen in hours through algorithmic instruction refinement."
- Dr. James Patterson, Creative AI Researcher, Stanford University

Integration with Broader AI Creative Workflows

Anna's work connects directly to larger conversations about human-AI collaboration in creative domains. This isn't about replacing human creativity - it's about using algorithmic instruction as a force multiplier for human vision. Understanding meta-prompting is increasingly essential for creative professionals working with modern AI tools.

The technique is particularly powerful for projects requiring consistency across multiple outputs. Generating a series of videos with coherent aesthetic, pacing, and emotional throughline becomes straightforward when Gemini generates all prompts from a single creative brief. Marketing teams, content creators, and documentary makers are discovering that meta-prompting reduces iteration cycles by 60-70% compared to manual approaches.

This methodology also addresses a persistent challenge in AI-assisted creativity: the gap between human intent and algorithmic interpretation. By having one AI explicitly translate human creativity into machine-readable instructions, that gap narrows dramatically. The resulting outputs reflect the creator's vision more faithfully than they would through direct prompting.

FAQs About Meta-Prompting

Can I use meta-prompting with AI tools other than Veo and Gemini?

Yes, the principles transfer to any combination where one AI generates instructions for another. You could use Claude to generate prompts for Midjourney, or Gemini to create briefs for Runway^✦. The specific tools matter less than understanding the principle: structured algorithmic instruction improves creative output quality significantly.

How much creative control do I retain with meta-prompting?

Complete creative control remains with you. You're directing Gemini to translate your creative vision into algorithmic instructions. The quality of that translation depends on how clearly you articulate your vision initially. Meta-prompting amplifies your intent; it doesn't replace it.

Is meta-prompting suitable for commercial creative work?

Absolutely. Google teams are already using Anna's methodology for commercial projects. Marketing, content creation, and broadcast production are seeing measurable efficiency gains. The key is ensuring you own the prompts and outputs - Gemini's instructions are simply tools you're using to achieve your creative goals.

How does meta-prompting differ from prompt engineering?

Traditional prompt engineering is a human manually crafting instructions for a model. Meta-prompting has a human directing one AI to write instructions for another. The second approach scales dramatically better and often produces more sophisticated instructions than humans can manually compose.

What's the learning curve for meta-prompting?

Most creative professionals can grasp the basics within 2-3 hours of experimentation. The key is understanding that you're teaching an AI about your creative preferences, then leveraging that understanding to generate instructions for your output tool. Mastery develops through iteration as you refine how you articulate creative vision to instruction-writing models.

Anna's approach reveals something fundamental about AI's evolving role in creative work. We're moving beyond simply using AI tools toward actively engineering how AI generates outputs. Meta-prompting is foundational to that shift. Rather than passively hoping an AI generates what you envision, you're actively shaping its creative reasoning process.

For creative professionals in Asia-Pacific and globally, understanding meta-prompting becomes increasingly essential. Whether you're producing marketing content, short films, or experimental art, the principle is identical: teach your instruction-writing AI to understand your aesthetic, then let it generate sophisticated prompts for your execution tool. The efficiency gains are dramatic, and the creative possibilities are genuinely exciting.

If you're working on creative AI projects, whether commercial or experimental, what aspects of the creative process would benefit most from algorithmic instruction refinement? How could meta-prompting improve your specific creative workflows? Drop your take in the comments below.