DeepSeek published seven official Creator Mode prompts. Here's what each one teaches about prompting an agent that builds its own plugins.
What are DeepSeek Harness Creator Mode prompts? #
DeepSeek Harness’s Creator Mode prompts are the seven official example prompts DeepSeek published to show how its agent app builds working plugins from a conversation. Each prompt describes a plugin, where it should appear in the app, and what it should do, and the agent writes the code, inspects the running app for available methods, and installs the result as a plugin tied to the user’s profile. DeepSeek released these after demoing Creator Mode publicly, partly to answer the obvious question: what did people actually type to get those results?
TL;DR #
- DeepSeek published seven prompts in English and Chinese on a single page with copy buttons, covering everything shown in its Creator Mode demo video.
- The prompts double as a prompting guide : read closely, they show a repeatable pattern for getting an agent to build a self-contained plugin instead of a vague answer.
- DeepSeek Harness is an open source agent app (MIT licensed, over 269,000 GitHub stars at time of writing) built around the idea that “everything is a plugin,” currently at release candidate 0.211 RC5.
- Creator Mode adds two read-only inspection tools that let the agent look at the running app’s loaded modules and available methods before writing any plugin code.
- The clearest lesson across all seven prompts is to specify where the plugin lives (left sidebar, right sidebar, or opened from the chat), since six of the seven do exactly that.
- DeepSeek itself admits some demo plugins took several rounds of conversation to refine, which matters for anyone paying per token to run Creator Mode.
- Pairing Creator Mode with a flat-rate model subscription and a parallel task runner can make iterating on plugins cheaper and faster than running everything through a single pay-per-token API.
pass the audit.
How does Creator Mode actually work? #
DeepSeek Harness runs in different modes. Standard mode is a normal coding agent that reads and writes files like most agent harnesses. Creator Mode is different because it’s aware of the app it’s running inside.
When Creator Mode is active, the agent gets two read-only tools: one to list what’s currently loaded in the running app, and one to check the exact methods and types available to call. This means the agent isn’t guessing at an API surface, it’s inspecting the live app state before it writes anything. Once it has a plan, it packages the result as a plugin bundle and installs it through the app’s built-in plugin manager. Because the plugin is tied to your profile, it survives a restart instead of disappearing at the end of the session.
This is the mechanism behind DeepSeek’s “everything is a plugin” tagline. The app isn’t just generating code you paste somewhere, it’s extending itself in place, live, while you’re talking to it.
What do the seven official prompts actually ask for? #
Each of the seven prompts corresponds to something shown in DeepSeek’s original Creator Mode demo:
- 3D landmark gallery : a left sidebar plugin rendering 3D models of the Eiffel Tower, the Great Wall of China, the Forbidden City, and one more landmark of the model’s choosing.
- Transformer visualizer : an interactive 3D demonstration explaining the computation inside a transformer, with an explicit instruction to keep panel titles and toolbars clear of the native macOS window buttons so they don’t overlap.
- Pixelated interface (DSH pixelated) : a plugin that restyles the entire app, offering five selectable visual styles including a pixel art look.
- Poster design workbench : opened from the conversation, built using the shader background pattern and font from the DeepSeek Harness website, with customizable title, text, and shader colors, and an instruction to keep the default page minimal.
- Video editing workbench : a dark-themed editor with multiple video uploads, a large preview panel on top and a timeline below, editable and exportable, with natural-language editing commands available in the chat.
- Paper reader : a right sidebar plugin for up papers that preserves original layout, fonts, and images, with a floating menu on text selection offering highlight, add note, or ask-in-conversation options.
- Conversation statistics dashboard : a plugin built on the user’s actual conversation data, showing token usage, conversation counts, cost estimates, a token usage timeline, per-conversation cost breakdowns, and CPU activity.
DeepSeek is upfront that a single copy-pasted prompt might not reproduce the demo exactly. Several of the showcased plugins were refined over multiple rounds of conversation, so these prompts are starting points rather than one-shot templates.
What do these prompts teach about prompting an agent to build plugins? #
Read as a set, the seven prompts reveal a consistent structure that’s more useful than any single example on its own.
made for the enterprise.
Say where it lives. Six of the seven prompts specify placement: three live in the left sidebar, two open from the conversation, one lives in the right sidebar. Only the statistics dashboard omits this, and it’s notably the one case where the agent has to guess the plugin’s location on its own.
Name the exact problem, don’t just say “fix it.” The transformer visualizer prompt includes a specific instruction about avoiding overlap with native macOS window buttons, the kind of detail that only gets added after a first attempt visibly failed. The lesson: when a plugin looks broken, describe the precise defect in the next prompt rather than restarting from scratch.
