dsh-think-zh：用系统提示词固定 Agent 的中文推理
dsh-think-zh 是 DeepSeek Harness 的零依赖 bundle 插件，通过 systemPrompt section 注入高优先级语言规则，默认固定推理通道为简体中文，并可配置 target、verbatim 与自定义 instruction。可在 Harness Web 插件页或 plugin_manager install_bundle 安装，配置写在 cordis.patch.yml。本文介绍安装、配置项、日志量化验证方式及工具输出导致的已知边界。
Read Moredsh-prompt-editor：在「更多设置」里统一管理全局系统提示词
dsh-prompt-editor 是 AGImentu 维护的 DSH 客户端插件，在「更多设置」中提供一页用于查看、修改并保存所有 agent 共用的全局系统提示词。它通过独立段落名 prompt-editor:global-instructions 注册系统提示词，避免直接改 personaPrefix 被 dsh-persona 遮蔽而不生效的问题。保存写入当前 profile 的 cordis.patch.yml，Loader 立即重组，下一轮对话即生效；清空保存可移除自定义。支持插件管理器本地路径安装、CLI github:AGImentu/dsh-prompt-editor 及手动 pnpm link。需注意本机路由无额外鉴权、分 profile 配置及宿主代码改动需重启客户端。MIT 0.1.0。
Read Moredsh-user-memory：用户编写的分级长期记忆注入 DSH 系统提示词
dsh-user-memory 是 XinzeQing 维护的 DSH 记忆类插件：将用户编写的 rules-must、agent、profile、rules-should 等 Markdown 注入系统提示词，提供中文面板编辑；AI 仅能通过 memory_suggest 提议，用户采纳后才写入。可选临时记忆跨会话交接，运行时只读、零依赖。安装通过 profile package.json 的 link 依赖与 cordis.patch.yml 配置。
Read Moredsh-prompt-inject：把自定义文字块注入 DSH 系统提示词
dsh-prompt-inject 是 azazo1 维护的 DSH 工作流插件，在系统提示词组装阶段注入用户维护的文字块或外部 Markdown 文件，支持段落顺序与 15 条 Codex/Claude Code 来源预设；默认不注入内容。本文介绍安装命令、配置字段、典型用法及与内置提示词的关系，并说明插件无网络访问、随 dsh 进程权限运行。
Read More提示词专家：把含糊需求落成可测、可迭代的提示词
SkillHub 社区 Skill「提示词专家」（@user_8d26dabd/qclaw-prompt-engineer，v2.1.4）面向提示词输出不稳定、需求边界不清的场景，将含糊目标整理为更可执行的提示词，并提供约 5–10 条测试用例与迭代方法以对齐预期。无需 API Key（SkillHub 标注）。本文基于 SkillHub 页面与安装文档说明 skillhub CLI 安装步骤，以及 Cursor、Claude Code、Codex 等常用 skills 目录示例，并归纳适用场景与使用注意。
Read More预设工坊：为 DeepSeek Harness 管理人设、系统提示词、工具与技能
预设工坊是专为 DeepSeek Harness (DSH) 设计的 Agent 配置管理工具，用于集中管理人设、系统提示词、工具、技能及预设方案。它允许用户按需组合能力，精简工具与技能以优化任务结果、权限和消耗，并支持预览最终请求组成。主要功能包括预设方案的新建与同步、人设变量编辑、工具与技能的动态绑定、JSON 工具描述编辑、配置导入导出以及会话预设回退机制。该工具支持 `dsh web` 或桌面端，通过 `plugin_manager` 安装。典型用法涵盖新建方案、编写含变量的提示词、组合工具技能、预览检查及保存使用。需注意，该工具不替代安全审批，修改配置不可撤销已发生操作，且部分代理层
Read Moredsh-personalization-local：DeepSeek Harness 个性化设置插件
DeepSeek Harness 插件 `dsh-personalization-local` 可通过配置称呼、职业及回复偏好，将信息自动注入系统提示词，实现 AI 按设定工作。该插件支持 Web 和桌面版安装，需确保 DSH 版本为 `0.2.0-rc.2`。配置持久化存储，卸载不影响历史会话，旨在通过轻量化配置统一 AI 行为。
Read More@gw/dsh-text-inject: DSH Text Injection Plugin
The `@gw/dsh-text-inject` plugin addresses the challenge of managing persistent prompts and project background context in DeepSeek Harness (DSH). It manages content through a standalone Markdown file (`inject.md`), supporting hot reloading and historical backups. The plugin appends content to the system prompt using `## [system]`, injects session context using `## [context]` with `trigger` keyword gating, and provides metadata controls such as `order` and `kind`. It is easy to install, does not access the network, and is privacy-preserving, making it suitable for scenarios requiring frequent prompt adjustments.
