Turn Any Book Into an AI Skill You’ll Actually Use  🔗 **Source:** [Turn Any Book Into an AI Skill You’ll Actually Use](https://www.youtube.com/watch?v=8PM_t3NWrAE) > [!note]+ Document Notes > tags: [skill-building, ai-agents, zettelkasten, deliberate-practice, knowledge-management, behavior-change] > author: Vicky Zhao > source: YouTube > reference: https://www.youtube.com/watch?v=8PM_t3NWrAE > date created: 2026-08-31 11:00 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Turn Any Book Into an AI Skill You'll Actually Use > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [The video under analysis](https://www.youtube.com/watch?v=8PM_t3NWrAE) — Vicky Zhao, 28:22, published 2026-08-27 - [*The 1-Page Marketing Plan*, Allan Dib](https://successwise.com/) — the worked example throughout - [Shu-Ha-Ri](https://en.wikipedia.org/wiki/Shuhari) — the Japanese mastery progression she builds the argument on - [Codex](https://openai.com/codex/) — the agent she builds the skill in ## 📝 Summary - **What happened.** Vicky Zhao argues that the standard way people turn books into AI skills — upload a PDF, or feed the model a summary, or just name the famous book — produces vague, useless skills, because each step is an abstraction of an abstraction. She proposes a different construction: a skill must encode **when** to use a method, **how** to make the change, and **record what happened**, forming a feedback loop. She grounds the discipline in **Shu-Ha-Ri**, admits her own failure to follow it, then builds the skill live in Codex against Allan Dib's *The 1-Page Marketing Plan*, tests it on her own course sales page, and finally upgrades it by pointing the agent at her Obsidian Zettelkasten to find her personal knowledge gaps. - **Why it matters.** This is the best articulation I have seen of why most "turn X into a prompt/skill" content produces disappointing results, and the diagnosis — compression destroys the specificity that makes a method actionable — is correct and generalizes well beyond books. The second half is more valuable still: pointing an agent at your own accumulated notes so the skill is personalized to *your* gaps rather than to the book's generic structure. - **Key stakeholders.** Anyone building agent skills. People with an existing note-taking system who have not connected it to their agents. Allan Dib, whose book is the running example and who she credits generously. Christo, credited for the "just follow the advice" observation. Codex, Claude Code, and ChatGPT Work, which she repeatedly urges viewers to adopt over chat interfaces. - **Risks and tradeoffs.** The advice to give agents rich context rather than compressed summaries is right in direction and unbounded in the video — context windows are finite and attention degrades over long contexts, neither of which is mentioned. Pointing an agent at an entire personal Zettelkasten is a meaningful privacy decision presented as a pure technique. And one moment in the live demo has her overriding a true-or-false agent finding without investigating, inside a video whose whole thesis is evidence-based feedback loops. - **Bottom line.** The strongest video in this batch and the one with the most durable ideas. The abstraction-layers diagnosis, the three-part skill anatomy, and the Zettelkasten-as-context move are all immediately usable. Take the framework; supply your own limits on context size and your own discipline about not talking the agent out of findings. ## 📖 Main Content **The problem she is actually solving.** She opens with two problems, and is careful to set the first aside. Problem one is remembering what you read, which she says note-taking in Obsidian solved for her. Problem two is bigger: *"Just because you can remember what's inside the book doesn't mean you use it."* Her framing of learning is behavioral — *"learning doesn't happen until we change our behavior"* — and everything downstream follows from that premise. Reading without behavior change she calls *"intellectual fantasizing about a better world,"* which is a good line and, as an epistemological claim, a contestable one I return to below. **The abstraction problem, which is the video's best diagnosis.** The standard advice is to hand an AI a book summary and ask for a skill. She traces what that actually does. A person has experience and expertise; they abstract it into a book — that is abstraction one, and a helpful one, since it makes the knowledge transmissible. A summary compresses 200 pages into a page of key frameworks — abstraction two, and here *"you lose some of the nuance, some of the examples. We don't just learn from abstract concepts. We need to see how they play out specifically."* Then you ask a model to distill that summary into a skill — abstraction three, arguably four. The output is predictably generic: *"it's not surprising that, at the end, we get something very average... something that doesn't have nuance, because we keep on stripping it of nuance so that it's easier for us to digest."* The turn is the sharp part: compression exists for *human* cognitive limits, and applying it before handing material to a machine is a category error. *"specifically for AI, we don't want to do this because AI is really good at digesting information. So we don't want to give it an abstraction of an abstraction."* That reframing — summaries are a human affordance, not a machine one — is the single most transferable idea in the video, and it explains a large class of disappointing AI outputs that people usually blame on the model. **The three-part anatomy of a working skill.** She specifies what a skill must encode. **When** to use the method: not "whenever," but the identified situation and the specific behavior you are trying to change — *"Why did you reach for that book in the first place?"* **How** to make the change, which is the method itself. And **record and review**: *"did it actually contribute to a change? If it changed, what changed? Did it change for the better, or did it change for the worse?"* The three combine into situation → method → new action → reviewed result, which is a feedback loop. Her argument for why this is newly practical is that AI removes the cost that historically made personal feedback loops collapse — *"it is good at tracking things, it is good at remembering things, it is good at surfacing things for you."* That is true and underappreciated. The friction of maintaining a personal evidence log is what kills most deliberate-practice systems, and an agent that logs and surfaces patterns removes most of it. **Shu-Ha-Ri, and an unusually honest failure story.** She imports the three-stage mastery progression from Japanese martial arts and tea ceremony. **Shu** (obey): follow the teacher's framework exactly. **Ha** (break): explore the framework's edges, bring in other teachers, customize. **Ri** (move away): having built genuine experience, develop your own approach. Her claim is that people skip straight past Shu, and she indicts herself specifically. Starting her business, she used *The 1-Page Marketing Plan* as a marketing bible. The book has nine grid cells with instructions for filling each. *"But me, you know, not having experience, I was like, okay, great, there are these nine things. But then, you know, secretly, I'm not comfortable doing some of them. So I was making up excuses for why I'm not going to do them."* She then tried to copy Apple's approach instead, and *"I was just such a mess... in the end I was getting nowhere."* This is the video's most credible passage, and the mechanism she identifies is precise: skipping Shu is usually not intellectual disagreement, it is discomfort wearing the costume of intellectual disagreement. The excuse arrives already dressed as strategy. She then supplies the right analogy — *"this is how scientific experiments work. You're not going to change all the variables at once"* — which makes the case for Shu as experimental control rather than as deference. That framing rescues the concept from sounding like obedience for its own sake. **AI as emotionless support.** The reason she wants a machine in this loop is that *"learning is actually a very emotional experience because we're getting into uncertainty... We are brushing up against our ego."* An agent has no ego to protect and no discomfort to rationalize around, so it can hold you to the method *"without us relying on staying disciplined and staying rational all the time."* This is a genuinely good argument, and it is the missing justification most AI-productivity content skips: the machine's value here is not capability, it is that it does not flinch. **Two build paths, demonstrated live.** The no-notes version is a plain conversational prompt to an agent naming the book, the situation ("whenever I am doing any marketing activities inside my business"), the method (findable online for a well-known book), and the two things to track — did I follow Dib's method, and what was the result. She is explicit that the prompt is casual: *"this is not a complicated prompt. I'm just making it up on the spot."* That leads to a claim worth isolating: *"If you use agents as opposed to the chat version... ChatGPT needs better prompting. But if you use agents, the prompt is less important. Context is everything with agents."