AI Reading Skill for OpenAI Codex and ChatGPT: Understand Books Faster, Think Deeper, Put Ideas to Work
English | 简体中文
Open-source AI book analysis · Feynman learning · Critical reading · Context-aware application
Reading a book can stretch across days or weeks—not only because of the pages, but because it takes time to find what matters, untangle difficult concepts, question convincing claims, and decide what any of it means in real life.
AI Reading is an open-source AI reading and book-analysis Skill for OpenAI Codex and ChatGPT. It helps you focus on the right questions, explain ideas in your own words, examine claims from multiple angles, and turn useful insights into decisions, actions, or content.
A typical AI helps you get a summary faster. AI Reading helps you reach your own understanding and judgment faster.
It compresses aimless searching and repeated reorganization—not the experience of reading itself.
Open Codex and paste:
Use $skill-installer to install the AI Reading Skill from:
https://github.com/yanz86808-beep/ai-reading-skill/tree/main/skills/ai-reading
After installation, start a new conversation:
Use $ai-reading to read Thinking, Fast and Slow. Explain its full structure, core ideas, disputes, and real-world applications in one report.
No API key or additional runtime dependency is required. The host may need web-search permission when current facts or external evidence must be verified.
Instead of treating every chapter and idea as equally important, AI Reading starts from your goal and identifies the concepts, claims, and questions most worth pursuing.
When you simply ask it to read a book, AI Reading first treats the book-wide master framework as the first core knowledge point: what problem the book solves, which modules make up its answer, and how those modules interact and feed back into one another. It then unfolds three to five core modules attached to that framework and uses evidence appropriate to the book type: complete cases for practical books, key scenes and textual evidence for literature, argument chains for philosophy, event and source chains for history, and study design plus current consensus for science. The report ends with matching understanding aids, a pressure test, and an appropriate next step instead of forcing every book into a business action plan. Evidence strength, outdated claims, and unverified details are consolidated in a closing audit.
It asks you to explain difficult ideas in plain language, corrects gaps, and examines claims from multiple independent angles. The goal is not to attack the author, but to understand why an idea works, when it works, and where it may fail.
The same idea can mean very different things to a manager, teacher, recruiter, creator, or individual reader. AI Reading uses only the role, goal, authority, and constraints you explicitly provide to determine what the idea means for you, what can be applied, what needs adapting, and what should not be copied blindly.
Find the question → Explain it clearly → Examine it from multiple angles → Put it to work
| Stage | Reader-facing purpose | What happens |
|---|---|---|
| 1. Find what matters | Decide where your attention is worth spending | Clarify the book, reading goal, available material, reader context, and the questions worth following. |
| 2. Explain it clearly | Turn “I think I understand” into an explanation you own | Use the Feynman method: explain one concept at a time so a six-year-old could follow, then correct and restate it. |
| 3. Examine it from multiple angles | Avoid accepting a single convincing viewpoint too easily | Test why the claim holds, what evidence supports it, which conditions it depends on, and where its boundaries are. |
| 4. Put it to work | Turn understanding into a usable result | Create one relevant checklist, knowledge card, training outline, decision aid, or channel-ready draft. |
The stages work as one continuous journey, but each can also run independently. Ask for stage 2, 3, or 4 directly; AI Reading will establish only the minimum reliable foundation needed instead of forcing you to restart from stage 1.
If you simply say “read this book,” the default is one complete report. If you say “guide me step by step” or “start with stage 1,” it switches to an interactive journey and waits for your own answer during Feynman learning.
If you explicitly ask to test or automatically run all four stages, AI Reading enters evaluation mode. It uses clearly labeled fixture answers to test correction and state continuity without presenting them as your answers or claiming that you learned the concepts.
AI Reading shortens the path from “I encountered this book” to “I can explain, evaluate, and use its ideas” by reducing four common forms of wasted effort:
- Reading without a question: identify the issues most relevant to your goal before spending equal time everywhere.
- Mistaking familiarity for understanding: expose gaps through explanation and correction instead of repeated passive rereading.
- Accepting claims without boundaries: compare multiple independent perspectives in one structured pass.
- Starting over after finishing: connect insights to your context and produce a usable outcome during the same workflow.
For practical nonfiction, management, psychology, business, and popular science, this can make the route to usable understanding much shorter. For literature, the Skill preserves ambiguity and reading experience instead of treating speed as the only goal.
AI Reading does not generate the same application advice for everyone. It can use information you explicitly provide, such as:
- your role and the decision you are facing;
- the outcome you want from the reading;
- what you can and cannot influence;
- organizational, ethical, legal, or practical constraints;
- the format you ultimately need.
For example, the claim “default options influence choices” raises different questions for a product manager designing an interface, a recruiter collecting candidate data, and an individual changing a habit. AI Reading changes the questions and application boundaries—not the book's original claim—to match the reader's actual situation.
Use $ai-reading to read Thinking, Fast and Slow. Explain the whole book first, then audit which claims remain reliable.
Use $ai-reading to help me decide whether this book is worth reading.
Run stage 1 only and tell me which three questions deserve the most attention.
