Human Skill Tree: 33 AI Agent Skills for Structured Learning

The gap Human Skill Tree is aimed at

A general chat model will answer a request to "teach me calculus" with a summary. The README frames this as the central problem: agents gained tool use, code execution and browsing, but they carry no pedagogical structure, so they "know everything but teach nothing." Human Skill Tree is the prompt layer meant to sit between the learner and the model.

The intended audience is broad and the README names four cases directly: a 35-year-old professional whose degree is going stale, a 10-year-old who will graduate into automated knowledge work, a PhD student questioning a narrow five-year investment, and a first-generation college student without mentors. Those are different problems wearing one name. The professional wants a catch-up path, the child needs a curriculum, and the student without mentors needs the unwritten rules of a field. A skill library can serve all three only if the skills themselves differ in kind, which is why the repository ships 33 of them rather than one tutor prompt.

The design bet is that behaviour, not knowledge, is what a model lacks. That bet is testable and the README leans on published evidence for it, citing a PNAS randomized controlled trial in which GPT-4 tutoring improved high school math performance by 48 to 127%, with gains disappearing when students were given answers instead of hints.

How the skills change an agent's behaviour

The mechanism is instruction, not infrastructure. A skill is a set of directions the agent loads before responding, and the README's side-by-side example shows what changes. Asked to teach calculus, an unmodified model produces a textbook passage. With the skill loaded, it first asks what a function means to the learner, whether slope is familiar, and whether the goal is exam preparation or depth, then builds from the answer.

The second example is the more interesting one. Asked to fix a bug, the default model names the line and hands over the correction. The skill version asks what the learner expected, what actually happened, and which line looks suspicious, then suggests adding a print statement. That is deliberate withholding, and it matches the PNAS finding the README cites: hints instead of answers preserved learning gains.

The same pattern recurs across the skill set. The README lists active recall, spaced repetition and Socratic dialogue as the techniques the skills encode, alongside culturally aware scenario simulation for social skills. The repository layout supports this reading: skills/ holds the skill definitions, app/ holds the web demo, and scripts/ and docs/ sit alongside them. The learning logic lives in text files that the agent reads, not in a service that scores you.

Installing it into Claude Code, Cursor or Gemini CLI

The README points to an installation section and lists Claude Code, Cursor, ChatGPT Codex and Gemini CLI as compatible clients, but the excerpt does not reproduce the exact commands. What can be confirmed from the repository layout is that skills live in the skills/ directory at the top level, which is where a client-compatible skill file would be read from. Clone the repository first and inspect that directory before wiring anything into a client.

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

Listing skills/ is the first real use, and it matters more than it sounds. With 33 skills covering 800+ subjects across K-12, career and social intelligence, the folder is the only reliable way to see what you actually get and which file corresponds to the subject you want. The README also links a hosted demo at humanskilltree.yechengzhang.com, which is the faster way to judge whether the tutoring style suits you before touching a client.

Once you have identified a skill file, the general pattern for agent clients is to make it available to the agent and then invoke the skill by name in a session. The repository states compatibility with Claude Code, Cursor, ChatGPT Codex and Gemini CLI, and the README carries badges for each. Treat the badge as a claim of compatibility rather than a tested guarantee; the excerpt does not include per-client setup steps, so follow the installation section in the README for your specific client.

Where the skill-library approach breaks down

A skill file cannot remember you. Spaced repetition requires a schedule, and a schedule requires state: what you reviewed, when, and how well it went. The README cites research on algorithmically optimized review intervals from a study of 12 million Duolingo learners, but the repository as described is a collection of agent skills plus a web demo, not a scheduling engine. If your goal is retention over months, the agent will need to reconstruct context each session, and the quality of that reconstruction depends on the client's memory features, not on this project.

The second limitation is the one the README's own evidence implies. Withholding answers works, and it is also the behaviour users complain about. A learner who wants the fix, not the Socratic route, will find the debugging skill actively obstructive. That is a design choice, not a defect, but it means the tool is wrong for anyone using an agent as a reference rather than a tutor.

Third, the range is a liability as much as a selling point. Thirty-three skills spanning a 10-year-old's curriculum and a professional's career pivot cannot be equally deep. The README advertises 800+ subjects, and breadth at that scale usually means each path is a structured starting point rather than a complete course. Anyone expecting a finished syllabus for a specific exam should check the relevant skill file before assuming coverage.

How this differs from a conventional learning app

The obvious comparison is a spaced-repetition application such as Anki, and the difference is architectural rather than cosmetic. Anki owns the schedule. It stores your cards, computes review intervals, and tells you what to study today. Human Skill Tree owns the conversation. It shapes how the model responds in the moment, and it delegates scheduling either to the model's context or to the learner.

That trade produces opposite failure modes. Anki will keep you reviewing material you have already mastered if your cards are badly written, and it will never explain a concept you failed to understand when you made the card. Human Skill Tree can explain anything on demand, and it will not tell you that you are three days overdue on quadratic equations. The README's own framing, that AI agents have no pedagogical structure, is a claim about the conversation layer, and the project answers that claim at the conversation layer.

The second alternative is simply the model without skills. That is a real comparison, not a straw man: the README's side-by-side table exists precisely because the default behaviour is competent and fast. If you want an answer, the default is better. The skills exist for the case where the answer is the thing you are trying to avoid.

Licence, maintenance and what an upgrade costs you

The licence is the first thing to resolve before adoption, and the repository contradicts itself. The README carries an AGPL-3.0 badge and links to a LICENSE file, while the repository metadata reports NOASSERTION. AGPL-3.0 is a strong copyleft licence with a network-use clause, which matters if you plan to fold these skills into a hosted product rather than use them locally. Read the LICENSE file itself rather than the badge, and treat the discrepancy as unresolved until you do. This is a description of the files, not legal advice.

The last push to the default branch was on 2026-03-25, and the only release listed is v1.0.0 from 2026-03-08, described as 31 skills. The README currently advertises 33, so the skill count has moved since that release without a corresponding tag. That is a small but concrete signal about how releases are cut here: the repository moves ahead of its version numbers.

Upgrade cost is low in the mechanical sense. Skills are files, so pulling a newer version means replacing text, not migrating a database. The real cost is behavioural. A skill update can change how the agent questions you, and if you have built a study routine around a particular Socratic style, a rewrite of that skill file is a change you will feel. Pin the version you start with, and diff skills/ before pulling again.

Editorial conclusion

Adopt Human Skill Tree if you already work inside an agent client such as Claude Code, Cursor, Gemini CLI or a ChatGPT Codex environment and you want the tutoring behaviour to be consistent across sessions instead of retyped each time. Do not adopt it if you need progress tracking, a gradebook or an offline study app; the README describes prompt skills and a separate web demo, not a learning management system. Before installing, confirm which licence actually governs the repository, since the badge says AGPL-3.0 while the repository metadata reports NOASSERTION, and open skills/ to check that the K-12 through career range matches the subject you intend to study.

Frequently asked questions

Official sources

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