* This is a real and useful distinction — an agent that can read your files, run tools, and iterate compensates for underspecified instructions in a way a single chat turn cannot. It also functions as the video's recurring product pitch, which she repeats at least four separate times. Codex produces a skill with phases matching the book's structure, a decision-focused question per phase, and — because she asked it to track results — an explicit list of minimum evidence per grid cell (core message, offer mechanism, proof, objection, CTA, claims boundary). She then tests it by pasting a URL to her own systems-thinking course sales page with no further instruction, and the agent correctly infers this is a test. **The findings are real, which is what makes the demo persuasive.** Source attribution and UTM persistence unverified — *"Yes, actually, this is missing, so I need to set it up."* No path for interested non-buyers — *"True. I haven't really thought about that."* No referral mechanism. A demo where the tool finds genuine gaps in the presenter's own live business is far better evidence than a staged example, and she does not edit out the parts where it makes her look unprepared. **One moment worth flagging.** The agent reports that cell six is broken and checkout renders zero. She responds: *"I've checked this before—but it doesn't seem to be a problem. I'm not sure what happened. Maybe it's a browsing issue."* She asks it to re-check, then tells it *"Checkout is working fine. So update your analysis."* She is very likely right about her own checkout. But inside a video arguing for evidence-driven feedback loops, this is the pattern that breaks them: an agent surfaces a finding, the human's prior says no, and the finding is overwritten rather than investigated. She had a hypothesis — a browsing issue — and never tested it. The correct move was one step further: confirm *why* the agent saw a zero. It is a small moment and I would not raise it except that the video's entire thesis is about not letting your priors dictate what counts as evidence. **The Zettelkasten upgrade, which is the most actionable part.** The bonus section is where the video earns its length. Her insight is that the *method* is identical for everyone who reads a book, but the *gaps* are not: *"maybe you have experience in branding, maybe someone else has experience in accounting... even though the method is the same, where our knowledge gaps are and why we find this method helpful will be different."* So she directs Codex at her Obsidian Zettelkasten folder to find her own notes on Dib, analyze what she repeatedly highlights and annotates, and infer her knowledge gaps. It surfaces her recurring preoccupations — focus, leverage, direct response, nurture, the nine boxes — and she confirms focus and leverage are exactly what she wants to work on and were not visible in the generic skill. The division of labor she states explicitly is the part to keep: *"I'm getting the agent to do the work for me because I've already done my part of the job, which is use my judgment, pick out what's important, document it, take notes, process the information."* Human judgment produced the notes; the agent does retrieval and synthesis over them. That is a defensible allocation and an argument for note-taking that survives the arrival of capable models — your notes are not a worse search index, they are a record of what *you* found important, which no general model can reconstruct. The regenerated skill asks five diagnostic decisions rather than mechanically walking the nine grids — which constraint binds across leads, sales, delivery, retention; which single cell is the bottleneck; what customer transition should happen; which audience, offer, channel, CTA to focus on; and *"What are we deliberately not solving yet?"* She ties the focus emphasis to a "64/4" heuristic (Pareto applied twice: 4% of causes producing 64% of impact). **Contested and uncertain.** *"Learning doesn't happen until we change our behavior"* is stated as settled and is not — it is a behaviorist position that excludes conceptual and declarative learning by definition. It works well as an operating principle for skill acquisition and poorly as a general theory of learning. *"AI is really good at digesting information"* is used to justify feeding it more raw material, with no mention of context limits or attention degradation over long inputs — the advice has a ceiling the video never locates. And pointing an agent at an entire personal knowledge base is a real privacy and data-handling decision that is presented purely as technique. ### 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=0) Two problems, and which one this video is about - Problem 1: remembering what you read — she considers this solved by note-taking in Obsidian and sets it aside. - Problem 2, the bigger one: remembering ≠ using. *"learning doesn't happen until we change our behavior."* - Reading without behavior change = *"intellectual fantasizing about a better world."* Memorable, and a contestable claim stated as settled. #### ⏱️ [00:58](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=58) The abstraction problem - Extracting frameworks is *"the preparatory step. It's not even the first step."* Application is the work. - Names the standard bad advice: upload the PDF, feed a summary, or just name a famous book the model already knows. #### ⏱️ [02:03](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=123) Four levels of abstraction - Experience → book (helpful abstraction) → summary (loses nuance and examples) → model-generated skill from that summary → generic output. - Key reframe: **summaries are a human affordance, not a machine one.** Compression exists for our cognitive limits; applying it before handing material to a model is a category error. - The single most transferable idea in the video. #### ⏱️ [03:23](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=203) What a skill must encode — three parts - **When**: the specific situation and behavior, not "whenever." *"Why did you reach for that book in the first place?"* - **How**: the method itself, from the source. - **Record and review**: did it change anything, for better or worse. #### ⏱️ [04:33](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=273) The loop - situation → method → new action → reviewed result. - Source can be a book, podcast, or video — the shape is format-independent. #### ⏱️ [05:55](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=355) Why feedback loops are newly practical - Historically collapse because humans are slow and inconsistent at logging. - Agents remove the friction: *"good at tracking things... remembering things... surfacing things for you."* - The evidence accumulates over time and answers: keep the strategy, improve it, or pivot. #### ⏱️ [07:07](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=427) Shu-Ha-Ri - **Shu** (obey) — follow the sensei's framework exactly. - **Ha** (break) — explore edges, bring in other frameworks, customize. - **Ri** (move away) — build your own approach on accumulated experience. - Claim: almost everyone skips Shu. #### ⏱️ [09:37](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=577) Why skipping Shu fails — credits Christo - Successful people *"just follow advice and just do it"* — not because the expert is right, but to generate their own evidence. - *"I probably need to do it a few times before I actually have credible data."* #### ⏱️ [10:52](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=652) Her own failure — the most credible passage - *The 1-Page Marketing Plan*, nine grid cells, used as a marketing bible when starting her business. - *"secretly, I'm not comfortable doing some of them. So I was making up excuses for why I'm not going to do them."* - Then tried copying Apple. *"I was just such a mess... in the end I was getting nowhere."* - **Mechanism worth keeping:** skipping Shu is discomfort wearing the costume of intellectual disagreement. The excuse arrives pre-dressed as strategy. #### ⏱️ [11:53](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=713) Shu as experimental control - *"this is how scientific experiments work. You're not going to change all the variables at once."* - Reframes obedience as establishing a baseline. Rescues the concept from sounding like deference. #### ⏱️ [13:06](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=786) AI as emotionless support - Learning is emotional — uncertainty, ego, fear of not being able to learn. - The agent has no ego to protect, so it holds you to the method *"without us relying on staying disciplined and staying rational all the time."* - The best-argued justification in the video: the machine's value is that it does not flinch, not that it is capable. #### ⏱️ [14:17](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=857) Build path 1 — no notes - Works in Codex, Claude Code, any agent. - Prompt names: the book, the situation, the method source, and two tracked things (did I follow it / what was the result). - *"this is not a complicated prompt. I'm just making it up on the spot."* #### ⏱️ [15:44](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=944) Agents vs chat — context over prompt - *"ChatGPT needs better prompting. But if you use agents, the prompt is less important. Context is everything with agents."