Human Skill Tree: 33 Agent Skills That Turn a Chat Model Into a Tutor

The gap Human Skill Tree claims to fill

The README frames the problem as an asymmetry. AI agents gained Skills, MCP servers and tool use, so they can run code, query databases and drive browsers. Humans did not get an equivalent structure. The repository's opening question is blunt: "AI got its skill tree. But what about humans?"

The audience it names is specific. A 35-year-old professional whose degree is dating, a 10-year-old who will graduate into a world where AI does most knowledge work, a PhD student wondering whether five years in a narrow field was the right investment, and a first-generation college student without mentors to explain the unwritten rules. Those four cases map to different parts of the project: K-12 subjects, career skills, and what the README calls social intelligence.

What the project actually ships is narrower than the framing suggests. It is a collection of 33 ready-to-use AI agent skills, plus a demo web app in app/ and documents in docs/. The skills are the product. The app is a way to browse them. Anyone expecting a learning management system with cohorts, grading and analytics will be disappointed by the repository layout, which has no server component beyond the demo.

How the 33 skills change a model's behaviour

The mechanism is not a fine-tune or a retrieval index. It is the Agent Skills standard, the same convention Anthropic published for Claude, applied to pedagogy. Each skill is a self-contained instruction set that a compatible client loads when the task matches. The README describes the standard in its badge row and links to anthropics/skills.

The README's side-by-side comparison is the clearest statement of intent. Asked to teach calculus, a raw model produces a Wikipedia-style wall of text. With the skill loaded, it first asks what a function is in the learner's own words, whether slope is familiar, and whether the goal is exam prep or depth. Asked to fix a bug, the raw model points at line 12 and hands over corrected code. The skill version asks what the learner expected, what actually happened, and which line looks suspicious, then suggests adding a print statement.

That second example is the whole design in miniature. The model withholds the answer and asks for a prediction first. The README grounds this in a randomized controlled trial published in PNAS, which it cites as finding that GPT-4-based tutoring improved high school math performance by 48 to 127 percent, and that students without structured pedagogical design became dependent on the AI and showed reduced skill acquisition. The repository treats that finding as the reason the skills exist at all. It also cites Dunlosky et al. (2013) and Donoghue and Hattie (2021) on practice testing and distributed practice, and Tabibian et al. (2019) on optimized spaced repetition from Duolingo data.

One caveat is visible in the README itself. The citation for the tutoring trial is given as Bastani et al. (2025), PNAS 122(26). The percentage range is quoted from the README, not verified here. Treat the number as the project's own summary of that paper.

Installing Human Skill Tree in Claude Code, Cursor or Gemini CLI

The README does not print a copy-paste install command in the section available here. It states compatibility through badges for Claude Code, Cursor, ChatGPT Codex and Gemini CLI, and links to the Agent Skills standard. The practical route is to clone the repository and point your client at the skills directory.

git clone https://github.com/24kchengYe/human-skill-tree.git
cd human-skill-tree
ls skills/

The listing is the first real check. You should see one directory per skill. If the directory is empty or the clone failed, nothing downstream will work, because the client has nothing to load.

For a client that reads skills from a project folder, the repository itself is the working directory. Start the client there and the skill files sit inside its search path.

cd human-skill-tree
claude

The README does not document a package-manager install, a global CLI, or an environment variable for a skills path. If your client expects skills in a fixed location such as a user-level skills folder, you are copying or symlinking the contents of skills/ there yourself. That step is not described in the README, so verify it against your client's own documentation rather than assuming a convention.

For a first real use, pick a skill that matches a question you already have and ask it the way the README's calculus example does. The observable difference is the first reply: instead of an explanation, you get diagnostic questions about what you already know. If you get the wall of text, the skill did not load.

