AI-Compass: a curated AI knowledge repository you install as a coding agent skill

What AI-Compass is trying to solve, and for whom

The AI tooling space produces new models, frameworks and papers faster than most developers can track them. AI-Compass answers that with curation rather than automation. The README describes it as a project that "将为社区指引在 AI 技术海洋中航行的方向" and explicitly rejects the label of a simple resource collection, calling itself a systematically organized learning ecosystem. The audience list is broad and worth reading literally: AI beginners, working developers, product managers, researchers, enterprise teams and job seekers. That breadth is also the first thing to be sceptical about. A repository that serves a beginner looking for Prompt engineering basics and a researcher looking for RLHF training frameworks is making a bet that a directory tree can satisfy both. The nine-module structure is how it attempts that. Blog, Code, foundational knowledge, technical frameworks, applied practice, products and tools, learning resources, community platforms, and enterprise open source each get their own top-level directory, and the numbered prefixes in the repository root make the intended reading order visible without opening any file.

The numbered directory tree is the actual product

The mechanism here is not a program. It is a naming convention plus a set of index files. Look at the repository root and the structure reads like a syllabus: 0.AI导航工具集, 1.1 Prompt工程 through 1.4 LLM合集-多模态, then 2.0 Embedding模型, 2.1 LLM训练框架, 2.2 LLM推理框架+部署, 2.3 LLM评估框架 and 2.4 RLHF. Applied architecture gets its own band with 3.0 MCP+A2A, 3.1 RAG+workflow, 3.2 Agent, 3.3 DeepSearch, 3.4 GraphRAG, 3.5 NLP2SQL, 3.6 AI Popular Framework and 3.7 AI Robot. Later bands cover competitions, courses, Python, databases, visualization, ML, CV, recommender systems, RL and knowledge graphs, then platforms, academic tools, articles, forums, software and interview material. Enterprise open source is split by vendor: 10.Datawhale, 10.paddle, 10.华为开源, 10.腾讯, 10.阿里开源. The numeric prefix is doing real work. It means a reader can navigate the repository in a file browser without a search function, and it means the maintainer can insert a new category without renaming everything downstream. The cost is that the numbering is not semantically meaningful. Nothing about "3.4" tells you it is GraphRAG unless you already know, and the README's own table of contents presents the modules in a different order than the directory prefixes imply.

Installing AI-Compass as a knowledge-base skill

The README's headline install path is unusual: instead of telling you to run pip or npm, it gives you a sentence to paste into a coding agent. The prompt instructs the agent to clone the repository if it is not already present, then follow Install.md to install it as a local knowledge-base skill usable by both Codex and Claude Code, and finally to report the install path, the verification result and whether a client restart is needed.

如果本机还没有 AI-Compass，请先 clone https://github.com/tingaicompass/AI-Compass.git，然后按照 https://raw.githubusercontent.com/tingaicompass/AI-Compass/main/Install.md 把它安装为我本机 Codex 和 Claude Code 都能使用的本地知识库 Skill。完成后告诉我安装路径、验证结果，以及是否需要重启客户端。

The repository root contains install.sh, SKILL.md, AGENTS.md and an agents/ directory, which is consistent with the prompt: the shell script performs the installation and the skill metadata files describe the skill to the agent. The README says the prompt causes the agent to clone the repository, run the root-level install script and verify the symlink. It does not document what the script writes, where the symlink points, or how to undo it.

If you would rather look before you leap, clone the repository yourself and read the script first.

git clone https://github.com/tingaicompass/AI-Compass.git
cd AI-Compass
less install.sh

Reading install.sh before executing it is the only way to know what the symlink targets and whether the script touches files outside the cloned directory, because Install.md is referenced by URL in the prompt rather than reproduced in the README. Once installed, the README says you can ask the agent about AI techniques, models, projects and the weekly updates directly, which implies the agent is reading from the cloned tree rather than from a packaged index.

