Guojiz Open-Source Suite: Claude Desktop Patch, GitHub as AI Learning OS, Word Matching, and Bilibili Subtitle Extraction

Overview

GitHub user Guojiz (https://github.com/Guojiz) maintains four focused open-source projects that, taken together, cover the AI-assisted personal productivity loop:

• claude-desktop-tweak-models — broaden model choice
• Repo-as-Review-OS — manage long-term learning
• word-match — practice vocabulary in short sessions
• bilibili-ai-subtitle-extractor — harvest knowledge from video

A unifying philosophy runs through all four: *pragmatic, lightweight, and never reinvent the wheel.*

---

1. claude-desktop-tweak-models — Bypass the wall, don't break the house

What it solves

Anthropic's Claude Desktop has a polished UI and features like Artifacts, Projects, and MCP, but its front-end enforces a model-ID whitelist limited to Anthropic's own models. This script removes that restriction so users can plug in any compatible third-party provider.

How it works

A PowerShell script that:

1. Locates the Claude Desktop installation under the Windows App Package directory. 2. Identifies the front-end JavaScript model-ID validation logic. 3. Patches the validation so any model ID passes. 4. Repairs the Electron ASAR integrity hash to avoid tampering detection. 5. Backs up the original files and exposes a one-click restore.

Design highlights

• No API key handling — only front-end validation is modified.
• No custom gateway — uses the official Third-Party Provider path already supported by Claude Desktop.
• Auto backup and rerun prompt after desktop updates.

Example use cases

• Developer who prefers Claude Desktop's UI but wants GLM-5.2 behind a Claude-compatible gateway.
• Enterprise with an in-house Claude-compatible model serving a polished client to staff without rebuilding a front end.

Limitations

• Windows only (Mac/Linux file layouts differ).
• Patches are overwritten by Claude Desktop auto-updates.
• Unofficial; Anthropic may change validation at any time.
• Requires admin rights and modifies installed software.

Verdict

Not piracy — simply enabling the third-party provider mechanism that the official client already supports.

---

2. Repo-as-Review-OS — Turn GitHub into an AI learning brain

The most substantial project in the suite (230+ commits, 20+ documents).

Core question

How do you turn AI from a chat tool into a long-term learning assistant?

Core model: goal → source → split → model → practice set → spaced repetition → automation → new understanding

1. Goal — concrete and measurable, stored in goals/. 2. Source — structured notes from papers, videos, textbooks (URL, key concepts, doubts). 3. Split — AI breaks the goal into dynamic, progress-aware sub-tasks. 4. Model — reusable problem-solving patterns extracted from sources, stored in models/. 5. Practice set — exercises generated against weak points, not random. 6. Spaced repetition — Ebbinghaus-style intervals tuned to real performance. 7. Automation — AI schedules tasks, tracks completion, updates dashboards, queues reviews. 8. New understanding — each review cycle feeds the model library, closing the loop.

Why GitHub?

• File tree shows complete learning structure.
• Writeback lets AI persist changes, not just suggestions.
• Git history captures learning trajectory.
• Timestamps record when each concept was studied.
• Diffs reveal how understanding evolves.
• GitHub Actions automate reminders.
• The repo survives chat windows closing.

Skill system

skills/repo-as-review-os/SKILL.md acts as the router; sub-skills cover goal management, source processing, spaced repetition, and dashboard updates. Any AI agent reading the SKILL file can take over the repository.

Limitations

• Steep cold-start; not for casual note-takers.
• Misread goals derail the entire pipeline.
• Requires Git/GitHub familiarity.
• Privacy-sensitive users may hesitate to host learning records on GitHub, even in a private repo.

Verdict

Not another note system — an operating system that turns AI into a learning-process agent.

---

3. word-match — Vocabulary for fragmented time

A lightweight web app: open and play.

Mechanics

• Pairing — click English word to its Chinese meaning; correct pairs disappear.
• Adaptive sampling — frequency rises for weak words, falls for mastered ones.
• Instant feedback — wrong answers show the correct match immediately.

Use cases

Commutes, queues, elevator waits, last-minute vocabulary sweeps.

Limitations

• Chinese-English only; no examples, no pronunciation.
• Persistence depends on browser storage.
• No deep SRS integration like Repo-as-Review-OS.

Verdict

A focused vocabulary drill for short pockets of time — a natural complement to Repo-as-Review-OS.

---

4. bilibili-ai-subtitle-extractor — Knowledge extraction from Bilibili video

Many AI agents default to downloading video and running ASR. Bilibili already serves subtitles — both uploader-uploaded human transcripts and AI-generated ones. This project teaches agents to read what already exists.

Guiding principles

1. Prefer Bilibili's existing subtitle data — no video download, no ASR. 2. Human/uploader subtitles first; AI subtitles as fallback. 3. Confirm the target language before picking a track. 4. If a target-language subtitle exists, do not translate another track on top. 5. Don't download the video first. 6. Don't run ASR if data already exists. 7. Use chapter markers and timestamps from the video description.

Agent-agnostic design

Delivered as an Agent Recipe rather than a platform-specific Skill — usable by ChatGPT, Claude, Codex, local agents, or even a plain curl script. This avoids obsolescence as AI tooling evolves.

Pipeline

1. Parse the BV ID from the URL. 2. Call Bilibili's video info endpoint for aid, cid, title, description. 3. Read the subtitle list from the danmaku view endpoint. 4. Smart-select: human subtitle > AI subtitle > translated subtitle. 5. Download the subtitle JSON. 6. Merge short lines, segment by description chapters, output readable text. 7. Export TXT/Markdown.

Limitations

• Bilibili's APIs and subtitle fields may change.
• Only Bilibili is supported — YouTube has stronger anti-bot defenses.
• If a video has no subtitles, the project (correctly) refuses to guess.

Verdict

The value is not "downloading subtitles" (yt-dlp can do that) — it's teaching AI a clear decision policy: which subtitle, why, how to structure output.

---

Shared design philosophy

1. Pragmatic over perfect — extend existing tools instead of replacing them. 2. Lightweight over comprehensive — a PowerShell script, plain Markdown, a single web page, a curl recipe. 3. Open over closed — MIT-licensed, documented Skill files, agent-agnostic recipes. 4. Solve real pain points:

• "Claude Desktop only runs Claude models" → tweak-models
• "AI chats forget my learning progress" → Repo-as-Review-OS
• "I have five minutes on the subway" → word-match
• "How do I extract knowledge from a Bilibili video?" → subtitle-extractor

---

Reference

• claude-desktop-tweak-models — https://github.com/Guojiz/claude-desktop-tweak-models
• Removes Claude Desktop's model-ID whitelist; PowerShell patch + ASAR hash repair; Windows only.
• Repo-as-Review-OS — https://github.com/Guojiz/Repo-as-Review-OS
• GitHub-based AI learning OS; goal→source→split→model→practice→SRS→automation→understanding; full Skill system.
• word-match — https://guojiz.github.io/word-match
• Browser pairing game; adaptive sampling; Chinese-English vocabulary.
• bilibili-ai-subtitle-extractor — https://github.com/Guojiz/bilibili-ai-subtitle-extractor
• Agent-agnostic subtitle recipe; human subtitles preferred; no ASR, no video download.