What is this article talking about?
I turned Naval Ravikant’s books and several podcast interviews into a digital-clone skill package you can chat with directly, more like talking to a friend than listening to a book-club host. This article uses that real case to break down the complete method for turning books, videos, and PDFs into AI consultants: organizing your sources, distilling the thinking, writing the consultant settings, and verifying the fit.
· People who read books, watch interviews, and follow expert content, and want that knowledge to stick and stay usable
· People with industry reports, handouts, or course series on hand who want to turn them into a decision-making advisor
· People who want to turn a teacher or author they admire into "a clone they can consult"
· A four-step method, from organizing sources to writing the consultant settings
· A skeleton for the consultant-settings document you can fill in directly
· A real iteration case: how a fit score of 65 got patched up until it was actually usable
Step one: clean up your sources
A consultant's quality is capped by the quality of its sources. Books can be split into chapters; for videos, get a verbatim transcript first; for PDFs, confirm the text layer is readable, and when you need precise content, convert it to plain text or Markdown, because PDFs hide a lot of invisible layout code that interferes with how the AI reads.
Two details matter especially. First, if a source is incomplete, note it clearly: which passages you have actually read and which are just index information, so the model doesn't fill the gaps with plausible-looking content. Second, multi-speaker material has to keep the speakers strictly separate. When I was organizing one of Naval's group interviews, I spelled it out in the skill package: Naval's own views must be kept distinct from the cases and context of the other guests in the room, or the consultant will attribute other people's words to Naval.
Step two: what you distill is the thinking, one layer beyond a summary
The key to a consultant-type AI is drawing out the way the source makes judgments. As you read, keep these questions in mind: When the author faces a problem, what do they look at first? What categories do they use? How do they make trade-offs? Which risks do they flag?
Take the Naval clone. I split the sources into a topic index: fundamentals of wealth, specific knowledge, leverage and accountability, long-term games and compounding, focus, luck, and patience. Each topic has its own reference file, and when a new interview comes in, it gets folded into the matching topic and cross-linked. That way, when the AI answers a question, it follows his thinking framework instead of re-reading the whole book.
When the volume is large, there's another practical trick: dumping hundreds of thousands of words on the AI at once noticeably degrades output quality. Feed it in batches and layers, giving the topic index first and loading the relevant chapters only when needed. The results are far more stable.
Step 3: Write the consultant role as a reusable setting
Once you've distilled the thinking, write it up as a consultant-settings document and drop it into a skill package, a project description, or your knowledge base, so any AI can activate this consultant later:
Different sources map to different tasks: a book can become a reading advisor, an industry PDF a decision advisor, a video series a course teaching assistant. Write the settings clearly and the same batch of material can produce value again and again.
Step four: verify the fit — a consultant is something you iterate
Don't rush to use the first version. I ran a serious check on the Naval clone: I took an AI-related topic, had the clone answer it, and compared that against what he actually said in his interviews. The fit came out to only 60–65%, and the "how to learn AI" section scored just 3.4.
The reason for the low score was very concrete: on tool-related topics the clone kept circling back to a philosophical framework, but Naval himself gets genuinely excited about these topics and gives concrete, hands-on advice. So I filled three gaps: how he actually uses AI to learn things, the specifics of cross-checking across multiple AIs, and his answer to AI anxiety. After patching those in, I added one rule to his speaking style: on tool topics, keep the excitement and give concrete steps.
This is another difference between a consultant and a summary: a summary is done once it's written, but a consultant has to be verified, iterated, and versioned. Every reinforcement written into the update log brings its answers closer to the real person's thinking. In the end it's a clone you've assembled, and it reflects how deeply you understand this batch of material.
How to get started: pick one small source first
Start with a single PDF or one video; don't take on an entire body of work right away:
- Organize the source: get clean text and mark which parts you've read and which are only index information.
- Distill the judgment rules: what the author looks at first, how they categorize, how they make trade-offs, which risks they flag — write it into a one-page note.
- Write a consultant setting from the skeleton above, hand it to the AI, and try asking three questions you've actually run into.
- Verify: compare the consultant's answers against the source, score the fit, and fill whatever is missing.
Once a small source works end to end, scale up to books and course series. If you want to see the finished product, I've put the Naval clone skill package on the official skill-package download page (both Chinese and English), and the English version is also open-sourced on GitHub, so you can reference the structure directly.