Using the repository without an agent

The agent path is optional. The README states that if you are using the repository as a knowledge base, you should look first at weeklyHighlights/latest.md and weeklyHighlights/INDEX.md, and that new weekly entries should update both files following weeklyHighlights/TEMPLATE.md. That is a concrete convention with a concrete failure mode: the weekly digest is only as current as the last person who remembered to update two files in sync. The README lists twenty-three past issues in a collapsed section, each covering a batch of releases such as Qwen3-Max, DeepSeek-V3.2, Sora 2 and Hunyuan 3D. The titles are the useful part; they tell you which week covered which model family, so you can jump straight to the issue that mentions the thing you are trying to place. The repository also accepts resource submissions through GitHub Issues using a markdown template with fields for tool name, homepage, category, one-line summary, core features, audience, licence or pricing status and submitter notes. The README states that the project does not sell ranking positions and that maintainers sort submissions by content quality and usefulness. That is a stated policy, not a verifiable one, and there is no described review SLA.

Where AI-Compass stops being the right tool

The repository has no retrieved releases and no stated licence. For an engineering team that needs to vendor a component, pin a version or clear a legal review before shipping, that is disqualifying on its own: you cannot audit a licence that is not declared. The README does not document rollback for install.sh, does not describe what happens if you run it twice, and does not say how to uninstall the skill from Claude Code or Codex. If your workflow depends on reproducible environments, an install path whose only documented verification step is "the agent tells you it worked" is weaker than a package with a lockfile. There is also a language boundary. The README, the module names and the weekly digests are in Chinese, with an English README-EN.md present at the root but no indication in the README of how completely it mirrors the primary file. Finally, the repository is a curated index, not an implementation. If you need a library that embeds text, serves a model or runs a graph query, AI-Compass will point you at one; it will not be one. Treating it as a dependency rather than a reading list is a category error, and the numbered directories will not save you from it.

How it compares to a framework like LangChain

The nearest thing people reach for when they want "the AI ecosystem in one place" is a framework such as LangChain, and the difference in approach is total. LangChain is executable Python that you install as a package, import, and call; its abstractions are code objects with versioned releases and changelogs. AI-Compass is a directory tree of links, markdown files and weekly digests, plus an optional skill install that makes the tree readable by a coding agent. LangChain answers "how do I build this"; AI-Compass answers "what exists, which category does it belong to, and what changed this week." The trade-off runs both ways. A framework gives you a stable interface and takes on upgrade churn; a curated repository gives you breadth and takes on staleness, because every link can rot and every weekly issue can go unread. If you want to know which inference framework to evaluate, the 2.2 LLM推理框架+部署 directory is a reasonable starting point. If you want to serve a model today, you still need to pick one of the things listed there and read its own documentation.

Maintenance, licence and what to check first

The repository is not archived, and the last push was on 2026-09-08, which is recent enough that the weekly cadence described in the README appears to be holding. That is the strongest maintenance signal available here, and it is worth stating precisely: recency of the last push, nothing more. There are no retrieved releases, so there is no version history to inspect and no upgrade path to plan around. Upgrading means re-cloning or pulling the default branch, and because the install path is a symlink to the clone, whatever state your working copy is in is the state your agent reads. The licence field is unknown, and the README does not name one. That matters if you intend to redistribute the content, mirror the weekly digests internally or bundle the repository into a product. It does not matter much if you are reading it. The README does state that commercial tools may be submitted but must disclose pricing, free tiers or trial terms, and that paid ranking is not offered, which is a content policy rather than a licence grant. Verify the licence file directly in the repository root before you rely on any assumption about reuse.

Editorial conclusion

Adopt AI-Compass if you read Chinese and want a single repository that maps LLM training, inference, RAG, Agent and RLHF material into one directory tree, or if you use Claude Code or Codex and want the repository reachable as a local skill through install.sh. Do not adopt it if you need an English-first reference, a versioned API, a package you can pin, or a licence you can verify before redistribution. Before you clone anything, open the repository root and confirm three things: whether a LICENSE file exists at all, what install.sh actually does on your machine, and whether the numbered directories are maintained as link lists or as working code. The README's own install prompt tells the agent to report the install path, the verification result and whether a client restart is needed, so treat those three answers as the acceptance test rather than trusting the badge row.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. tingaicompass/AI-Compass on GitHub