DeepSeek Harness Creative Mode: Build Plugins by Describing Them

DeepSeek Harness Creative Mode: Plugin Development and Mode Architecture

DeepSeek Harness (DSH) introduces a "Creative Mode" that enables users to create plugins by describing requirements in the chat interface. The mode loads the plugin development skill, reads the runtime, writes plugin files, installs the plugin, and hot-reloads it within the same session — after a page refresh the new plugin is immediately available.

Plugin Creation Examples

Token Usage and Cost Statistics Plugin

The author created a plugin to track token consumption and cost per session, motivated by a previous comparison where token usage varied 13x across models (DeepSeek, Kimi K3.1, Opus 5.5, GPT-6). Using one of seven official prompt templates, the plugin was created in a single attempt. After creation, the plugin appears in the sidebar under "Usage Statistics" and shows detailed per-session breakdowns.

Paper Reader Plugin with ArXiv Support

The author adapted the official "Paper Reader" prompt to accept ArXiv links or IDs directly, eliminating the manual download step. The prompt used:

Create a paper reader accessible from the right sidebar, supporting paper uploads and direct ArXiv links, preserving original layout, fonts, images, and other content without loss. Show a floating menu on text selection with three options: highlight, add note, and ask in conversation. The default question prompt language should match the interface language.

The resulting plugin opens from the sidebar, renders the paper with original formatting preserved, and provides a floating menu on text selection for highlighting, note-taking, and asking questions in the chat.

Thunder Big Thinking Analyzer Plugin

To observe DeepSeek's reasoning process (which is folded and scrolls quickly), the author created a plugin that extracts and displays the thinking steps. The plugin adds an entry at the top of the conversation, allowing side-by-side bilingual viewing of the reasoning. The author notes translation speed is slow when using Qwen3.6 flash for translation.

DSH's Four Modes Compared

DSH operates in four modes, each defined by a preset configuration that determines which tools and capabilities the agent can use. The author examined the underlying preset files and summarized the differences:

Feature                          | Standard | PTC      | Minimal | Creative
---------------------------------|----------|----------|---------|---------
File read/write, terminal, web search | ✅       | ✅       | Terminal + image only | ✅
Skills, planning, goals, todos   | ✅       | ✅       | ❌      | ✅
Sub-agents                       | ✅       | ✅       | ❌      | ✅
Workflows                        | ✅       | ❌       | ❌      | ✅
Scheduled tasks                  | ✅       | ✅       | ❌      | ✅
Context compression              | ✅       | ✅       | ✅      | ✅
Runtime inspection               | ❌       | ❌       | ❌      | ✅
Plugin development guidance      | ❌       | ❌       | ❌      | ✅
Plugin install & marketplace     | ❌       | ❌       | ❌      | ✅
System prompt                    | Coding agent + runtime context | Same as Standard | Fixed one-liner, no runtime context | Same as Standard

Mode Details

Standard Mode

The default mode: the agent directly calls tools to read/write files, search the web, run commands. It includes Skills, planning, sub-agents, and workflows. Suitable for everyday coding, file processing, and data organization.

PTC Mode (Programmatic Tool Calling)

Instead of calling tools one by one, the agent writes a TypeScript program using run_code that loops, branches, and invokes tools concurrently. Intermediate results are filtered and merged inside the program before being passed to the model. This is ideal for batch-processing hundreds of files under the same rule, keeping the context clean. The trade-off is that workflow tools are disabled. The official documentation notes that Standard Mode can also achieve batch processing via scripts; PTC only changes the invocation style and token savings depend on the specific task.

Minimal Mode

Only a persistent terminal and an image-reading tool are available. The system prompt is fixed to "You are a helpful software engineer assistant" with no Skills, no planning, and no runtime context injection. It serves as an experimental baseline for model comparison tests. The agent can still use the command line to read/write files and run scripts, but lacks assistance for managing long-running tasks. The author found it less beginner-friendly despite the reduced toolset.

Creative Mode

Builds on Standard Mode's full toolset and adds three capabilities: runtime inspection (viewing DSH's internal plugins and services), built-in plugin and agent preset development guidance, and plugin management/marketplace tools. Crucially, the plugin installation tool is only enabled in Creative Mode — the other three modes have it disabled. Therefore, creating plugins is only possible in Creative Mode.

Creative Mode Can Also Create Modes

Beyond plugins, Creative Mode can create new modes. Plugins add capabilities (tools, service connections, sidebar entries), while modes define which tools are available and how the agent works for a given task. Users can ask Creative Mode to combine tools and prompts into a custom mode (e.g., a "Writing Mode") that can also use their own plugins. The new mode appears in the mode list when creating a new task. Plugins may require a DSH restart depending on what they modify; new modes are available immediately for new tasks.

Architecture Note

DSH's architecture is based on the Cordis plugin framework — sidebar entries, tools, Skills, and even UI themes are all plugins. The "Everything is a plugin" design enables the extensibility demonstrated by Creative Mode.