Help me understand sunk cost with the Feynman method.
Start directly at stage 2 and ask me one question at a time.
Examine the claim that a company should enroll employees in AI behavior monitoring by default.
Use multiple independent angles and tell me where the claim stops being reasonable.
Turn what I have learned into a 45-minute training outline for my team.
Use only ideas that have already been clarified and examined.
Automatically test $ai-reading across practical, literary, scientific, and philosophical books. Run all four stages with labeled fixtures without waiting for my answers.
See examples/prompts.md for more prompts, evals/ai-reading/test-cases.md for public acceptance scenarios, and the report quality rubric for complete-report evaluation.
- Nonfiction, management, psychology, business, and practical-method books
- History, biography, philosophy, and popular science
- Literature, fiction, and multiple textual interpretations
- Close reading of lawfully provided excerpts
- Feynman learning, critical reading, decision support, reading notes, and content assets
AI Reading does not require an uploaded copy of the entire book. A title, author, ISBN, cover, concept, table of contents, lawful excerpt, or reading question can be enough to begin, depending on the requested stage.
Input options and evidence conditions
| Input | Required | Details |
|---|---|---|
| Book or concept | Yes | Provide a title, author, ISBN, cover, or concept. If the identity is ambiguous, AI Reading first asks for the author or edition. |
| Reading goal | Recommended | Understand a concept, solve a problem, examine an argument, experience a literary work, or create an asset. |
| Stage | No | If omitted, AI Reading produces a complete reading report. You can also request stage 1–4 directly. “Next stage” continues the current reading state. |
| Source material | No | Provide a lawful excerpt, table of contents, chapter, or key pages when available. |
| Reader context | No | Provide a role, problem, action scope, or constraint. Sensitive traits are not inferred. |
| Output controls | No | Choose concise, standard, or deep output; spoiler preference; target channel; or asset type. |
Evidence coverage is always explicit:
- When task-relevant text is available, it is treated as primary evidence.
- When only public sources are available, authors, publishers, papers, and authoritative institutions are prioritized.
- When only partial excerpts are available, conclusions remain limited to those passages.
- When a new or obscure book cannot be verified, AI Reading asks for a table of contents or lawful excerpt instead of inventing content.
The absence of pasted book text is not itself a blocker. The workflow pauses only when the book or goal cannot be identified, or no reliable material can support the requested stage.
Output modes and stage completion rules
- Complete report mode (default for broad requests): provide a book-wide master framework as the first core knowledge point, three to five attached core modules, type-appropriate supporting evidence, key tensions or misconceptions, understanding aids, a Feynman self-test, pressure test, credibility audit, and an appropriate next step.
- Guided mode: proceed stage by stage when the user explicitly asks for guidance, and wait for the reader's answer during Feynman learning.
- Single-stage mode: run the requested stage directly and establish only its minimum reliable foundation.
- Focused report mode: answer a specified question with research, self-test material, multi-angle examination, and asset recommendations while marking interactions that have not occurred.
- Evaluation mode: when explicitly requested, run all four stages with labeled fixture answers and verify book switching, skipped or revisited stages, and claim-version continuity. Fixtures never count as reader mastery.
- Length levels: concise usually uses up to three core frameworks; standard uses three to five; deep adds disputes, counterexamples, and source comparison.
For stage 3, simple methods usually use two perspectives, causal or generalized claims use three, and high-risk claims affecting individuals, organizations, institutions, and ethics use four. The perspectives must be meaningfully different rather than renamed versions of the same objection.
Skipped stages are never presented as completed. Every output states its actual evidence coverage and distinguishes verifiable book content, author commentary, third-party interpretation, and analysis.
Reliability is a foundation of the workflow, not its headline:
- No invented chapters, cases, experiments, data, page numbers, or quotations
- No chapter-by-chapter reconstruction or extensive text that substitutes for the original book
- No claim of full-book coverage when only partial or public material was available
- No inference of sensitive personal traits from reader context
- No substitution for professional medical, legal, or financial advice
See source-and-copyright-policy.md for the complete policy.
Clone or download the repository, then copy only skills/ai-reading into the user-level skill directory:
git clone https://github.com/yanz86808-beep/ai-reading-skill.git
mkdir -p ~/.agents/skills
cp -R ai-reading-skill/skills/ai-reading ~/.agents/skills/ai-readingSome environments use $CODEX_HOME/skills, commonly ~/.codex/skills; $skill-installer selects the configured location automatically. For repository-level discovery, copy the folder to .agents/skills/ai-reading inside the target repository.
skills/ai-reading/ Installable Skill
evals/ai-reading/ Forward-test cases
examples/ Curated prompts
scripts/ Deterministic release validation
Run the dependency-free validator:
ruby scripts/validate_skill.rbGitHub Actions runs the same validation on pushes and pull requests. Local runs/ and reviews/ contain development artifacts and are excluded from the public release.
Contributions are welcome. See CONTRIBUTING.md, report security or privacy issues through SECURITY.md, and review the stable v0.1.0 release.
Released under the MIT License.