* - Real distinction: file access, tool use, and iteration compensate for underspecification in a way one chat turn cannot. - Also the video's recurring product pitch — repeated at least four times. #### ⏱️ [16:26](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1021) Reviewing what Codex produced - Phases match the book; one decision-focused question per phase. - Because she asked it to track results, it derived minimum evidence per cell: core message, offer mechanism, proof, objection, CTA, claims boundary. - *"we don't want to abstract away the things that we're checking for."* #### ⏱️ [18:11](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1091) Live test — a URL and nothing else - Pastes her systems-thinking course sales page with no instruction; the agent correctly infers it is a test. #### ⏱️ [19:25](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1165) Real findings on her own live business - UTM persistence and source attribution unverified — *"Yes, actually, this is missing."* - No path for interested non-buyers — *"True. I haven't really thought about that."* - No referral mechanism. Delivery cell unverifiable (agent lacks backend access) — correctly noted as a limit. - ⚠️ **The moment to flag:** agent reports checkout renders zero; she says *"Maybe it's a browsing issue,"* never tests that hypothesis, then instructs *"Checkout is working fine. So update your analysis."* Probably correct about her own site — but this is the exact pattern that breaks evidence-driven loops, in a video about evidence-driven loops. #### ⏱️ [20:41](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1241) Closing the loop - Will return post-launch with real results so the agent can track improvement across launches. #### ⏱️ [21:55](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1315) Build path 2 — Zettelkasten as context - The method is universal; the *gaps* are personal. Branding vs accounting vs product experience → different gaps from the same book. - Directs Codex at her Obsidian Zettelkasten folder to find her Dib notes and infer gaps from what she repeatedly highlights. #### ⏱️ [23:18](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1398) What it found - Recurring preoccupations: focus, leverage, direct response, nurture, the nine boxes. She confirms focus and leverage are the real targets and were invisible in the generic skill. - **Division of labor, stated explicitly:** *"I've already done my part of the job, which is use my judgment, pick out what's important, document it, take notes, process the information."* Human judgment makes the notes; the agent does retrieval and synthesis. - The strongest available argument that note-taking survives capable models: notes record *what you found important*, which no general model can reconstruct. #### ⏱️ [24:29](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1469) The personalized skill — five decisions - Biggest gap identified: *"the bridge from learning to application"* — which is the skill's own reason to exist. - Five diagnostic questions: which constraint binds (leads/sales/delivery/retention); which single cell is the bottleneck; what customer transition should happen; which audience/offer/channel/CTA to focus on; *"What are we deliberately not solving yet?"* - Shift from mechanical application to diagnosis. Evaluates each cell at three levels: understanding, decision, measurement evidence. #### ⏱️ [25:44](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1544) 64/4 - Pareto applied twice: *"What 4% of the causes is going to give me 64% of the impact?"* #### ⏱️ [27:11](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1631) Close - Over time, patterns emerge in which parts of a method work for you — that is the on-ramp to Ha and Ri. - Free resource lead magnet; a final push to move from chat to agents. ## 📚 References - [Vicky Zhao, "Turn Any Book Into an AI Skill You'll Actually Use," 2026-08-27](https://www.youtube.com/watch?v=8PM_t3NWrAE) — the source under analysis - [Allan Dib, *The 1-Page Marketing Plan*](https://successwise.com/) — the nine-grid framework used as the running example - [Shu-Ha-Ri](https://en.wikipedia.org/wiki/Shuhari) — the mastery progression underpinning the argument - [Codex](https://openai.com/codex/) — the agent used in the live build - [Claude Code](https://claude.com/claude-code) — named as an equivalent target for the same technique - Christo — fellow YouTuber, credited for the "just follow the advice" framing; no link given in the video ## 🗣️ Paraphrase - **Opening.** People who still read books know the problem: we reach for a book because we want to improve something, we highlight and annotate, and it feels productive. But afterwards, first, it is hard to remember what we read — note-taking in Obsidian addressed that. The second and larger problem is that remembering is not using, and learning has not occurred until behavior changes. Otherwise it was only intellectual fantasizing about a better world. - **The plan.** She will show how she turns books into skills usable inside her AI, which then holds her to practicing them. Extracting frameworks is only preparation, not even the first step; the work is figuring out how to apply them. - **Why the obvious approach fails.** People assume you can upload a PDF, rely on the model already knowing a famous book, or feed it a summary and request a skill. Technically possible, but the result will not be useful. - **Abstraction layers.** Expertise from experience is abstracted into a book — a useful first abstraction. A summary compresses hundreds of pages into a page, losing nuance and the specific examples we need to see a concept play out. Asking a model to build a skill from that summary is an abstraction of an abstraction of an abstraction, roughly four levels deep, so the generic result is unsurprising. Since AI digests information well, feeding it pre-compressed material is the wrong move. - **Three requirements.** A successful skill must establish when to use the method — tied to the specific behavior you wanted to change and why you reached for the book — then how the change is made, then how results get recorded and reviewed, since only that reveals whether anything actually improved. - **The loop.** Situation, method from a book or podcast or video, resulting new action, reviewed outcome. AI makes such loops practical because it tracks, remembers, and surfaces, supplying evidence over time about whether to keep, improve, or pivot a strategy. - **Shu-Ha-Ri.** From Japanese martial arts and tea ceremony. Shu means obey — follow the teacher's method exactly. Ha means break — explore the framework's edges, incorporate other teachers, customize. Ri means move away — develop your own style from accumulated experience. Most people skip Shu. - **Why that matters.** Per a conversation with fellow YouTuber Christo, those who succeed simply follow the advice, not because the expert is certainly right, but because doing the prescribed steps generates the evidence needed to decide what comes next, usually over several attempts. - **Her mistake.** She used *The 1-Page Marketing Plan* as a marketing bible when starting out but, lacking experience, conflated all three stages: privately uncomfortable with some of the nine steps, she invented reasons to skip them and tried copying Apple instead. The result was a mess and no progress, because customization offers infinite options once you have no baseline. - **The correction.** Start at Shu: pick the book, do the work, follow the method even when you think you know better, and repeat it enough to see results — as in scientific experiments, where you do not change every variable at once. Only with feedback-loop evidence do you move to breaking and departing. - **Why AI belongs here.** It provides emotionless support. Learning is emotional — uncertainty, ego, fear of never learning the skill — and a system without emotion can guide the process without depending on the learner remaining disciplined and rational. - **Building it, path one.** Tell the agent — Codex, Claude Code, any of them — the book, the situation in which it applies, where the method can be found, and the two things to track: whether the method was followed, and what resulted. The prompt need not be elaborate, because with agents context matters more than prompting, unlike chat interfaces. - **What Codex produced.** A skill matching the book's phases, a decision-focused question per phase, and per-cell minimum evidence — core message, offer mechanism, proof, objection, CTA, claims boundary — because tracking was requested. - **Testing it.** She supplies only a URL to her course sales page; the agent infers a test. It reports that much of the method is followed, flags unverified source attribution and UTM persistence, notes no path exists for interested non-buyers, and finds no referral mechanism. It cannot assess delivery without backend access. A checkout reading is disputed and she has it re-check. - **Closing the loop.** After the current launch she will report results so improvement can be tracked across launches. - **Path two — notes.