Preface¶

After finishing the README for an open-source project, the product introduction, or the course syllabus, the most time-consuming task is rarely writing another long article. Instead, it is adapting the same core facts into formats suitable for Xiaohongshu, Zhihu, WeChat Official Accounts, Moments, short-video platforms, and overseas platforms respectively. Copy-pasting directly will make your post look like an ad on Moments, clickbait on Zhihu, and an unscripted monologue on video platforms.

DeepSeek Harness (dsh) wraps models, tools, skills, and interfaces as plugins. Community maintainer lhmd developed a development and runtime plugin called dsh-promotion-toolkit: it first extracts core viewpoints, citeable evidence, risk boundaries, and original links from the source text, then adapts the content to native reading formats for each platform. The community directory categorizes it under “Development & Runtime”, with a one-sentence description: “Turn any of your ideas into native promotional content for every platform.”

This article is organized after cross-checking with the community directory page, GitHub repository README, package.json, cordis.patch.yml, src/index.js, runtime Skills, and official DeepSeek Harness documentation: covering what it is, what the tool actually does, how to install it, how to use it, and its usage boundaries.

What is it¶

dsh-promotion-toolkit is a native DeepSeek Harness plugin with the npm package name @lhmd/dsh-promotion-toolkit, current version 0.1.0, MIT license, and main language JavaScript. The repository is maintained by lhmd at lhmd/dsh-promotion-toolkit. As of 2026-08-18, it has 9 stars on GitHub; the community directory page showed 8 stars at the time of crawling, so refer to the repository page for the latest count.

It does not solve the problem of “generating a shorter summary”, but rather tasks like turning product introductions, open-source READMEs, articles, courses, event descriptions, or unorganized thoughts into multi-platform promotional drafts that can be directly edited. The README outlines the workflow in three steps:

Any idea / original text / README
            |
            v
Core Idea + Evidence + Risk Boundaries + Original Link
            |
            v
17 platform-native long-form articles, posts, video scripts, and CTAs

The repository follows DSH’s profile-bundle structure: cordis.patch.yml integrates the package into the configuration layer, src/index.js exports apply(ctx) and registers the tool viral_kit, src/skill.js registers the runtime Skill viral-kit, and skills/viral-kit/SKILL.md is a human-readable mirror. Peer dependencies declare @deepseek-ai/dsh-tools and cordis.

Two key points need clarification. The viral_kit tool itself calls the local heuristic engine buildKit(), with the return value containing generatedBy: "dsh-promotion-toolkit/local-heuristics": it splits sentences, scores them, extracts keywords, and fills them into Chinese and English platform templates, without calling a large language model (LLM) internally. The runtime Skill, by contrast, requires the agent to first use this tool for the initial pass, then polish the output according to platform conventions, and must not add unprovided user data, metrics, dates, prices, or links. The complete case in the repository was generated end-to-end using deepseek-v4-pro, with a generation timestamp of 2026-08-13T18:19:52Z.

The core philosophy of the official DeepSeek Harness repository is “everything is a plugin”. deepseek-harness-plugin.com is a community-curated plugin directory, not officially affiliated with DeepSeek or Magic Square, and should not be treated as an official app store.

Core Features¶

Extract facts first, then adapt to platform entry points¶

The plugin’s design focus is not on shortening a paragraph 17 times, but first fixing “what can be said”, then modifying “how to say it”. The tool outputs structured JSON with the following fields:
- coreIdea: Core viewpoints extracted from the original text
- evidence: Evidence sentences that need to be retained
- shareability: A 0-100 score and signal list (e.g., whether it contains numbers, comparisons, questions)
- titles: Three title angles
- hook / quote: Opening hook and quoteable short sentences
- platformDrafts: Completed drafts for 17 platforms
- publishText: Completed draft for the currently selected platform
- imagePrompt, hashtags
- riskFlags: Pre-publication verification items such as numbers, quotes, and absolute statements

If sourceUrl is provided, it will be appended to the CTA of each platform; no legitimate http(s) links will be invented if none are provided. The default input limit is 24000 characters, which can be modified in the tool parameters or cordis.patch.yml via maxInputChars.

17 platforms categorized by reading habits¶

cordis.patch.yml divides platforms into two groups. Chinese platforms default to Chinese output: xiaohongshu, zhihu, wechat, wechat_moments, weibo, bilibili, douyin, kuaishou. Global platforms default to English output: x, threads, linkedin, reddit, tiktok, youtube, medium, facebook, instagram. Explicitly passing language as zh, en, or auto will override the default setting. The default platform is Xiaohongshu, and the default language is Chinese.

The README specifies the output format conventions for each platform:

The Skill also specifies: do not copy drafts from one platform directly to another when dedicated drafts exist; use more visual symbols on Xiaohongshu, limit emojis to a maximum of two on Moments, and use emojis sparingly on other platforms.

Built-in reproducible examples¶

The README states that the examples were fully generated using deepseek-v4-pro, covering 4 long-form topics in both Chinese and English, with 17 platform versions per topic. The four topics in examples/live-demos.md are:
1. Product launch long article
2. Indie developer log
3. Creator workflow
4. Promotional实战 for this repository (the source material is the repository itself)

The complete structured results are in examples/live-demos.json, the online bilingual display page is at lhmd.top/dsh-promotion-toolkit, and the source files are in the docs/ directory. These are demonstration drafts provided by the repository, and you still need to cross-check them against the original text using riskFlags before publishing. Do not treat the effect descriptions in the demo copy as verified data.

