Why you should take notes if you use AI - CourtneyR.dev's 2nd Brain

Why you should take notes if you use AI 🔗 **Source:** [Why you should take notes if you use AI](https://www.youtube.com/watch?v=FtJYJF8QodM) > [!note]+ Document Notes > tags: [context-engineering, note-taking, ai-agents, knowledge-management, obsidian, second-brain] > author: Vicky Zhao > source: YouTube > reference: https://www.youtube.com/watch?v=FtJYJF8QodM > date created: 2026-02-02 00:00 > > > [!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 Do not quote or attribute any other line to the named author. > > > > **Repaired 2026-08-31:** frontmatter previously credited this video to Eleanor Konik with a > > 2024-12-28 date. The video is by **Vicky Zhao**, published 2026-02-02. The body analysis below > > is from the original pass and is retained; the Conclusion and Refactor Appendix were missing > > and have been added. > > # 🏷️ Title: Why you should take notes if you use AI >  ## 🧰 Action Inbox - [H1] Set up note-taking system for AI context — immediate priority for AI users #setup - [H2] Document tacit knowledge into explicit notes — convert mental models to shareable content #documentation - [H3] Create frameworks for AI judgment criteria — establish quality standards for outputs #frameworks ## 🧭 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 - [Claude Code Setup](https://claude.ai) - [Obsidian](https://obsidian.md) ## 📝 Summary - Eleanor argues that AI users must develop note-taking systems to move from prompt engineering to context engineering - Why it matters: Quality AI outputs depend on providing rich, personal context rather than generic prompts - Key audience: Knowledge workers using AI tools who want deeper, more personalized results - Risks: Without notes, users outsource thinking and get generic outputs; with notes, they can think deeper - Bottom line: Notes transform AI from a generic tool into a personalized thinking partner ## 📖 Main Content The core argument centers on the evolution from prompt engineering to context engineering. While prompt engineering focuses on crafting better requests to AI, context engineering involves providing rich background information that enables more sophisticated AI assistance. Eleanor demonstrates this through a practical example comparing generic ChatGPT responses about career advice versus Claude's analysis of her 39,000-token Obsidian vault. The difference is dramatic: generic advice versus personalized insights that synthesize her documented thinking patterns. The framework includes six categories of information for AI context: role/goal/audience (traditional prompt engineering), plus inputs/source of truth/judgment criteria (context engineering additions). This systematic approach transforms AI from a basic assistant into a thinking partner that can identify patterns and synthesize ideas from personal knowledge bases. ## 🗒️ Notes #### ⏱️ [00:00](https://youtu.be/FtJYJF8QodM?t=0) — [H1] The shift from prompt to context engineering - Key idea: AI effectiveness depends more on providing rich context than crafting perfect prompts - Evidence: "We're moving away from prompt engineering and the increasing importance of context engineering" #### ⏱️ [01:15](https://youtu.be/FtJYJF8QodM?t=75) — [H2] Converting tacit to explicit knowledge - Key idea: Notes transform internal knowledge into shareable, AI-readable formats - Evidence: "Turn tacit knowledge that just exists in our head into explicit knowledge so we can share with other people" #### ⏱️ [02:45](https://youtu.be/FtJYJF8QodM?t=165) — [H3] The three layers of communication - Key idea: Communication has evolved from internal thoughts to written sharing to AI collaboration - Evidence: "Now we're adding another layer - it's not just us and other people but there's also an AI" #### ⏱️ [03:30](https://youtu.be/FtJYJF8QodM?t=210) — [H2] Six categories for AI context - Key idea: Effective AI interaction requires six types of information beyond basic prompting - Evidence: "Six categories we have to think about... role, goal, audience, constraints, inputs, source of truth, judgment" #### ⏱️ [04:45](https://youtu.be/FtJYJF8QodM?t=285) — [H3] Defining inputs for AI systems - Key idea: Explicitly specify what data the AI can and cannot access from your knowledge base - Evidence: "Tell it explicitly what's the input it can use... everything in my vault except for these tags" #### ⏱️ [05:20](https://youtu.be/FtJYJF8QodM?t=320) — [H3] Establishing source of truth hierarchy - Key idea: Specify which documents should override AI's general training when conflicts arise - Evidence: "What kind of documents outrank its generic information... I don't want