** The method is identical for everyone, but knowledge gaps differ by background. She has Codex read her Obsidian Zettelkasten for her notes on the book and infer gaps from what she repeatedly highlights. It surfaces focus, leverage, direct response, nurture and the nine boxes; she confirms focus and leverage as her real targets. She frames this as the agent doing retrieval and design while she has already done the judgment work of selecting, documenting and processing. - **The result.** The upgraded skill names her biggest gap as the bridge from learning to application, and asks five focusing decisions covering the binding constraint, the bottleneck cell, the customer transition, the audience/offer/channel/CTA, and what is deliberately not being solved — evaluating each cell for understanding, decision, and measurement evidence. - **Close.** With a repeatable process, patterns emerge over time about which parts of a method work, which is the path through Shu-Ha-Ri toward applying book knowledge as your own. A free resource on agents is offered, with a final encouragement to move beyond chat interfaces. ## 📰 Newsletter **Why your "turn this book into a prompt" skill came out so bland.** Vicky Zhao has the best diagnosis I have read for a failure mode most of us have hit and blamed on the model. Here it is. An expert has experience. They compress it into a book — **abstraction one**, and a good one, because now it is transmissible. You grab a summary, so 200 pages become a one-pager of frameworks — **abstraction two**, and now the examples are gone, which matters because *"we don't just learn from abstract concepts. We need to see how they play out specifically."* Then you hand that summary to a model and ask for a skill — **abstraction three**, maybe four. And you are surprised the output is generic. The reframe that makes this click: **summaries exist for human cognitive limits, not machine ones.** We compress because we cannot hold 200 pages. A model can. Pre-compressing before you hand it over is a category error, and it explains a whole class of disappointing outputs. **What a skill actually needs.** Three parts, and most skills have only the middle one. *When* to use it — tied to the specific behavior you wanted to change, not "whenever." *How* to make the change — the method. And *record and review* — because otherwise you have no idea if it worked. Together those make a feedback loop, and her argument for why this is newly practical is right: personal feedback loops historically collapse because logging is tedious, and an agent that tracks and surfaces removes exactly that friction. **Then she brings in Shu-Ha-Ri**, and does something I did not expect — indicts herself with it. Shu is obey: follow the framework exactly. Ha is break: explore edges, mix in other teachers. Ri is move away: build your own. Everyone skips Shu. Her story: starting her business, she used Allan Dib's *The 1-Page Marketing Plan*, nine grid cells, clear instructions. And — *"secretly, I'm not comfortable doing some of them. So I was making up excuses for why I'm not going to do them."* Then she tried copying Apple's marketing instead. *"I was just such a mess."* That mechanism is worth carrying around: **skipping Shu is usually discomfort wearing the costume of intellectual disagreement.** The excuse arrives already dressed as strategy. And her fix reframes obedience as experimental control rather than deference — *"you're not going to change all the variables at once."* Do the method as written enough times to get a baseline, *then* customize. **The best argument in the video is about why you want a machine in this loop at all,** and it is not capability. Learning is emotional — uncertainty, ego, the fear you may never get it — and an agent has no ego to protect and no discomfort to rationalize around. It holds you to the method *"without us relying on staying disciplined and staying rational all the time."* The value is that it does not flinch. **Now the part to actually steal.** The method in a book is identical for every reader. Your *gaps* are not — someone with branding experience and someone with accounting experience need different things from the same nine grids. So she points Codex at her Obsidian Zettelkasten, tells it to find her notes on the book, and infer her knowledge gaps from what she keeps highlighting. It surfaces focus and leverage. She confirms those are exactly right and were invisible in the generic skill. Her framing of the division of labor is the strongest argument I have seen that note-taking survives capable models: *"I've already done my part of the job, which is use my judgment, pick out what's important, document it, take notes, process the information."* Your notes are not a worse search index. They are a record of what **you** found important, and no general-purpose model can reconstruct that. **Two caveats she does not give you.** "Give the AI more context, not less" is right in direction and unbounded in the video — context windows are finite and attention degrades over long inputs. And pointing an agent at your entire personal knowledge base is a genuine data decision presented purely as technique. **One moment I keep thinking about.** Mid-demo, the agent reports her checkout renders zero. She says *"maybe it's a browsing issue,"* never tests that hypothesis, and tells it *"Checkout is working fine. So update your analysis."* She is probably right about her own site. But in a video whose entire thesis is evidence-driven feedback loops, that is the pattern that breaks them — a finding surfaced, a prior asserted, the finding overwritten without investigation. Worth watching for in your own sessions, because I do it too. ## ✅ Action Items #### DevRel - Developer advocacy: rewrite our "prompt library" guidance around the abstraction-layers argument. Most published prompt advice hands models pre-compressed material. `#action/h1` - Education: workshop on skill anatomy — when/how/record — with a live build. This framework teaches better than any prompt-engineering checklist. `#action/h1` - Developer experience: connect our documentation corpus to agents as raw context rather than as summaries. Test whether output quality rises. `#action/h1` - Developer marketing: not applicable to this source. `#action/h3` - Community management: Shu-Ha-Ri is the right frame for contributor onboarding — follow the contributing guide exactly first, propose improvements after you have shipped. `#action/h2` - Internal/external/cross-project: the "record and review" third of a skill is missing from nearly every published skill. Propose it as a convention. `#action/h1` #### OSPO - Policy & Governance: write guidance on pointing agents at personal or internal knowledge bases — what may be indexed, by which agents, under what retention. `#action/h1` - License Compliance: encoding a copyrighted book's framework into a distributable skill has an unexamined licensing dimension. Worth a position before people share these publicly. `#action/h1` - Security Management: whole-vault agent access is a data-exposure surface that current tooling does not model. `#action/h2` - Software Inventory: not applicable to this source. `#action/h3` - Developer Support: office hours on building skills from internal methodology docs using this three-part anatomy. `#action/h2` - Community Engagement: not applicable to this source. `#action/h3` - Strategic Alignment: the feedback-loop pattern applies to org practices, not just personal ones — encode a practice, track adherence, review outcomes. `#action/h2` #### Community - Maintainers: your CONTRIBUTING.md is a Shu document. Write it so it can be followed exactly before it is customized. `#action/h2` - Contributors: build a skill from your project's contribution guide that checks your PR against it before you open it. `#action/h1` - Sponsors: not applicable to this source. `#action/h3` - Ecosystem: a shared convention for skills that record and review their own outcomes would raise quality across every published skill library. `#action/h2` #### WordPress - Core: not applicable to this source. `#action/h3` - Core Performance: not applicable to this source. `#action/h3` - Core AI: the when/how/record anatomy is a good specification template for any ability WordPress ships. The record-and-review third is the one everyone omits. `#action/h1` - Core Program Team: not applicable to this source. `#action/h3` - Mobile: not applicable to this source. `#action/h3` - Playground: Playground is an ideal place to run a skill against a real site and record results without risk. `#action/h2` - Community: meetup format — build a skill from the WordPress coding standards and test it on your own plugin. `#action/h2` - Training: Learn WordPress should teach skill authoring using this anatomy; it is more durable than prompt tips. `#action/h1` - Marketing: not applicable to this source. `#action/h3` - Support: not applicable to this source. `#action/h3` - Design: not applicable to this source. `#action/h3` - Documentation: **the strongest item here.