Read Moredsh-user-rules: DeepSeek Harness User Rules Configuration Plugin
The `dsh-user-rules` plugin provides a convenient way to manage system prompt rules in the DeepSeek Harness web interface. Without modifying any code, users can add, edit, delete, and reorder rules via the card list on the settings page. These rules are injected into the system prompt as `user-rules`, and support real-time editing with immediate effect and persistent storage, significantly improving development efficiency.
Read Moredsh-djy-xttsc: Global System Prompt Injection Plugin
The dsh-djy-xttsc plugin provides a visual “Big Fat Fish Instructions” configuration in the settings page to inject a global system prompt into all model requests. It supports hot reloading, persistent storage, and version compatibility, ensuring the instruction takes effect across sessions, workflows, and subagents. Installation requires command-line operations, while usage only involves enabling the option in the settings panel or modifying the instruction text. Note that the plugin content is injected as-is and conflicts with other injection sources in the same environment should be avoided. If the configuration does not take effect, enable diagnostic mode to troubleshoot.
Read Moredsh-system-prompt: DeepSeek Harness System Prompt Editing and Rollback Tool
`dsh-system-prompt` is a plugin for DeepSeek Harness that allows managing system prompts on the settings page, solving the issue of being unable to easily roll back changes. It supports three modes—default, instruction appending, and full replacement—and provides features such as loading defaults, previewing, undoing changes, and restoring defaults. The plugin has a 100,000-character text limit, stores its configuration in DSH settings, and retains those settings after uninstallation.
Read Moredsh-instruction-memory: DSH Long-Term Instruction Memory Plugin
`dsh-instruction-memory` is a plugin for DeepSeek Harness (DSH), designed to maintain stable system instructions through a plugin-based architecture. It allows users to manage long-term instructions in settings and supports automatic injection, multi-instance concurrency protection, external file synchronization, and data export and import. Users must install it via the command line and configure its priority and applicable scenarios on the settings page. The plugin requires Node.js >= 20, and the injected content is limited to 40,000 characters. Before installation, ensure compatibility with the host version and be aware of security risks when running in non-loopback address environments.
Read Moredsh-weneedfirst: DeepSeek Harness Plugin, Shaping the Chain of Thought
DeepSeek Harness (DSH)’s `dsh-weneedfirst` plugin is built on the core philosophy of “everything is a plugin,” aiming to govern the model’s chain of thought by injecting a global system prompt. It requires the model to think in English and start each reasoning block with a specific phrase (default: “What we need is”). No build step is required after installation; the plugin is automatically activated via `dsh bundle patch`, but the web application must be restarted after configuration changes or plugin updates. Stateless and lightweight, it supports customizing the starting phrase via a configuration file and is well suited to shaping the model’s reasoning behavior.
Read Moremodole/dsh-plugin-system-prompt-laoboshi: Safely Editing Persistent System Prompts in DSH Web
This article introduces the DeepSeek Harness (DSH) plugin `modole/dsh-plugin-system-prompt-laoboshi`. Maintained by Modole, the plugin addresses the need to securely and persistently manage system prompts in the DSH Web interface. Developers can add it to the current profile using an install command and then edit the prompts in the interface after restarting `dsh web`. When using it, ensure it runs with the current user's permissions and be sure to review the source code.
Read Moredsh-prompt-persona: Editing DSH System Prompt on Settings Page
@xilin3/dsh-prompt-persona is a DSH client plugin maintained by Xilin3, taking over the editing and persistence of deployment:persona on the web settings page. It supports three injection modes: replace, append, and off, along with real-time preview of the complete system prompt, optimistic concurrency saving based on revision, and template variables such as {{model}}, {{cwd}}, and {{provider}}. After installation, simply restart the DSH web to adjust the agent's role settings directly on the settings page, without modifying the Harness itself or manually writing a cordis.patch.yml file.
Read MoreCustom-First Control Prompt: Injecting System Segments and Reference Dialogues Before Persona
@wm-coders/dsh-custom-first-control-prompt is a DSH deployment-side prompt prefix plugin maintained by WM-CODER. It registers ordered system prompt segments before the persona through sections and prepends user/assistant reference conversations as real alternating messages to each ordinary dialogue request via history, intercepting through llm/stream with zero session log writing. The static content facilitates prefix KV cache reuse. Installation is done via `dsh plugin --profile web add`, with configuration written to `cordis.patch.yml` for targeted overrides or saved through the panel. It is suitable for scenarios requiring unified deployment rules and reference conversations; note the token duplication overhead, that it takes effect after a restart, and consider the plugin process permissions.
Read More