Introduction

DeepSeek Harness (DSH) extends capabilities via plugins. In agent workflows, links and local audio/video are often processed into one-time summaries, but subsequent steps require cleaning short links, extracting subtitles or transcribing, saving evidence, and importing into a note library. Here is an introduction to learn-skill.

Overview

learn-skill is a DSH skill plugin provided by the gtbwpkwjnb-alt/learn-skill repository. It is an Agent-native learning pipeline: taking in share text or local audio/video, passing through short link/promotional parameter cleaning, subtitle-first and recoverable transcription, keyframe OCR, and full-text Map-Reduce-Verify to output hierarchical knowledge cards, which can be imported into Obsidian, Siyuan, or saved as local Markdown. The license is MIT.

Core Features

Input and Extraction

  • Network adaptation: Automatically detects GFW and filters available platforms.
  • 8 Platform Support: Douyin · TikTok · Bilibili · YouTube · Podcasts · WeChat · Xiaohongshu · Local Files.
  • Batch Learning for Douyin Homepages: Automatically enumerates /aweme/post pagination, deduplicates by post ID, and reports page count discrepancies.
  • Chinese Interface OCR: PP-OCRv6 (isolated environment) → Legacy PaddleOCR → Tesseract.

Analysis and Output

  • Current Agent Evidence Analysis: Full-text segmentation Map → Reduce aggregation → Local evidence Verify; uses the current agent’s session model after the skill is activated, without relying on external model APIs.
  • Hierarchical Output: Chapter breakdown + Mermaid mind map + 3-paragraph summary; knowledge cards can include chapters, mind maps, highlights, terminology, and flashcards.

Tasks and Import

  • Recoverable Tasks: SQLite task ledger + artifact manifest + resumable downloads; cleans up local artifacts by default after successful import.
  • Dual Knowledge Base Import: Obsidian (default) + Siyuan fallback + Local Markdown.
  • Safety Boundaries: The main script only performs local extraction and evidence saving, and does not read or invoke external model APIs.

Installation and Usage

First, install the plugin in DSH:

dsh plugin --profile web add github:gtbwpkwjnb-alt/learn-skill

Restart the web interface after installation:

dsh web

After installation, the skill appears in Settings → Plugins and is lazily loaded as needed. The plugin shell is provided by SKILL.md in the repository root, while cordis.patch.yml + index.js serve as the plugin shell; it has no npm dependencies and can be installed directly from GitHub. The skill format is a kebab-case name + name/description frontmatter, consistent with the dsh skill contract.

System Dependencies:

  • ffmpeg
  • Python 3.10+

To enable corresponding import or extraction capabilities, configure the following variables:

Variable Purpose
SIYUAN_TOKEN Required for importing to Siyuan
BILI_COOKIE To obtain high-quality subtitles
OBSIDIAN_VAULT Obsidian vault path

Typical Usage

As an AI Skill

You can hand links to the skill, for example:

zhixi https://www.bilibili.com/video/BV1GJ411x7h7
zhixi https://v.douyin.com/xxxxx/
learn this video https://b23.tv/xxx

Command Line

Single processing:

python zhixi-learn.py "https://www.bilibili.com/video/BV1GJ411x7h7"

Batch processing:

python zhixi-learn.py "url1" "url2" "url3"

Douyin creator homepage:

python zhixi-learn.py "https://www.douyin.com/user/..."

Extract content only:

python zhixi-learn.py "url" --extract-only

Preview:

python zhixi-learn.py "url1" "url2" --dry-run

Resolve short links without accessing the network:

python zhixi-learn.py "https://b23.tv/xxxxx?spm_id_from=333" --dry-run --no-resolve-links

Files and Retention Policy

  • Artifacts during processing are located in learn-output/_tasks/<task_id>/.
  • After importing to Obsidian, the note is located at <Vault>/learn/<topic>-<YYYY-MM-DD>.md.
  • Local artifacts are cleaned up by default after successful import; use --keep-local if retention is needed.

Applicable Scenarios and Notes

Suitable for the following scenarios:

  • Frequently extracting content from Douyin, TikTok, Bilibili, YouTube, Podcasts, WeChat, Xiaohongshu, or local files.
  • Want to import processing results into Obsidian or Siyuan.
  • Need to batch process Douyin homepage links.
  • Want to use the current agent’s session model for evidence analysis within the DSH skill workflow.

Notes before use:

  • The plugin runs with the permissions of the current dsh process; check the repository source code and license before installing.
  • The license is MIT.
  • The main script only performs local extraction and evidence saving, and does not read or invoke external model APIs; the current Agent evidence analysis uses the current agent’s session model.
  • The tested compatible version is dsh v0.1.1-rc.2 (2026-08-22); preview versions iterate quickly, so refer to the official documentation for API changes.
  • Network adaptation automatically detects GFW and filters available platforms.

Links

Repository:

https://github.com/gtbwpkwjnb-alt/learn-skill

Plugin catalog directory page:

https://www.skillhub.cn/plugins/gtbwpkwjnb-alt/learn-skill

This directory page comes from the plugin catalog and has not been directly confirmed in the gathered materials.