** Handbooks are Shu documents feeding agent context. Publishing them as clean, complete, uncompressed text — not summaries — directly improves agent output quality for everyone. `#action/h1` - Photos: not applicable to this source. `#action/h3` - TV: not applicable to this source. `#action/h3` - Polyglots: not applicable to this source. `#action/h3` - Accessibility: build a skill from WCAG 2.2 AA that checks work and records which failures recur — the recurrence data is the valuable part. `#action/h1` - Meta: not applicable to this source. `#action/h3` - CLI: not applicable to this source. `#action/h3` - Hosting: not applicable to this source. `#action/h3` - Tide: the record-and-review pattern is what Tide scoring should be — track whether flagged issues actually get fixed over time. `#action/h2` - Openverse: not applicable to this source. `#action/h3` - Themes: not applicable to this source. `#action/h3` - Plugins: a skill built from the plugin guidelines that checks a submission before review would cut reviewer load measurably. `#action/h1` #### The WPCC - Programs: not applicable to this source. `#action/h3` - Projects: a shared library of skills built from WordPress standards documents — coding standards, plugin guidelines, accessibility handbook — is concrete and useful. `#action/h2` - Funding experiments: not applicable to this source. `#action/h3` ### FAIR - Technical independence: the technique is agent-agnostic — she names Codex and Claude Code interchangeably. Skills built this way are portable, which is the independence argument in practice. `#action/h2` - Policies and covenants: not applicable to this source. `#action/h3` - Productized governance artifacts: governance documents are Shu documents. Building skills that check adherence and record outcomes turns a charter into something measurable. `#action/h1` ## 🧪 Hypotheses 1. **Skills built from full source text outperform skills built from summaries, measured by specificity of the checks they generate.** The abstraction-layers argument predicts this directly, and it is the video's central empirical claim. `#atom/literature/h1` 2. **Skills that include a record-and-review component produce behavior change; skills without it do not, regardless of quality.** The loop, not the instruction, is the active ingredient. `#atom/literature/h1` 3. **Agent output quality is more sensitive to available context than to prompt phrasing; chat output is the reverse.** Her explicit claim, and testable. `#atom/literature/h1` 4. **Personal note corpora yield materially different skill personalization than the same book without them.** Demonstrated once, on herself, with no control. `#atom/literature/h2` 5. **Practitioners who follow a framework unmodified for N iterations before customizing outperform those who customize immediately.** The Shu claim — widely asserted across domains, rarely measured. 6. **Humans override correct agent findings at a measurable rate when the finding contradicts a prior belief about their own work.** Visible once in this video's own demo, from a presenter arguing for evidence-driven loops. ### 📊 Methodology - **H1 (source vs. summary).** Same book, same task, three conditions: full text, publisher summary, model-generated summary. Generate skills from each; have blind raters score the resulting checks for specificity and actionability. Also count concrete checkable items per skill, which is objective and needs no rater. Directly runnable this week. - **H2 (loop is the active ingredient).** Randomize participants to a skill with tracking versus the identical skill without it. Measure behavior change at 30 days by observable output, not self-report. Confounder: the tracking condition adds contact frequency, so include a third arm with tracking prompts but no method to separate attention from mechanism. - **H3 (context vs. prompt).** 2×2: casual prompt / careful prompt × agent with file access / chat without. Fixed task battery, blind quality scoring. Prediction: large main effect for context, small for prompt in the agent condition, reversed in chat. This is a clean, cheap experiment and nobody appears to have run it. - **H4 (personal corpora).** Within-subject: build a skill with and without vault access for the same person and book, then have that person rate which better targets their actual gaps. Subjective by necessity — the construct is personal relevance — so report as preference data, not accuracy. - **H5 (Shu discipline).** Hardest to test, most valuable if answered. Cohort of learners adopting a defined framework, randomized to follow-exactly-for-N-iterations versus customize-freely. Objective outcome measure required. Attrition will be the main threat, since the follow-exactly arm is less pleasant. - **H6 (override rate).** Instrument agent sessions and count instances where a user asserts a correction without verification, then check ground truth. Ethically requires consent and is more feasible as self-audit: review your own transcripts for "actually that's fine, update your analysis" and see how often you were right. ### ❓ Questions - Where does "give it more context" stop being true? Context windows are finite and attention degrades over long inputs — the video's advice has a ceiling it never locates. - Is *"learning doesn't happen until we change our behavior"* a useful operating principle or a general theory? It works for skill acquisition and excludes conceptual learning by definition. - What are the licensing implications of encoding a copyrighted book's framework into a shareable skill? She uses Dib's nine grids; the skill effectively redistributes the framework. - What is the actual data-handling posture when you point an agent at a whole Zettelkasten? What is retained, what is sent upstream, what is used for training? - How does a skill built against one model behave under the next version? Her skills are prose; prose tuned for one model steers the next differently. - Would the checkout finding have held up if she had investigated instead of overriding? The one unresolved thread in her own demo. - Does the record-and-review component actually get used after the video ends? She promises to return with launch results. Follow-through on personal tracking systems is historically poor. - Can this run in reverse — an agent reading your notes to tell you which book to read for the gap you actually have? ### 🔮 Future Research - **Optimal context density for agent tasks.** The video argues against pre-compression, correctly, and stops there. Somewhere between a one-page summary and a full corpus there is an optimum that varies by task and model, and nobody has mapped it. - **Personal knowledge bases as personalization substrate.** Using accumulated notes to infer a learner's gaps is a genuinely new capability and the most novel idea in this video. Whether inferred gaps match assessed gaps is directly testable and would validate or undercut the whole technique. - **Whether AI-mediated feedback loops survive contact with real life.** The friction argument is sound in principle. Longitudinal data on whether people actually maintain agent-tracked practice loops past 90 days does not exist. - **The emotional-labor argument for agents.** "Emotionless support" is the most interesting under-examined claim here — that the machine's value in learning is affective neutrality rather than capability. Testable against human-coached and self-directed controls. - **Human override of correct machine findings.** A specific, measurable failure mode visible in this video's own demo, relevant to every agent workflow, and largely unstudied. ### 💡 Implications - **Practice.** Two things to do this week. Stop feeding models summaries when you have the source. And add a record-and-review component to a skill you already use — the tracking, not the instruction, is what changes behavior. - **Policy.** Organizations pointing agents at internal knowledge bases need a stance on scope and retention before the practice spreads. The technique is compelling enough that it will spread whether or not the policy exists. - **Security.** Whole-corpus agent access is a data-exposure pattern that current tooling does not model — no scanner reports "an agent read your entire notes directory." The value is real and so is the surface. - **Education.** Shu-Ha-Ri is a better frame for teaching frameworks than most instructional design offers, and the video's own contribution — that skipping Shu is discomfort disguised as disagreement — names the actual failure mode. That belongs in any curriculum teaching a method. - **Funding models.