What happened
DeepSeek’s collection presents seven prompts from its DeepSeek Harness Creator Mode video. The source says users can copy a prompt into DeepSeek Harness and select “Creator Mode” to try recreating the plugins. It does not provide a release date, price, usage quota, or a version number for Creator Mode in the supplied material. Nor does it claim that the resulting plugins will match the video exactly: some were refined through several rounds of conversation, and DeepSeek explicitly warns that a single prompt may not reproduce what viewers saw.
The examples specify a range of interfaces and tasks. One asks for a plugin, reachable from a left sidebar, displaying 3D models of the Eiffel Tower, Great Wall of China, Forbidden City, and another famous building. Another requests an interactive 3D demonstration game explaining computation inside a Transformer, with a specific instruction to keep the panel title and toolbar clear of macOS window buttons. A third asks for five whole-app visual styles, including a pixelated option. These are not merely requests for a topic; each includes a placement, interaction, or presentation constraint.
Two prompts describe creative workbenches. The poster tool should be accessible from a conversation, use a shader background pattern and English font from the DeepSeek Harness website, and let users customize the title, text, and shader colors. Its default title is “Everything is a plugin,” and the requested default page is minimal. The video editor should support uploading multiple videos, editing and exporting them, with a large preview above a timeline. It should also accept natural-language requests in the conversation for importing and editing video.
The remaining prompts cover document reading and usage visibility. The paper reader, placed in a right sidebar, should preserve uploaded papers’ original layout, fonts, images, and other content; selecting text should expose highlight, note, and ask-in-conversation actions. The statistics plugin should use actual conversation data to show token usage, conversation counts, estimated costs and budgets, a usage timeline, per-conversation usage and costs, and CPU activity and utilization. The source gives these as requested features, not measured results or confirmed performance claims.
What the archive adds
This is a new fact on a known DeepSeek Harness story, not a wholly new product thread. DeepSeek Harness ships dsh-v0.2.2-alpha.2 with plugin updates, six new UI languages and an experimental Codex/Claude CLI subagent tool covered an alpha update and an experimental subagent tool. The current source is narrower: it exposes seven prompts behind Creator Mode demos. The archive also has DeepSeek publishes DeepSeek Harness on GitHub, a related earlier item about the product itself.
Trend signal
The examples weakly support agent-ready-software-exposes-capabilities: they envision plugins as distinct, reachable capabilities such as a paper reader, video editor, and usage dashboard. But these are prompts describing demos, not evidence that the capabilities are reliably shipped, permissioned, or composable. Treat this as an illustration of a product direction, not validation of the pattern’s operational requirements.
Why now
The source’s timing is tied to a Creator Mode video: DeepSeek has made the seven demo prompts available so viewers can try recreating the showcased plugins. The supplied material does not explain a broader launch schedule or connect the release to another product change.
Why it matters
For builders evaluating prompt-driven interface generation, the collection offers unusually concrete starting specifications. The prompts name where an interface should live, what controls it needs, and how it should behave: a document reader must preserve layout and expose actions on selected text; a video editor needs a preview, timeline, import, editing, and export; a usage dashboard should break costs down by conversation. That level of detail makes it easier to adapt a request to a real workflow than to begin from a vague instruction to “build a plugin.”
The collection also reveals the gap between a demo prompt and a dependable product. A prompt can specify a desired result, but the source itself says some examples took several rounds of refinement. Teams should therefore assess Creator Mode by testing the full loop: whether it produces a working artifact, how much correction it takes, whether requested data and actions are actually available, and whether the result survives use beyond a demonstration. The published prompts are useful inputs to that evaluation, not proof of its outcome.
Limits and uncertainty
The source publishes requests for seven demos; it does not establish that every requested feature works, that generated plugins are production-ready, or that outputs are reproducible across runs. It provides no benchmark, success rate, latency, cost, access restriction, or usage limit. The statistics prompt asks for actual conversation data and CPU utilization, but the collection does not explain the data interfaces, permissions, or accuracy behind such a plugin. DeepSeek’s own warning that some demos needed multiple rounds is a material limitation: one prompt may not recreate the video’s result.
Reader takeaway
The useful deliverable is not evidence that Creator Mode can reliably build any app from one sentence. It is a set of seven specific interface briefs that builders can inspect, reuse, and test. Start with the prompt closest to your workflow, then measure how much iteration is needed and verify the generated tool’s behavior before treating it as usable software.