A diagnosis-first STEM / AI-CS tutor skill that finds the learner's current knowledge gap before teaching.
English | 中文
User Guide · Command Surface · Examples · Changelog
Universal Diagnostic Tutor Skill is a Markdown-only AI tutor behavior layer for STEM, mathematics, programming, AI/CS, and exam review. It diagnoses the learner's gap, chooses the next best teaching step, and checks evidence before advancing.
The difference: it is not answer-first. It identifies the subject, concept, prerequisite, notation, method, or reasoning gap before choosing a compact teaching move. It is not a course platform, database, RAG system, or hidden-memory service.
| Where you use AI | Start here |
|---|---|
| Ordinary ChatGPT, Gemini, DeepSeek, Doubao, Kimi, or Qwen chat | Copy the Lite Prompt |
| Codex or Claude Code-style agent | Use the Full Skill and follow the install guide |
| Codex with visible Skill entrypoints | Choose the most relevant tutor-* entrypoint |
| Custom bot or API prompt | Choose an adapter in Portability |
New to the project? Read the User Guide. Installation and updates live in INSTALL.md, while entrypoint details live in COMMAND_SURFACE.md.
| Capability | What it does |
|---|---|
| Diagnosis-first tutoring | Locates the subject, concept, prerequisite, notation, method, or reasoning gap |
| Learning Architecture | Clarifies broad goals, builds a compact knowledge map, and selects one next step |
| Practice & Mastery Loop | Generates targeted practice, waits for an answer, grades qualitatively, repairs mistakes, and checks readiness |
| Skill entrypoints | Exposes focused tutor-* doors into one shared Tutor system |
| STEM Exam Track | Supports university STEM, postgraduate math, and CS review without prediction or score promises |
| Topic Scan + Trusted Resources | Uses reliable learning resources when they improve the current teaching step |
| Knowledge Link Cards | Explains one to three strongly related concepts when they block the current task |
| Learning State Cards | Creates visible, copyable checkpoints for continuation without hidden memory |
| Cross-platform adapters | Packages smaller prompt versions for custom bots, ordinary chat, and API-style use |
The strongest current coverage is university-level STEM and AI-CS: calculus, linear algebra, probability, discrete mathematics, programming, algorithms, machine learning, systems, networks, physics, signals, and engineering foundations. The Tutor remains useful across other learning domains, but it is not positioned as a generic answer bot.
| Need | Use |
|---|---|
| General tutoring | universal-diagnostic-tutor |
| Learning path / study plan / exam route | tutor-learn-path |
| Practice / grading / mistakes / gap diagnosis | tutor-practice |
| State cards | tutor-state-card |
| Resources | tutor-resource-scan |
| Visual learning | tutor-visualize |
V1.9.2 simplifies the public command surface into six canonical Tutor
entrypoints. Older intents such as /study-plan, /mistake-review, and
/diagnose-gap remain supported as text aliases. These aliases are not
guaranteed native commands in every host. See the
Command Surface for practical examples.
Goal clarification -> Knowledge map -> Teach one concept -> Practice
-> Learner answer -> Qualitative grading -> Mistake repair
-> Visible state update -> Readiness decision -> Next step
The full chain is used only when the learner needs it. A quick factual question does not trigger a giant workflow, and a practice turn normally stops after one targeted exercise to wait for the learner's answer.
V1.9 closes the gap between explanation and demonstrated understanding. The Tutor can generate one targeted exercise, wait for the learner's attempt, preserve correct reasoning, identify the earliest meaningful mistake, select a focused repair, update visible learning state when useful, and decide whether to advance, review, step down, diagnose again, or continue practicing.
Readiness is evidence-based. An explanation or one lucky correct answer does not establish mastery. Knowledge Link Cards are limited to strongly related blocking concepts and usually contain only one to three compact cards.
Instead of immediately solving a vector problem, the Tutor first distinguishes "parallel" from "equal components," identifies scalar multiples as the missing idea, teaches one compact step, and asks the learner to apply it before moving on. See EXAMPLES.md for concise comparisons and teaching flows.
| Document | Purpose |
|---|---|
| User Guide | Beginner-friendly setup and usage tutorial |
| Command Surface | Tutor entrypoints and text shortcuts |
| Install | Installation, updates, and copied-Skill synchronization |
| Portability | Full Skill, custom bot, Lite Prompt, and API prompt choices |
| Examples | Short public examples of diagnosis-first tutoring |
| Evaluations | Behavioral evaluation cases |
| Quality Rubric | Scoring criteria for tutoring quality |
| Failure Taxonomy | Known failure classes and repair targets |
| Changelog | Release history |
The root READMEs are landing pages. Detailed tutorials belong in the linked
documents, and implementation guidance remains in
skills/universal-diagnostic-tutor/.
- No hidden memory, automatic learner profile, database, RAG/vector store, or backend infrastructure.
- No official grading claims, score guarantees, exam prediction, leaked materials, cheating, or 押题.
- No claim that every platform natively supports Skills or slash commands.
- No replacement for professional medical, legal, financial, tax, or safety advice.
- No copied textbooks, answer bank, course platform, or persistent gradebook.
Learning continuity uses visible, user-controlled Learning State, Profile, and Task Cards. Platform adapters are prompt packaging and may be less capable than the Full Skill.
Released under the MIT License.