Give it numbers, not adjectives. The pixelated interface prompt asks for exactly five styles, not “several” or “a few.” Numbers are checkable. Adjectives aren’t, and vague quantities tend to produce inconsistent or incomplete results.
Tell it what to leave out, and set defaults. The poster workbench prompt explicitly says to keep the default page minimal, with no unnecessary subheadings, and sets a specific default title. Agents tend to add filler unless told not to, so an explicit negative constraint plus a default value closes that gap.
Connect the plugin back to the conversation. The video editing workbench isn’t just a standalone panel, it accepts natural-language editing commands typed in the chat. A plugin that the agent can also drive through conversation is more useful than one that’s purely a static UI.
Spell out interactions step by step. The paper reader prompt details exactly what happens when you select text (a floating menu with three specific options) and even specifies that the default question prompt should match the interface language. These are details that only surface after actually using the tool, which is why including them upfront saves iteration.
Ask for real data, not plausible-looking data. The statistics dashboard prompt insists the numbers come from the user’s actual conversation history. Without that constraint, it’s easy for an agent to produce a dashboard that looks correct but is populated with fabricated figures.
Put together, a solid Creator Mode prompt tends to cover five things: where the plugin lives, what it does, the specific numbers and defaults involved, what to explicitly exclude, and how it ties back into the conversation.
Is Creator Mode worth the token cost? #
The catch with Creator Mode is baked into DeepSeek’s own documentation: plugins shown in the demo took multiple rounds of conversation to get right. Each round means the agent re-reading the running app’s state, rewriting the plugin, reinstalling it, and then responding to follow-up corrections. On a pay-per-token API, that back-and-forth adds up quickly, especially for more ambitious plugins like the 3D gallery or the video editor.
One way to manage that cost is to run Creator Mode against a flat-rate model subscription rather than metered API billing, since Creator Mode’s iterative nature rewards a setup where repeated rounds of conversation don’t carry a per-token penalty. DeepSeek Harness supports adding a custom model API under its settings, so it’s possible to route Creator Mode traffic through any OpenAI-compatible endpoint rather than a single default provider.
- 1 Secure
- 2 Regulation-compliant
- 3 Auditable
- 4 Scalable
1 SecureAdjusters can't export patient data✕ Export blocked
2 CompliantSSN
•••-••-4821✓ HIPAA ✓ SOC 2 3 Auditable09:14 · M. Chen approved09:12 · J. Ruiz edited09:07 · A. Patel viewed
4 Scalable12,480users online
It’s also worth separating “plugin that lives inside DeepSeek Harness” from “general coding task.” Not everything needs to be a Creator Mode plugin. Ordinary web apps, repo reviews, or scheduled jobs are better suited to a standalone coding agent running outside the app, with Creator Mode reserved for things that specifically benefit from being embedded in DeepSeek Harness itself, like UI themes or panels that read the app’s own runtime state.
What are the current limitations? #
DeepSeek Harness is still labeled a developer preview and ships release candidates frequently, so behavior can shift between builds. DeepSeek’s own note that a single prompt may not reproduce the demo results means users should expect iteration, not one-shot perfection. Plugins that restyle the entire app, like the pixelated interface, carry some risk of breaking the app’s visual consistency, though the built-in plugin manager provides a way to disable or remove a plugin if that happens. Anyone running Creator Mode heavily on a capped subscription or usage tier should also expect to hit rate limits with larger, more capable models faster than with lightweight ones.
Frequently Asked Questions #
What is DeepSeek Harness?
DeepSeek Harness is DeepSeek’s open source agent application, built around the principle that “everything is a plugin.” It’s MIT licensed and has passed 269,000 stars on GitHub, with active release candidates shipping regularly.
What makes Creator Mode different from standard mode?
Creator Mode adds two read-only tools that let the agent inspect the running app, listing loaded modules and checking available methods and types, before it writes and installs a plugin through the app’s plugin manager.
Will copying one of the seven official prompts exactly reproduce DeepSeek’s demo?
Not necessarily. DeepSeek states that several of the demoed plugins required multiple rounds of conversation to refine, so the published prompts are reliable starting points rather than guaranteed one-shot results.
Do all seven prompts specify where the plugin should appear?
No. Six of the seven specify placement (left sidebar, right sidebar, or opened from the conversation). The conversation statistics dashboard prompt is the exception, leaving placement up to the model.
Why does Creator Mode cost more in tokens than typical coding prompts?
Because each plugin build involves the agent reading the app’s runtime state, writing code, installing the plugin, and then handling follow-up refinement rounds, all of which consume tokens on every iteration rather than in a single pass.