** Not applicable to this source, and worth saying rather than stretching. ### 🧠 Learning Styles Assessment - **Reading level estimate:** spoken, conversational; Flesch-Kincaid Grade Level roughly 8–9, Flesch Reading Ease around 62–70. Vocabulary is plain, sentences are long and self-interrupting in the way speech is, and the two imported terms (Shu-Ha-Ri, Zettelkasten) are both defined on introduction. #### 👀 Visual learners - Present: strong. On-screen abstraction-layer diagram, the Shu-Ha-Ri progression, live screen recording of the Codex build, the generated skill document, and the actual findings against her sales page. - Improve: the three-part skill anatomy — when/how/record — is the video's most portable idea and never appears as a graphic. A single card would make it screenshot-able and shareable, which is how ideas travel. #### 🔊 Aural learners - Present: clear delivery, deliberate pacing, verbal signposting between sections ("first of all," "the second thing," "as a bonus"). - Improve: frequent self-interruption and restarts ("What is—what is this thing," "We are brushing up against") make the audio harder to follow than the ideas warrant. Chapters would let a listener re-enter at the framework sections. #### 🗣️ Verbal learners - Present: the video's greatest strength. *"intellectual fantasizing about a better world"* and *"an abstraction of an abstraction of an abstraction"* are both memorable and precise. Both imported concepts are defined before use. - Improve: "64/4" arrives at 25:44 with no explanation beyond the arithmetic — a viewer unfamiliar with Pareto-squared will not follow why it matters. #### 🧩 Kinesthetic learners - Present: excellent. She builds the skill live, tests it against a real asset, and shows the unedited findings including ones that make her look unprepared. A viewer can replicate the whole thing. - Improve: no starting prompt is provided in text form, so viewers must transcribe from speech. The single highest-value addition would be the prompt in the description. #### 🧠 Logical learners - Present: the argument is genuinely well-constructed — diagnosis (abstraction layers), specification (three parts), theory (Shu-Ha-Ri), evidence (her own failure), demonstration (live build), extension (personal notes). Each stage motivates the next. - Improve: two premises are load-bearing and unexamined — the behaviorist definition of learning, and "more context is better" with no ceiling. And the checkout override is a logical lapse inside an argument about evidence. #### 🧑‍🤝‍🧑 Social learners - Present: credits Allan Dib warmly and Christo by name; shares a genuinely unflattering personal failure. - Improve: skills are framed as a solo artifact. Nothing about sharing them, whether someone else's skill can be trusted, or how a team might maintain a library — which is where this goes next. #### 🧘‍♀️ Solitary learners - Present: ideally suited. The whole method is individual, and the Zettelkasten extension rewards solitary accumulated work. - Improve: no guidance on which books are worth this treatment. The process is non-trivial and applying it to every book would be a poor use of it. #### 🌿 Naturalistic learners - Present: nothing environmental. - Improve: the Shu-Ha-Ri progression is developmental and the feedback loop is ecological in shape — she has the vocabulary available and does not use it. A minor missed enrichment, not a flaw. ### 📊 Accessibility Assessment *Assessment of the video as published on YouTube.* #### Screen reader semantics - Present: descriptive, accurate title. Standard YouTube chrome supplies landmarks. - Improve: substantial unnarrated screen content — the generated skill document, the findings list, the Zettelkasten analysis — is read only partially aloud. She summarizes rather than reads, so a non-sighted viewer gets the gist but not the artifact. WCAG 2.2 AA 1.2.3 Audio Description or Media Alternative. #### Keyboard navigation - Present: the YouTube player is keyboard operable. - Improve: not applicable to the content itself. #### Contrast and readability - Present: screen recordings at native resolution; the diagrams appear high-contrast. - Improve: Codex output is small monospace text on screen. Verify legibility at phone sizes, where much of a 28-minute video is watched. WCAG 2.2 AA 1.4.3 Contrast (Minimum), 1.4.4 Resize Text. #### Text resizing and reflow - Present: YouTube handles player reflow. - Improve: recorded terminal and editor UI does not reflow. Zooming into the region under discussion is the standard remedy and is used inconsistently. WCAG 2.2 AA 1.4.10 Reflow. #### Media and captions - Present: auto-captions available, and noticeably more accurate than the Lipsky video's — proper nouns like Codex, Obsidian, and Zettelkasten survive. - Improve: "Shu-Ha-Ri" and "Allan Dib" are the terms most at risk in auto-captioning, and both are load-bearing. A corrected caption track would fix that cheaply. No audio description for the dense screen-work sections. WCAG 2.2 AA 1.2.2 Captions (Prerecorded), 1.2.5 Audio Description (Prerecorded). #### Forms and interactions - Present: not applicable to video content. - Improve: the linked free resource should be checked separately for labeled inputs, since it is gated behind a signup. #### Standards reference - Primary gap: **1.2.5 / 1.2.3** — a substantial portion of the value is on-screen text summarized rather than read, which disadvantages non-sighted viewers in an instructional video. Secondary: **1.4.4** (small monospace at mobile sizes). Captions are adequate, which is better than typical for this genre. ### 🌍 DEIB Assessment #### Cultural breadth - Present: genuinely good. Shu-Ha-Ri is introduced with its origin in Japanese martial arts and tea ceremony, and she offers the Japanese and Chinese senses of the characters rather than stripping the concept of provenance. Zettelkasten is likewise named and situated. - Improve: business examples are US-centric (Apple, Nike, Patagonia), and the economic assumption is unstated — Codex or Claude Code plus a paid agent tier plus an established Obsidian practice is a real cost and time floor. #### Inclusive language - Present: warm and self-deprecating; "wonderful dinosaurs who still read books" is affectionate rather than dismissive. Terms are defined on introduction, and she avoids treating agent use as an in-group marker. - Improve: "It is not technical, especially if you're non-technical. Don't worry, it's not just for coding" is reassuring but arrives after several minutes of terminal work that will read as technical to the audience it is aimed at. The reassurance and the demonstration are slightly out of step. #### Neurodiversity support - Present: consistent structure, explicit enumeration ("three things," "step one, step two, step three"), and concrete examples before abstraction. - Improve: frequent self-interruption and mid-sentence restarts raise processing load. No chapter markers in a 28-minute video that functions as a reference. A summary card of the three-part anatomy would serve re-entry. #### Representation - Present: credits Allan Dib and Christo by name, and shares a failure that reflects poorly on her own early judgment — which is a form of representation that helps beginners more than polished expertise does. - Improve: only two other people are named, both in an expert or peer role. No learner voices, and no acknowledgement that the accumulated-Zettelkasten path is unavailable to anyone without years of notes. #### Psychological safety - Present: **the strongest dimension.** She explicitly names learning as emotional — ego, uncertainty, fear that you may never learn the skill — and treats that as normal rather than as weakness. Her own failure story is told without self-flagellation, and the emotionless-support argument is framed as removing shame from the process rather than removing the human. - Improve: the repeated push toward agents ("just switch over," "you've got to at least try") edges toward pressure, and viewers on free tiers or without file-access agents are not told they cannot follow along. ### ✅ Logical Validity #### Strong reasoning - **The abstraction-layers diagnosis (02:03).** A causal chain in which each step is independently plausible and the conclusion follows: expertise → book → summary → skill, with nuance and examples stripped at each stage, therefore generic output. The supporting premise — *"We don't just learn from abstract concepts. We need to see how they play out specifically"* — does real work rather than decorating the claim. The best reasoning in the video. - **Correct identification of a category error (03:23).** *"specifically for AI, we don't want to do this because AI is really good at digesting information."* Recognizing that summarization is an accommodation to human limits, wrongly applied to a system without those limits, is a genuine insight and explains a broad class of failures usually blamed on model quality. - **Shu reframed as experimental control (11:53).** *"this is how scientific experiments work. You're not going to change all the variables at once."* This converts an appeal to tradition into a methodological argument. Following the framework unmodified establishes a baseline; without one, customization has nothing to be measured against. The reasoning stands independent of the martial-arts provenance, which is what makes it persuasive rather than decorative. - **Self-implicating evidence (10:52).** Using her own documented failure — inventing principled-sounding reasons to skip steps she was uncomfortable with — as the evidence for her thesis is both honest and unusually strong support, because it is costly to offer and identifies a mechanism rather than just an outcome. - **The emotionless-support argument (13:06).** A real argument for why a machine belongs in a learning loop, resting on a stated premise (learning is emotionally threatening) and a property of the tool (no ego to defend). Most AI-productivity content asserts usefulness; this derives it. - **Personalization premise (21:55).** *"even though the method is the same, where our knowledge gaps are and why we find this method helpful will be different."* Correct, and it justifies the Zettelkasten step rather than presenting it as a neat trick. #### Fallacies - **Stipulative definition presented as fact** — 00:00: *"learning doesn't happen until we change our behavior... it was just intellectual fantasizing about a better world."