Installation and Activation¶

The installation command provided on the community directory page is as follows, to be run in the DeepSeek Harness terminal:

dsh plugin add github:lhmd/dsh-promotion-toolkit

The dsh CLI will parse the plugin from GitHub and install it to the current configuration. For reproducible installations, the directory page recommends pinning the commit hash. The latest commit on the current main branch is c4e19742d4399e0282ab70829e412fb47bbb42a7 (2026-08-13):

dsh plugin add github:lhmd/dsh-promotion-toolkit#c4e19742d4399e0282ab70829e412fb47bbb42a7

The directory page also notes that the plugin runs with the permissions of the current dsh process, and may execute code during installation. Please review the source code repository and license before installing.

If you clone the repository locally and link it to a web profile, the README provides the following steps:

npm install -g @deepseek-ai/dsh
dsh plugin --profile web add /absolute/path/to/dsh-promotion-toolkit
dsh web

Replace the path with your local absolute path. Local validation commands:

npm test
npm run check:release
npm run demo

The official DeepSeek Harness repository is currently in developer preview, and the documentation notes that there may be breaking compatibility changes. The plugin’s package.json lists the version as 0.1.0. If the tool does not appear after installation, first verify the dsh version and whether the profile has loaded this bundle.

Typical Usage¶

Generate initial drafts via the command line¶

The repository provides a CLI with the entry point scripts/viral-kit.mjs, and the bin name in package.json is dsh-promotion-toolkit. The README example is as follows:

node scripts/viral-kit.mjs \
  --text "Paste your product introduction, article, course, event, or any idea" \
  --language auto \
  --platform xiaohongshu \
  --source-url https://github.com/lhmd/dsh-promotion-toolkit \
  --json

--platform determines which draft is selected for publishText; the full platformDrafts are still included in the output. --source-url is used for original links such as repository, product page, article page, or registration page. You can also read from standard input:

cat article.md | node scripts/viral-kit.mjs --stdin --json

--demo will run the script using a built-in Chinese product introduction from the repository, to confirm that the script works. Without --json, the script will print core viewpoints, evidence, titles, platform drafts, image prompts, tags, and risk items to the terminal.

Hand off to the agent in DeepSeek Harness¶

After successful installation, the tool viral_kit and Skill viral-kit will appear in the conversation. The tool parameters are:
- text (required): Original text to be packaged
- language: auto / zh / en
- platform: One of the 17 platforms above, or general
- sourceUrl: Standardized link
- maxInputChars: Input character limit

The Skill workflow is as follows: first guard the original text boundaries, extract viewpoints, evidence, audience, and standardized URLs; if the user specifies a platform, prioritize returning that platform’s draft, otherwise return the full matrix; Chinese platforms default to Chinese, global platforms default to English; then adjust the structure according to the platform, without altering the facts.

A reproducible usage scenario is: paste the README or product introduction into the conversation, specify the platforms you want to post to, and include the repository address. For example: “Turn this README into posts for Xiaohongshu and Zhihu, use the GitHub repository as the source link.” The agent should first call viral_kit, then polish the results based on the tool output, rather than skipping the tool and writing generic copy directly.

Applicable Scenarios and Notes¶

This tool is suitable for people who already have a verifiable original text: open-source maintainers releasing version notes, indie developers sharing product updates, courses or events that need to sync content across multiple platforms, and writers looking to split long articles into short posts and video scripts. It does not take responsibility for proving dissemination effects, nor will it fill in user counts, revenue, or dates when no data is provided in the original text.

Please note the following when using it:
1. The first pass of the tool is a heuristic template. When run separately via the command line, the output is a local template fill, not a final LLM-generated draft. To match the effect of the repository’s demo page, you need to cooperate with the model in DeepSeek Harness and rewrite according to the Skill constraints.
2. Verify before publishing. The riskFlags in the output will mark numbers, quotes, and absolute statements. Statements such as “saves time” or “improves efficiency” that appear in the demo copy should not be treated as verified product metrics directly.
3. You must provide the links. Without sourceUrl, the plugin will not invent repository addresses or registration pages.
4. Permissions and license. The plugin runs with the permissions of the current dsh process. Please read the source code and MIT license before installing. For reproducible installations, pin the commit hash mentioned above.
5. You still need to comply with platform rules. Each platform’s requirements for advertising identifiers, external links, and tag counts are subject to the platform’s current rules, and the plugin will not conduct additional audits.

Summary¶

dsh-promotion-toolkit turns “one original text, multiple entry points” into a DSH plugin: the local tool first extracts viewpoints and evidence, and the Skill then asks the model to adjust the structure for 17 platforms without altering the facts. For people who already have a README or product long text but are stuck on multi-platform rewriting, it provides a verifiable draft matrix, rather than a publish button that guarantees viral success.

Community Directory: https://deepseek-harness-plugin.com/zh-CN/plugins/dsh-promotion-toolkit/

GitHub: https://github.com/lhmd/dsh-promotion-toolkit

Online Demo: https://lhmd.top/dsh-promotion-toolkit/