Where Human Skill Tree is the wrong tool

The project has no persistence layer in the repository layout. There is no database directory, no schema, no account system. Spaced repetition is cited as a design principle, but a skill running inside a chat client has no reliable memory of what you reviewed last Tuesday unless the client itself provides that memory. The README's comparison table lists spaced repetition as a feature of the skill tree, and the skills can prompt a review schedule, but the schedule lives in the conversation or in whatever the host client stores. If you want an algorithmically optimized queue like the Duolingo study the README cites, this is not it.

Assessment is another boundary. The skills ask the learner to self-report what they know. A model cannot verify that a 10-year-old actually understands fractions rather than pattern-matching the questions. The README's own cited research warns about exactly this dependency risk, and a prompt-level guardrail reduces it without eliminating it.

The subject coverage is also uneven by construction. The README advertises more than 800 subjects and 33 skills, which means each skill has to generalize across a wide range. A skill that covers all of K-12 mathematics cannot carry the same depth as a single-subject curriculum. For exam preparation with a fixed syllabus and a fixed date, a structured course with graded problem sets will outperform a conversational tutor that has to infer the syllabus.

Finally, the licence. The README badge and the LICENSE entry point to AGPL-3.0, while the repository metadata reports NOASSERTION. Those two signals disagree, and the LICENSE file in the repository root is the one that governs. If you plan to embed these skills in a hosted product, read that file before you build.

How it compares with a general-purpose tutor prompt

The obvious alternative is a system prompt you write yourself: a paragraph telling the model to ask questions before answering, to use the Socratic method, and to quiz you later. That approach costs nothing and needs no repository.

The difference is packaging and consistency. A single paragraph degrades as the conversation grows and as you switch models. Human Skill Tree splits the instruction into 33 separate skill files, each scoped to a subject or capability, loaded only when relevant. That structure is the same one the Agent Skills standard defines, so a client that already supports skills can consume the repository without custom glue. Your own prompt is portable across anything with a text box; these skills are portable across clients that implement the standard, and the README lists Claude Code, Cursor, ChatGPT Codex and Gemini CLI among them.

The second alternative is a conventional online course. A course gives you a fixed sequence, deadlines and graded feedback from a human or a rubric. Human Skill Tree gives you an adaptive conversation with no deadline and no external grading. The trade is flexibility against accountability. If you finish courses and need the structure, the course wins. If you abandon courses and need a patient interlocutor at 11pm, the skill collection is the more realistic option.

Maintenance, licence and what the repository tells you about cost

The repository is not archived, and the last push to master was on 2026-03-25. The only release listed is v1.0.0, dated 2026-03-08, titled 31 Skills for Lifelong Learning. The README badge now reads 33 skills, so the skill count moved after the release tag without a new release. If you pin to v1.0.0 you get the 31-skill snapshot; if you track master you get the current count and whatever else changed since March.

That gap matters for upgrade planning. There is no changelog entry in the repository describing what changed between the release and the current README, so a team tracking master should diff the skills/ directory rather than assume additive changes. Skill files are plain text, which makes those diffs readable.

The licence situation deserves attention before adoption. The README displays an AGPL-3.0 badge, the repository metadata reports NOASSERTION, and the LICENSE file is the authoritative source. AGPL-3.0 is a strong copyleft licence with a network-use clause. Without giving legal advice: if you fork these skills into a hosted service, the obligation to publish source for modified versions is the question your counsel needs to answer, not a question the README settles. The repository also carries a CONTRIBUTING.md, so outside contributions are anticipated, which means the skill files may change shape over time.

Editorial conclusion

Adopt Human Skill Tree if you already use an Agent Skills compatible client and want the AI to stop handing over answers before the learner has tried. Do not adopt it if you need progress tracking, learner accounts or an offline course, because the repository is a skill collection plus a demo app, not an LMS. Before committing, open the skills/ directory and read one SKILL.md end to end, confirm the AGPL-3.0 terms in the LICENSE file suit your distribution, and check the last commit date on master, which was 2026-03-25.

Frequently asked questions

Official sources

1. 24kchengYe/human-skill-tree on GitHub
2. Issues
3. README
4. Releases