* This defines learning as behavior change, excluding conceptual and declarative learning by fiat rather than argument. It is a defensible operating principle for skill acquisition and a contested claim as general theory, stated as settled. **Fix:** "for the kind of learning I care about here — where you want a different result — nothing has happened until behavior changes." `#fallacy/unsubstantiated-claim` - **Unbounded premise** — 03:23 and throughout: *"AI is really good at digesting information,"* used to justify supplying maximal raw context. True relative to humans, and bounded in ways never mentioned — finite context windows, attention degradation over long inputs, and cost. The advice is directionally right and has a ceiling the video does not locate, which will produce failures viewers cannot diagnose. **Fix:** name the ceiling and how to tell you have hit it. `#fallacy/hasty-generalization` - **Overriding evidence to preserve a prior** — 19:25: the agent reports checkout renders zero; she offers an untested explanation (*"Maybe it's a browsing issue"*), does not test it, and instructs *"Checkout is working fine. So update your analysis."* She is likely correct about her own site. But this is precisely the failure the video exists to prevent — a finding surfaced, a prior asserted, the finding overwritten without verification — occurring inside the demonstration of an evidence-driven method. **Fix:** verify, then correct. The verification would have taken a minute and would have strengthened the demo either way. `#fallacy/begging-the-question` - **Generalization from n=1** — 21:55 to 25:44: the Zettelkasten personalization is demonstrated once, on herself, with no comparison against the un-personalized skill and no check on whether the inferred gaps match her actual gaps. She confirms them, but she is both subject and judge. **Fix:** acknowledge it as a single illustrative case. `#fallacy/anecdotal-evidence` - **Repeated product advocacy inside instruction** — 15:44, 16:26, 17:01, 27:11: four separate pushes to move from chat to agents, one self-flagged (*"I'm going off topic"*), closing with a lead magnet. The underlying claim (context matters more than prompting for agents) is sound and worth making once. Repeating it four times converts an argument into a funnel. `#fallacy/red-herring` - **Appeal to authority via tradition** — 07:07: Shu-Ha-Ri's martial-arts and tea-ceremony provenance is offered partly as warrant. She rescues this at 11:53 with the experimental-control argument, which is why this is a mild finding rather than a serious one — but for four minutes the framework's authority rests on its age. `#fallacy/appeal-to-tradition` #### Tone of voice Thoughtful, conversational, and unusually willing to be unimpressive. The register is a practitioner thinking out loud rather than an expert delivering conclusions — sentences restart, she catches herself going off topic, and she narrates uncertainty in real time (*"I'm not sure what happened"*). That texture costs polish and buys credibility, and it is the right trade for content whose central claim is that you should follow a method before you have mastered it. The self-deprecation is specific rather than performative. *"I was just such a mess"* attaches to a described sequence of decisions, not to a vague admission of imperfection, and the failure she reports — inventing principled reasons to avoid uncomfortable steps — is genuinely unflattering in a way that most creator vulnerability is not. The one register shift is the agent advocacy, which arrives four times and flattens into pitch. She half-acknowledges it (*"just a small plug,"* *"I'm going off topic"*), which is more honest than most, but the repetition still moves the video from argument toward funnel at those moments. The contrast is instructive: everywhere else she earns the claim, and there she asserts it. ### 🛡️ DARVO Analysis DARVO describes an accused party's response pattern — Deny, Attack, Reverse Victim and Offender. **No DARVO pattern is present.** This is instructional content built around the presenter's own admitted failure. There is no accuser, no allegation, and no defensive posture — and structurally, a video whose central evidence is self-indictment is close to the opposite of the pattern. Component by component: - **Deny:** Absent, and inverted. Rather than denying fault, she volunteers it in detail: *"secretly, I'm not comfortable doing some of them. So I was making up excuses for why I'm not going to do them now."* That is an admission of motivated reasoning against her own interest. The one moment resembling denial is the checkout exchange at 19:25, where she overrides an agent's finding without verifying — but that is a machine's report, not a person's accusation, and the failure there is epistemic (unverified prior beating surfaced evidence), not defensive. Filing it as DARVO would misclassify it, and the distinction matters: DARVO's deny protects the self against another person's claim, while this protects a belief against a tool's output. - **Attack:** Absent. Nobody is disparaged. Allan Dib is credited warmly and his framework treated as sound — notably, her failure story blames herself rather than the book, when blaming the book was available and easier. Christo is credited positively. Chat interfaces are argued against on capability grounds rather than dismissed. Even the "wonderful dinosaurs who still read books" line is affectionate toward its target. - **Reverse victim/offender:** Absent. No grievance is claimed anywhere. She positions herself as someone who made an avoidable mistake and is reporting the correction — the opposite of recasting oneself as wronged. The emotional content she does surface (ego, fear, uncertainty in learning) is framed as universal and normal, not as harm done to her. **Emotional cues and rhetoric.** Warm, self-deprecating, occasionally uncertain in real time. The persuasive pressure that exists runs toward product adoption (four pushes toward agents plus a closing lead magnet), which is commercial rhetoric, not defensive rhetoric. **Related fallacies.** The findings above — `#fallacy/begging-the-question` on the checkout override, `#fallacy/unsubstantiated-claim` on the definition of learning, `#fallacy/red-herring` on the repeated agent advocacy — capture the real weaknesses. All are reasoning and framing issues in good-faith instructional content with no adversary, a materially different class from DARVO. `#darvo/deny` `#darvo/attack` `#darvo/reverse` — all three assessed. None present. ## 🎯 Conclusion - **Bottom line.** The best-argued video in this batch. Its central diagnosis — that turning a book into an AI skill fails because you hand the model an abstraction of an abstraction of an abstraction — is correct, well-supported, and generalizes far past books; the reframe that **summaries are a human affordance rather than a machine one** explains a whole class of disappointing outputs usually blamed on the model. The three-part skill anatomy (when · how · record-and-review) is a better specification template than most prompt-engineering advice, and the record-and-review third is the part nearly every published skill omits. Shu-Ha-Ri is rescued from mere tradition by the experimental-control argument, and her own failure story identifies the real mechanism: skipping Shu is discomfort wearing the costume of intellectual disagreement. The Zettelkasten extension is the most actionable idea here and the strongest argument I have seen that note-taking survives capable models — your notes record what *you* found important, which no general model can reconstruct. Two premises go unexamined: the behaviorist definition of learning, and "more context is better" with no stated ceiling. And in one telling moment she overrides a true-or-false agent finding without investigating, inside a video about evidence-driven feedback loops. - **Next steps for viewers.** 1. Stop feeding models summaries when you have the source. This is the single highest-value change and it costs nothing. 2. Add a record-and-review component to one skill you already use. The tracking is the active ingredient; the instructions are not. 3. Point an agent at your existing notes on a book or topic and ask it to infer your knowledge gaps from what you repeatedly highlight. Decide first what you are comfortable exposing — this is a data decision, not just a technique. 4. Pick one framework you have been "customizing" and run it unmodified for several iterations. Ask honestly whether your customizations were reasoned or merely comfortable. 5. Watch for the override pattern in your own sessions. When an agent reports something that contradicts what you know about your own work, verify before you correct it — that is the exact hinge where an evidence loop becomes a confirmation loop. 6. Note the ceiling she does not mention: more context helps until the window fills and attention degrades. If output quality drops as you add material, you have found it. ## 🔄 Refactor Appendix — Turn Any Book Into an AI Skill You'll Actually Use ### 📚 Literature Atoms (H3) #### [H1] Summaries are a human affordance, not a machine one > [!quote] ID:rw_vzbook_h1 > "specifically for AI, we don't want to do this because AI is really good at digesting information. So we don't want to give it an abstraction of an abstraction." **Context & use**: - 03:23 — the turn in the abstraction-layers argument. - Compression exists for human cognitive limits; applying it before handing material to a model is a category error. Explains a large class of disappointing outputs usually blamed on the model. Bounded by context windows, which the video never mentions. **Tags**: #xref #articleseed #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=203) <!