> Author: Guojiz (GitHub: https://github.com/Guojiz) > Project collection: Claude tweak + AI learning OS + vocabulary study + Bilibili subtitle extraction > One-line summary: From "bypassing Claude's limitations" to "turning GitHub into an AI learning brain," these four projects share one design philosophy — pragmatic, lightweight, and never reinventing the wheel.
Why these four projects matter
Thousands of new GitHub projects appear daily; most are either too ambitious ("I'll build a platform") or too bare-bones (incomplete READMEs). Guojiz's four projects strike a rare balance: each solves a specific pain point and delivers maximum value in a minimal viable way.
They form a complete "AI toolchain":
- claude-desktop-tweak-models — access more AI models (tool entry)
- Repo-as-Review-OS — learn efficiently (knowledge management)
- word-match — learn in fragmented time (specific scenario)
- bilibili-ai-subtitle-extractor — extract knowledge from video (information source)
- Doesn't touch API keys — only the frontend validation
- No gateway building — it leverages Claude Desktop's official Third-Party Provider mechanism
- Automatic backup before every patch, with Restore support
- Update-aware — prompts you to re-run after Claude Desktop updates overwrite the patch
- Developer prefers Claude Desktop's UI but wants GLM-5.2: configure Zhipu's Claude-compatible gateway, set model ID to
glm-5.2, keep Artifacts/Projects/MCP features while calling GLM underneath. - Enterprise with in-house models: deploy a Claude-API-compatible service, patch the client for employees, and skip building a custom frontend.
- Windows only — Mac/Linux file structures differ
- Re-patch needed after every Claude Desktop update
- Unofficial — Anthropic could change validation at any time
- Security considerations — modifying installed software requires admin rights, even though the script is open source
- File tree — AI sees your full learning structure
- Writeback — AI writes files, not just suggestions
- History and timestamps — Git records your learning trajectory
- Traceable changes — watch your understanding evolve
- Automation output — GitHub Actions can run scheduled review reminders
- Persistent workspace — repos persist after the chat window closes
- High cold-start barrier — not for users who "just want to jot notes"
- Depends on AI capability — a misread goal derails the whole pipeline
- GitHub learning curve for Git novices
- Privacy concerns even with private repos
- APIs may change — subtitle fields can shift at any time
- Bilibili only — YouTube's anti-automation makes adaptation harder for now
- Depends on subtitle availability — will clearly tell the user if none exist
- Extend your Claude Desktop with tweak-models
- Let AI manage your learning with Review-OS
- Fill fragmented time with word-match
- Extract knowledge from Bilibili with subtitle-extractor
- claude-desktop-tweak-models: https://github.com/Guojiz/claude-desktop-tweak-models
- Removes Claude Desktop's model-ID whitelist; supports third-party LLMs
- PowerShell script; patches frontend validation; fixes ASAR hash
- Windows only; re-patch after updates
- Repo-as-Review-OS: https://github.com/Guojiz/Repo-as-Review-OS
- GitHub-based AI learning OS
- Core model: goal → source → split → model → practice → SRS → automation → understanding
- 230 commits, 20+ docs, complete skill system
- word-match: https://guojiz.github.io/word-match
- Pair-matching vocabulary learning with adaptive selection
- Instant-use web page, gamified
- bilibili-ai-subtitle-extractor: https://github.com/Guojiz/bilibili-ai-subtitle-extractor
- Bilibili subtitle extraction as an agent-agnostic recipe
- Human subtitles first; use existing data, don't detour
- Bilibili only; depends on API stability
claude-desktop-tweak-models: bypassing the wall without breaking the house
Claude Desktop is Anthropic's official desktop client — polished, but hard-limited to Anthropic's own Claude models. Want to use Zhipu GLM-5.2 or OpenAI GPT-5 inside Claude Desktop? Not officially supported.
Rather than building a new client, Guojiz patches the frontend validation logic of Claude Desktop to accept non-Anthropic model IDs.
Technical implementation
A PowerShell script automatically: 1. Locates the Claude Desktop install (Windows App Package directory) 2. Finds the model-ID whitelist check in the frontend JavaScript 3. Patches the check so any model ID passes 4. Fixes the Electron ASAR integrity hash (so tampering isn't detected) 5. Backs up originals and supports one-click restore
Key design choices:
Use cases
Limitations and risks
Verdict: Not piracy — the official client supports third-party gateways but adds an unnecessary frontend model-ID whitelist. This simply removes that unreasonable restriction.
