Googler details meta‑prompt technique that guides Gemini to craft Veo videos
The internal memo from Google’s AI team reveals a surprisingly hands‑on approach to getting Gemini to produce the kind of footage that powers Veo’s sports‑highlight reels. Rather than feeding the model a single, static instruction, a researcher named Anna has been experimenting with a two‑step prompting workflow. First, she asks Gemini to generate a set of instructions for itself; then she feeds those self‑generated cues back into the model to shape the final video output.
The method is still in trial mode, with no formal guidelines dictating how the prompts should be crafted. Still, Anna reports that certain patterns in her “meta prompts” consistently coax Gemini into delivering more nuanced, content‑rich results. The technique, while still experimental, appears to be the linchpin that transforms raw model capabilities into the polished, context‑aware clips seen in Veo’s latest releases.
The following excerpt explains why those self‑referential prompts matter.
But the prompts she uses to instruct Gemini on how to create its prompts are key. Anna's meta prompts inspire Gemini to produce richly detailed prompts for instructing a gen AI model. "There are no rules here -- we're experimenting -- but I've found a few things that help steer Gemini to really rich prompts," she says.
"You want to define a very specific task: 'write a detailed prompt that an LLM will understand.' And you want to be clear about your format and style: say, an 8-second stop-motion animation of paper-engineered scenes. Then give it constraints, like foil paper or shiny paper, rather than just general paper. Then let it do its thing." Depending on how a model responds to Gemini's prompts, you may want to tweak them, she says.
Add or change details about the sounds and textures you want to produce -- it's a collaboration. "I've found it helps to suggest the feeling you want to evoke," she adds. "Tell Gemini you want it to think about 'scenes which are satisfying to watch,' for example." With such instructions and the task of creating botanical art, Gemini delivered a prompt for an unfurling paper fern in which "the animation should be slow and mesmerizing, with each frond delicately unfolding in a gentle, rhythmic sequence." Veo understood the assignment.
Anna's ferns and feathers are not part of her core work: Day-to-day, she helps build the infrastructure and tools for Google DeepMind's researchers to scale their AI experiments. But it's something that gives her joy when she finds a spare 10 minutes, and she's happy to share the love. (She even created a deck to pass on her learnings.) Her biggest tip for Googlers… and anyone else who's listening?
Anna Bortsova’s work shows that a “meta‑prompt” layer can nudge Gemini toward richer descriptions, which in turn feed a generative model that assembles Veo‑style videos. The internal chat group where Googlers swap demos illustrates how quickly such tricks surface, yet the process remains informal—“there are no rules here,” she says, and experimentation drives the results. What she calls a handful of steering techniques appear to coax Gemini into producing “really rich” prompts, but the article offers no data on consistency or scalability.
It’s unclear whether these methods will translate beyond the team’s sandbox or how they compare with other prompt‑engineering approaches. The lack of formal guidelines suggests that success may hinge on individual intuition as much as on the underlying model. While the demonstrations are impressive, the broader impact on Google’s AI pipeline remains uncertain, and further testing will be needed to confirm whether the meta‑prompt concept can be adopted more widely.
Further Reading
- Googler details meta-prompt technique that guides Gemini to craft Veo videos - The Verge
- Ultimate prompting guide for Veo 3.1 - Google Cloud Blog
- Prompt design strategies for Gemini models - Google AI Developer Documentation
- Veo on Vertex AI video generation prompt guide - Google Cloud Documentation
- Inside Google’s plan to use Gemini for AI-generated sports highlights - TechCrunch
More in Research & Benchmarks
CognitiveLab unveils NetraEmbed, 150% accuracy gain, adds ColNetraEmbed
CognitiveLab’s latest release promises a tangible lift in how machines handle multilingual text.
Student AI models can inherit bias and harmful traits from teacher models
The paper titled “Subliminal Learning: How AI Models Inherit Hidden Dangers,” published under the...
70% of Creatives Fear Stigma as AI Drives Majority of Their Ideas – Anthropic
Seventy percent of creative professionals say they worry about being judged for leaning on...
Corporate AI agents favor simple workflows; 41.5% accept minute‑range latency
Corporate AI agents are being rolled out across enterprises not as fully autonomous bots but as...
Bright Data API Delivers Seamless AI/ML Integration and Anti‑Bot Protection
Why does the choice of a scraping tool matter for today’s AI projects? Teams building large‑scale...
Common Questions Answered
How does Anna’s two‑step prompting workflow help Gemini create Veo‑style videos?
Anna first asks Gemini to generate its own set of detailed instructions, then feeds those self‑generated cues back into Gemini. This iterative process guides the model to produce richer prompts that a generative system can use to assemble sports‑highlight reels similar to Veo’s output.
What is a “meta‑prompt” and why is it important for Gemini according to the memo?
A meta‑prompt is a higher‑level instruction that tells Gemini how to write its own prompts, effectively a prompt about prompting. By defining a very specific task and format, the meta‑prompt encourages Gemini to produce richly detailed cues, which improves the quality of the final video generation.
Which specific techniques does Anna recommend to steer Gemini toward “really rich” prompts?
Anna suggests explicitly stating the task, such as “write a detailed prompt that an LLM will understand,” and being clear about the desired format and style, for example specifying an 8‑point structure. These steering techniques help Gemini focus its output and generate more comprehensive instructions for downstream video creation.
What does the article say about the formalization of the meta‑prompt approach within Google’s AI team?
The article notes that the process remains informal, with no official rules governing the meta‑prompt technique. Googlers share demos in an internal chat group, and experimentation drives results rather than a standardized protocol.