-- atomicity: 1 idea ✓ --> #### [H2] A working skill encodes when, how, and record-and-review > [!quote] ID:rw_vzbook_h2 > "we want to know when to use the method... the second thing is how to make the change... Then the third thing is to record and review what happened." **Context & use**: - 03:23–04:33 — the skill anatomy. - Most published skills carry only the middle third. The record-and-review component is what converts an instruction into a feedback loop, and it is the part everyone omits. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=273) <!-- atomicity: 1 idea ✓ --> #### [H3] Skipping Shu is discomfort disguised as disagreement > [!quote] ID:rw_vzbook_h3 > "secretly, I'm not comfortable doing some of them. So I was making up excuses for why I'm not going to do them now." **Context & use**: - 10:52 — her *1-Page Marketing Plan* failure. - The mechanism, not just the outcome: the excuse arrives already dressed as strategy. Self-implicating evidence, which is costly to offer and therefore strong. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=652) <!-- atomicity: 1 idea ✓ --> #### [H4] Your notes record what you found important, not what exists > [!quote] ID:rw_vzbook_h4 > "I've already done my part of the job, which is use my judgment, pick out what's important, document it, take notes, process the information" **Context & use**: - 23:18 — the division of labor in the Zettelkasten step. - The strongest available argument that note-taking survives capable models. Human judgment produces the corpus; the agent does retrieval and synthesis over it. A general model can reconstruct the book but not your attention. **Tags**: #xref #articleseed #addendum **Highlight**: [view highlight](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=1398) <!-- atomicity: 1 idea ✓ --> #### [H5] AI's contribution to learning is affective neutrality, not capability > [!quote] ID:rw_vzbook_h5 > "AI doesn't have emotion, and we should use that to our advantage... without us relying on staying disciplined and staying rational all the time." **Context & use**: - 13:06 — why a machine belongs in the loop at all. - Most AI-productivity content asserts usefulness; this derives it from a stated premise (learning is ego-threatening) and a property of the tool (no ego to defend). The machine's value is that it does not flinch. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=8PM_t3NWrAE&t=786) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Compress for humans, not for machines **Claim** (one sentence, Courtney's voice): Summarize material for people who must hold it in working memory, and hand systems that have no such limit the uncompressed source instead. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds wherever a machine consumer has capacity a human reader lacks — agent context, search indexing, retrieval corpora. Bounded by finite context windows, attention degradation over long inputs, and cost; the ceiling shows up as output quality dropping when you add material. Does not hold where the compression itself encodes judgment you want preserved, which is exactly what [[#H4]] is about — a note is not a summary, it is a selection. **Links** (≥1 with annotation): - [[Context engineering]] — because extends - [[Progressive summarization]] — because contrasts on who the compression serves - [[Retrieval augmented generation]] — because applies-in corpus design <!-- atomicity: 1 idea ✓ --> #### [P2] Follow the method unmodified long enough to earn a baseline **Claim** (one sentence, Courtney's voice): Customizing a framework before running it as written leaves you with no baseline, so you cannot tell whether your changes helped. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for any framework you adopted because you lacked expertise in its domain — the condition that motivated adoption is the same one that disqualifies your early judgment about which parts to skip. Does not hold where you already have domain expertise and are integrating a new framework into an established practice, which is legitimate Ha. Diagnostic: if the steps you want to skip are the uncomfortable ones rather than the inapplicable ones, that is discomfort, not judgment. **Links** (≥1 with annotation): - [[Shu-Ha-Ri]] — because extends - [[Experimental control]] — because applies-in why the baseline matters - [[Cargo cult practice]] — because contrasts on following without understanding `proof_state: working` <!-- atomicity: 1 idea ✓ --> #### [P3] An unverified prior that overrides a surfaced finding turns an evidence loop into a confirmation loop **Claim** (one sentence, Courtney's voice): The moment you correct a tool's finding from memory instead of checking, your feedback loop stops producing evidence and starts producing agreement. **Evidence pointer**: [[#H2]] **Evidence**: > [!quote] ID:rw_vzbook_p3 > "I've checked this before—but it doesn't seem to be a problem. I'm not sure what happened. Maybe it's a browsing issue." **Scope & conditions**: Holds for any human-in-the-loop review — agent findings, linters, test failures, code review, monitoring alerts. Sharpest when the human has genuine expertise about the artifact, because that is when overriding feels most justified. Legitimate override exists: when verification cost exceeds the finding's value and you record the dismissal. The failure is the unrecorded, unverified one. **Links** (≥1 with annotation): - [[Alarm fatigue]] — because extends - [[Confirmation bias]] — because extends - [[Falsification]] — because contrasts on what evidence is for `proof_state: working` <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Point Codex at my Zettelkasten and have it infer knowledge gaps from what I repeatedly highlight. Directly applicable — I already have the vault and the agent. Decide scope first. #fleeting #inbox - [F2] Audit my existing skills for the missing third: which ones record and review outcomes? I suspect none do. #fleeting #inbox - [F3] The abstraction-layers argument applies to my own deep-read pipeline — am I feeding models summaries where I have source text? Check `article_template` inputs. #fleeting #inbox - [F4] Post idea: "compress for humans, not for machines" — the category error behind a lot of disappointing AI output. #fleeting #articleseed - [F5] Self-audit: grep my own transcripts for the override pattern ("that's fine, update your analysis"). How often was I actually right? #fleeting #inbox - [F6] Where does "more context" stop helping? Nobody in this genre names the ceiling. Worth measuring on a fixed task battery. #fleeting #inbox - [F7] Licensing question: encoding a copyrighted book's framework into a shareable skill. Does the skill redistribute the framework? #fleeting #inbox - [F8] Shu-Ha-Ri maps onto my credit-routing policy — follow the routing matrix as written before optimizing it. #fleeting #xref ### → Things 3 (max 2 actions) - [ ] Point an agent at the vault Zettelkasten to infer knowledge gaps on one topic — knowledge work #things - [ ] Add a record-and-review component to one existing skill and test it for two weeks — tooling #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Turn Any Book Into an AI Skill You’ll Actually Use.md|Turn Any Book Into an AI Skill You’ll Actually Use by Vicky Zhao]] ## Highlights - it an abstraction of an abstraction. So instead, what we'll do is make sure ([View Highlight](https://read.readwise.io/read/01m1c8p69yn3y4k5efpm7aqbgb)) - Note: The best diagnosis in the batch. Expertise → book (abstraction 1) → summary (abstraction 2, loses the examples) → model-generated skill from that summary (3, maybe 4) → predictably generic output. The reframe that makes it click: summaries are a HUMAN affordance, not a machine one. We compress because we can't hold 200 pages; a model can. Pre-compressing before handing it over is a category error, and it explains a whole class of disappointing output usually blamed on the model. Bounded by context windows, which she never mentions. - Tags: atom/zettel/z1, atom/literature/h1 - let's now bring in AI, because what AI is really good at doing is providing emotionless support. ([View Highlight](https://read.readwise.io/read/01m1c8qf7jjg4ekwg255aa9jha)) - Note: The best-argued justification in the batch for why a machine belongs in a learning loop — and it isn't capability. Learning is emotionally threatening (uncertainty, ego, fear you may never get it), and an agent has no ego to protect and no discomfort to rationalize around, so it holds you to the method "without us relying on staying disciplined and staying rational all the time." Most AI-productivity content asserts usefulness; this derives it. The machine's value is that it doesn't flinch. - Tags: atom/literature/h1