Repo-as-Review-OS: turning GitHub into your AI learning brain
The most complete of the four (230 commits, 20+ doc files). Core question: how do you turn AI from a "chat tool" into a "long-term learning assistant"?
Answer: treat GitHub as a writable timeline for learning — not for storing notes, but for letting AI agents continuously take over and improve your learning process.
Core pipeline: goal → source → split → model → practice set → spaced repetition → automation → new understanding
1. Goal — not vague "learn English" but measurable targets like "score 750 on SAT math," stored in goals/
2. Source — structured records of papers, videos, textbook chapters, mistakes: URL, key concepts, open questions
3. Split — AI dynamically decomposes goals into executable tasks as you progress
4. Model — reusable problem-solving patterns (e.g., "general approach to SAT probability problems") stored in models/
5. Practice set — targeted exercises generated from your weak points, not random question banks
6. Spaced repetition — Ebbinghaus-based scheduling, dynamically adjusted by actual performance
7. Automation — AI handles task distribution, status marking, dashboard updates, and review scheduling
8. New understanding — post-review takeaways feed back into the model library, forming a positive loop
Why GitHub?
Skill system
A skills/repo-as-review-os/SKILL.md main router skill, with sub-skills for goal management, source processing, spaced repetition, and dashboard updates — any AI agent reading SKILL.md knows how to take over the repo.
Limitations
Verdict: Not "another note-taking system" — it turns AI from a Q&A tool into an agent that manages your learning process. Its value lies in design philosophy: GitHub as an AI-writable learning timeline.
word-match: vocabulary study in fragmented time
Extremely lightweight — a web page, open and use.
Mechanisms: 1. Pair matching — English words on the left, Chinese definitions on the right; correct pairs clear, wrong pairs prompt correction 2. Adaptive selection — well-mastered words appear less often; weak ones more often, automatically exposing weak spots 3. Instant feedback — wrong answers show the correct pair to reinforce memory
Use cases: commutes, waiting in line, quick pre-exam vocab review.
Limitations: English–Chinese only (no examples, no pronunciation), possibly no data persistence across refreshes, no full SRS integration.
Verdict: Not a full language platform — it's "a quick vocab pass in fragmented time," an ideal complement to Repo-as-Review-OS.
bilibili-ai-subtitle-extractor: knowledge extraction from Bilibili videos
When AI agents get a Bilibili link, the usual approach is: download the video, run ASR, or wrestle with yt-dlp. But Bilibili already provides subtitle data — both uploader-submitted and auto-generated AI subtitles. This project lets agents read what already exists.
Core principles
1. Prefer Bilibili's existing subtitle data — no video download, no ASR 2. Human/uploader subtitles first, AI subtitles as fallback — quality first 3. Confirm the target language before picking a track — avoid pointless translation chains 4. Don't re-translate another track when target-language subtitles exist — avoid information loss 5. Don't prioritize downloading the video — lightweight first 6. Don't prioritize running ASR — why ignore existing data? 7. Use chapters/timelines from the video description — structured output
Agent-agnostic design
Built as an Agent Recipe (a roadmap for AI), not a platform-bound skill: ChatGPT, Claude, Codex, local agents — even a script that only runs curl commands — can use it. This avoids rapid obsolescence in a fast-moving AI tool landscape.
Technical flow
1. Parse the BV ID from the link 2. Call Bilibili's video info API for aid, cid, title, description 3. Call the subtitle-view API to list available subtitle tracks 4. Smart subtitle selection: human/uploader > AI > translated 5. Download subtitle JSON 6. Merge short sentences, segment paragraphs, incorporate description chapters 7. Output TXT/Markdown
Limitations
Verdict: The value isn't subtitle extraction itself (yt-dlp does that) but giving AI a clear decision framework: which subtitle to use, how to judge quality, how to organize output. Teaching AI to think, not just giving it tools.
The shared design philosophy
1. Pragmatic over perfect — tweak the existing client, don't rebuild it; give AI a roadmap, don't build a universal downloader; one web page, not a full app 2. Lightweight over comprehensive — a PowerShell script, pure Markdown + GitHub, one web page, curl-runnable 3. Open over closed — MIT licenses, an open skill system, agent-agnostic recipes 4. Solving real problems, not minting concepts — each maps to a genuine pain point
Closing: a high-quality solo-developer output model
Not "build a big platform" but "solve one small problem extremely well." As AI tools explode, many developers rush to scale up. Guojiz chose the other path: dig into specific scenarios and deliver value minimally and viably. The payoff isn't user counts or valuations — it's real problems solved and real user time saved.
If you use AI tools, all four are worth trying: