AI Learning Method

简体中文

An evidence-driven Agent Skill for learning any subject deeply enough to retrieve, distinguish, transfer, and use it later.

AI tutors often optimize for a smooth conversation: long explanations, quick praise, and a feeling of progress. This Skill optimizes for durable capability. It gives an AI tutor a macro curriculum loop, a micro mastery loop, explicit stage transitions, repair rules, delayed review, and auditable learning evidence.

The core idea

flowchart LR
    T[Target capability] --> M[Domain map]
    M --> U[Minimal unit]
    U --> R[Closed-book retrieval]
    R --> D[Diagnose and repair]
    D --> X[Discriminate]
    X --> G[Guarded transfer]
    G --> P[Defense or performance]
    P --> V[Delayed review]
    V -->|evidence updates| M

Loading

A lesson is not complete because the explanation ended. A learner advances when the required evidence exists.

What it prevents

• dumping an entire chapter before checking prerequisites;
• mistaking recognition or fluent notes for recall;
• endless Socratic questions without an explicit correction;
• praising a materially wrong answer as “basically correct”;
• skipping ahead after a failed transfer or faulty analogy;
• marking mastery from one immediate answer;
• losing the exact learning stage during a meta conversation or new session.

Evidence gates

Gate	What the learner must show
Retrieval	Reconstruct the idea without seeing the answer
Discrimination	Distinguish it from nearby concepts and traps
Transfer	Apply the underlying principle in a changed situation
Defense / Performance	Explain, derive, solve, create, or perform under challenge

Evidence is recorded separately. Passing one gate does not imply passing the others.

Quick start

Clone the repository or copy it into the skills directory supported by your AI tool. Then ask the agent to read SKILL.md:

Use ai-learning-method to help me learn probability. My goal is to reason correctly about uncertainty in real decisions, not just finish a textbook.

For a serious multi-session course:

1. Copy the files in assets/course-bootstrap/ into a private course folder.
2. Set the final capability, current level, time horizon, source standard, and desired depth.
3. Let the tutor build a compact domain map before drilling details.
4. Preserve state and evidence at the end of each meaningful session.

For a short question, use the lightweight Explore path. The Skill should not force a full course workflow onto every casual interaction.

Two nested loops

Macro: systematic learning

Target capability → domain map → dependencies → modules → integration → capstone or real use → feedback → map update.

Micro: mastery of one unit

Orient/teach → closed-book retrieval → diagnose → one minimal hint and one retry → explicit correction when needed → re-retrieval → discrimination → guarded transfer → defense/performance → delayed review.

Works across subjects

The method is domain-neutral. Evidence changes by subject:

• mathematics: derivation, error localization, novel problems;
• history: chronology, causality, sources, interpretations;
• primary texts: claim, context, textual evidence, competing readings;
• languages: production, comprehension, correction, delayed reuse;
• science: model, prediction, experiment, limitations;
• coding: implementation, debugging, explanation, changed constraints;
• business and professional learning: judgment, scenarios, artifacts, defense.

See references/domain-adapters.md.

Execution controls

• skill calibration / skill 校准: pause teaching and audit the current stage, evidence, and any illegal transition.
• Meta conversation: freeze the course state, discuss the process, then resume from the exact frozen point.
• Source fidelity: label source claims, interpretations, uncertainty, and dynamic facts.
• Error repair: incorrect analogies and failed transfers must be repaired before advancing.

Package layout

ai-learning-method/
├── SKILL.md                  # canonical Skill entry
├── SKILL.zh-CN.md            # complete Chinese version
├── references/               # engines, state machine, adapters, rationale
├── assets/course-bootstrap/  # bilingual course-state templates
├── evals/                    # regression prompts and baseline audit
├── docs/                     # integration and example session
└── manifest.json             # package metadata

Privacy

The repository contains no learner data. Keep course state, private sources, conversations, and credentials in the host project's private workspace. The Skill never treats publication or an external tool call as implicitly authorized.

Evaluation

evals/evals.json captures behavior regressions discovered in real learning workflows, including premature mastery, false praise, stage drift, faulty analogies, and broken freeze/resume continuity.

Run the package validator:

python tools/validate_package.py

This command checks package files, versions, bilingual pairs, and evaluation data. It does not run the prompts against a model or establish learning outcomes; those require separate recorded evaluation runs.

Contributing

Contributions are welcome, especially reproducible teaching failures, better domain adapters, and clearer evidence tests. Read CONTRIBUTING.md and use synthetic learning records only.

License

MIT © 2026 davidme6. Reuse, modification, distribution, and commercial use are allowed while the copyright and license notice are preserved.

Support the project

If the Skill helps you learn or teach more effectively, you can support its maintenance below. Support is voluntary and does not affect access or licensing.

PayPal	Alipay / 支付宝
shengyichaogg@gmail.com

Please verify the recipient before paying. See SUPPORT.md for details and other ways to help.

Related project: Personal AI Agent.