AI Consultant Workflow

Convert books, videos, and PDFs into conversational AI consultants

A book can be summarized or turned into a consultant who can be consulted at any time. The difference is that the latter saves not only the content, but also the author's judgment method, problem framework and usage context.

Published 2026-05-03 | Last updated 2026-05-03

What is this article talking about?

I combined Naval Ravikant’s writings and multiple podcast interviews into a digital avatar skill package that can be used to chat directly, conversing like a friend, more than a book club guide. This article uses this real case to dismantle the complete method of turning books, videos, and PDFs into AI consultants: sorting sources, extracting ideas, writing consultant settings, and verifying fit. Who is

suitable for?

· People who often read books, watch interviews, and follow expert content, and hope that the knowledge can be retained and used
· Those who have industry reports, handouts, and series of courses in hand and want to become a decision-making consultant
· People who want to help the teachers or authors they admire become "a doppelganger who can ask for advice"

What can you take with you?

· Four-step method from source collection to consultant setting
· The skeleton of a consultant setting document, fill it in directly
· Real verification iteration case: How to make up for the fit score of 65 until it can be used

The core difference is one sentenceThe abstract answers "What is this book about?" and the consultant answers "According to the ideas of this book, what should I do now?" To achieve the latter, what you extract must be the method of judgment, and there is still one layer missing just from the content and focus.

Step one: sort out the sources

The upper limit of consultant quality is determined by the quality of the source. Books can be split into chapters; videos must be obtained verbatim first; PDFs need to confirm that the text layer is readable, and it is recommended that precise content be converted into plain text or Markdown, because PDFs hide a large amount of invisible typesetting code, which will interfere with AI reading.

Two details are particularly important. First, if the source is incomplete, mark clearly in the notes: which paragraphs have been read and which are just index information, to prevent the model from filling in the blanks with seemingly reasonable content. Second, multi-speaker materials must strictly distinguish between speakers. When I was organizing a multi-person interview with Naval, I clearly stipulated in the skills package: Naval’s own views must be distinguished from the case context of other guests in the same room, otherwise the consultant will regard other people’s words as Naval’s.

The second step: the focus of extraction is thinking, which goes beyond the abstract level

The key to advisory AI is to sort out the judgment behind the material. When reading, keep these questions in mind: What does the author look at first when faced with a problem? What classifications are commonly used? How to choose? What risks will you be reminded of?

Taking Naval's clone as an example, I split the source into a topic index: basic concepts of wealth, exclusive knowledge, leverage and responsibility, long-term game and compound interest, focusing on luck and patience. Each topic has a reference file. When new interviews come in, the corresponding topics are supplemented and linked to each other. In this way, when the AI ​​answers questions, it follows its thinking framework and does not need to re-read the entire book.

There is another practical point of view when the volume is large: if you throw hundreds of thousands of words to AI at a time, the output quality will be significantly reduced. Provide it in batches and layers. Index the topics first and then load the corresponding chapters when needed. The effect is much more stable.

Step 3: Write the consultant role as a reusable setting

After the thinking is extracted, write a consultant setting document and put it into the skill package, project description or knowledge base, so that any AI will know how to activate this consultant in the future:

Consultant setting skeleton: 1. Source of information: Which book, which videos, and which chapters (list attached) 2. Role tasks: Reading consultant/decision-making consultant/course teaching assistant, what types of questions should be answered? 3. Answering style: How I speak, what metaphors I often use, and what I am excited about 4. Quotation rules: When answering, distinguish between "original text key points / organizer's inferences / AI suggestions" 5. Prohibited matters: How to respond to topics that are not covered by the source, and how to mark when unsure

Different sources correspond to different tasks: a book can become a reading consultant, an industry PDF can become a decision-making consultant, and a series of videos can become course teaching assistants. If the settings are clearly written, the same batch of materials can produce value repeatedly.

Step 4: Verify the fit, the consultant is iterative

Don’t rush to use it after finishing the first version. I conducted a serious verification on Naval's clone: ​​I took an AI topic and asked the clone to simulate answering, and then compared it with what he actually said in the interview. The result was that the fit was only 60% to 65%, and the "How to learn AI" section only got 3.4 points.

is very specific: the avatar will circle back to the philosophical framework when it comes to tool topics, but Naval himself is excited about these topics and can give specific operations. So I filled three gaps: the actual usage of learning things with AI, the specific operations of cross-validation across multiple AIs, and his solution to AI anxiety. After completing it, add a rule in the way of speaking: retain excitement on the topic of tools and give specific operations.

This is another difference between a consultant and a summary: the summary is over, and the consultant needs to verify, iterate, and record versions. Every time reinforcement is written into the update record, the consultant's answer will become closer and closer to my own thinking. After all, it is a clone that has been sorted out, and represents the depth of your understanding of this batch of materials.

The boundary between quotation and inference must be checked before public release.The most common mistake of consultant AI is to make the content that is not in the source appear to be said in the original text. Each output is clearly divided: the key points of the original text, the organizer's inferences, and the AI ​​suggestions. When publishing to the outside world, keep the source link and uncertainty mark so that readers can trace it back. This is also the most basic respect for the original author.

How you can start: Choose a small material first

Start with a PDF or a video first, don’t challenge the entire set of works:

  1. Organize the sources: Get the clean text and mark which ones have been read and which ones are just indexed.
  2. Extraction judgment rules: What should the author look at first, how to classify, how to choose, and what risks to remind, and write it into a page of notes.
  3. Write a consultant setting according to the skeleton above, throw it to the AI ​​and try to ask three questions you actually encountered.
  4. Verification: Compare the consultant’s answers with the original text, rate the fit, and make up for whatever is missing.

After running through the small materials, expand to books and series of courses. Want to see the complete finished product, I put the Naval clone skill packOfficial website skill package download page(both Chinese and English versions), the English version is also open source at GitHub, the structure can be referenced directly.

AI ConsultantKnowledge ExtractionDigital cloneSkill Pack DesignReading methodReference boundaryKnowledge Management