DeepSeek has published a collection of seven prompts for Creator Mode in DeepSeek Harness on its CDN — an agentic runtime where the model builds a ready-made application with an interface from a text request. Each prompt is fully disclosed and describes how to ask the model to create a specific tool: the cases can be reproduced and customized, rather than just evaluated based on others' demonstrations.

image

What happened

On November 2, 2026, DeepSeek published a page on cdn.deepseek.com featuring seven reference prompts for Creator Mode in DeepSeek Harness (DSH); each prompt is presented in both English and Chinese. The prompts describe how to ask the model to build a plugin: a 3D gallery of landmarks featuring the Eiffel Tower, the Great Wall of China, and the Forbidden City; an interactive 3D visualization of computations inside a Transformer; an app appearance switcher with five styles, including pixel art; a poster editor with a shader background and the headline "Everything is a plugin"; a video editor with a timeline, preview, and chat-based editing control; a scientific paper reader that preserves the original PDF layout and supports notes; and a statistics panel showing token usage, cost estimation, and CPU load. A separate announcement of the mode was posted on the DeepSeek Harness account on X with the slogan "Your Harness. You Define It. Meet Creator Mode."

Context

DeepSeek Harness (DSH) is an agentic runtime for DeepSeek models, and Creator Mode extends the agentic approach to interfaces: the slogan "Everything is a plugin" sets the rule that the interface is not designed in advance but emerges as a result of a text request. The public collection changes the format of verification: instead of a "look how it works" video, users receive source prompts in a "do it yourself" mode, which significantly increases the verifiability of claims about agentic generation. The selection of tasks is also telling: video editing, PDF reading with notes, and 3D graphics are composite interactive applications, meaning the task is inherently more complex than generating code snippets, and its failures are particularly informative. Another signal about the environment's architecture: one of the cases demonstrates that runtime metrics are available to plugins as regular data, which deserves separate verification.

Why this matters for the industry

For the industry, this is a rare case where agentic interface generation is documented with reproducible prompts: product teams receive a ready-made library of seven reference "text request — working tool" patterns that can be copied and adapted to their own scenarios right now. The shift affects the very requirements for developer tools and UX teams: full editors, readers, and visualizers are generated without manual interface design, not just code snippets. There is also a downside: if the interface is built from a description, the layer of "pretty UI on top of someone else's model" is devalued to near zero — this is a risk for wrapper teams and an opportunity for those building the runtime itself. If the cases are confirmed as reproducible, public prompt collections could become de facto benchmarks for agentic runtimes, where comparisons are made based on the number of iterations to a working tool and the percentage of successful reproductions without refinement.

Why this matters for users

The practical effect is immediate: with access to DeepSeek Harness, users simply need to select Creator Mode and paste any of the seven prompts to attempt to reproduce the tool — from a 3D model of a Transformer to chat-based video editing. The barrier to entry is measured by the prompt, not the code: the files are available in full in the public domain, and each can be modified for one's own task — this is a cheap way to check which internal tools and prototypes are better generated on demand rather than programmed. This is also the fastest way to personally verify how well claims about agentic interface generation match reality, without waiting for independent reviews.

What is currently unknown / limitations

The evidence base is limited to demonstrations: the sources do not include success metrics, the percentage of completed tasks, iteration statistics, or an indication of which specific model generated the cases. DeepSeek itself warns that some plugins were refined over multiple dialogue iterations — the showcase is assembled from the best results, and a single prompt may not exactly reproduce the demo from the video, meaning there is a selection effect in the collection. There is currently no data for production use: no information on generation costs, latency, or service level guarantees. The actual reproducibility of the "prompt — working application" combination across a wide class of tasks can only be judged by independent attempts to reproduce the published cases.

Sources

Author

Look at AI, editorial team