you to use them" #### ⏱️ [06:00](https://youtu.be/FtJYJF8QodM?t=360) — [H3] Creating judgment frameworks for AI - Key idea: Provide explicit criteria for evaluating output quality rather than relying on intuition - Evidence: "How does LLM judge what is good, what is bad? This is where we need to take tacit knowledge into explicit knowledge" #### ⏱️ [07:15](https://youtu.be/FtJYJF8QodM?t=435) — [H2] Framework limitations in AI outputs - Key idea: AI systems default to generic frameworks when not provided with specific alternatives - Evidence: "The problem with its frameworks is it keeps using the same ones and very generic ones" #### ⏱️ [08:30](https://youtu.be/FtJYJF8QodM?t=510) — [H3] Using examples to generate frameworks - Key idea: AI can distill quality criteria from good and bad examples when explicit frameworks don't exist - Evidence: "10 examples of things that are great and 10 examples that are terrible, get the LLM to create that framework" #### ⏱️ [09:45](https://youtu.be/FtJYJF8QodM?t=585) — [H1] Demonstration: Generic vs contextualized AI responses - Key idea: Context dramatically improves AI output quality and relevance - Evidence: Shows ChatGPT giving generic career advice vs Claude analyzing personal vault #### ⏱️ [12:00](https://youtu.be/FtJYJF8QodM?t=720) — [H2] Generic AI advice limitations - Key idea: Without context, AI provides broadly applicable but not personally actionable advice - Evidence: "I glaze over everything... It is very generic, this is best practice so to speak" #### ⏱️ [15:30](https://youtu.be/FtJYJF8QodM?t=930) — [H1] Contextualized AI synthesis breakthrough - Key idea: AI can identify patterns and articulate ideas better than manual review of personal notes - Evidence: "Fundamentally interested in building the case that creativity and structured thinking are not opposites" #### ⏱️ [18:45](https://youtu.be/FtJYJF8QodM?t=1125) — [H2] AI discovering implicit connections - Key idea: Context-aware AI can surface insights not explicitly documented but implied across notes - Evidence: "I didn't put that thought anywhere explicitly... having it go through 39,000 tokens... is gold" #### ⏱️ [21:30](https://youtu.be/FtJYJF8QodM?t=1290) — [H3] Personalized research recommendations - Key idea: AI can suggest relevant resources based on documented interests and knowledge gaps - Evidence: Shows AI recommending books and authors aligned with personal research interests #### ⏱️ [24:00](https://youtu.be/FtJYJF8QodM?t=1440) — [H1] The compound effect of documented thinking - Key idea: Notes enable ideas to build upon each other rather than starting from scratch each time - Evidence: "If you want ideas to compound and not just build sand castles then you have to document it" ## 📚 References - [Claude AI](https://claude.ai) - [Obsidian Vault](https://obsidian.md) - [Context Engineering Resources](https://claude.ai/setup) ## 🗣️ Paraphrase - AI users need personal knowledge systems because we're shifting from crafting better prompts to providing richer context for more sophisticated assistance - Note-taking transforms internal thoughts into external resources that AI can analyze, synthesize, and build upon in ways human memory and processing cannot match - The quality difference between generic AI responses and contextualized ones is dramatic - context turns AI from a basic tool into a personalized thinking partner ## 📰 Newsletter **Transform Your AI Experience with Smart Note-Taking** If you're using AI tools but feeling frustrated by generic outputs, you're missing a crucial piece: context engineering. Eleanor Konik demonstrates how a robust note-taking system transforms AI from a basic assistant into a personalized thinking partner. The key insight? We're moving beyond prompt engineering to context engineering. Instead of crafting perfect requests, focus on providing rich background information that enables sophisticated AI assistance. Eleanor's comparison is striking: generic ChatGPT career advice versus Claude's synthesis of her 39,000-token knowledge vault. **Your Action Plan:** - Document your tacit knowledge in explicit, searchable formats - Create frameworks for judging AI output quality - Establish clear inputs, sources of truth, and judgment criteria for AI systems This isn't about replacing thinking - it's about augmenting it. When AI can access your documented expertise, it becomes capable of pattern recognition and synthesis that amplifies rather than replaces human insight. Resources: [Claude AI Setup](https://claude.ai), [Obsidian](https://obsidian.md), [Note-taking Systems](https://fortelabs.co) ## ✅ Action Items #### DevRel - Developer advocacy: Create workshops on AI-assisted documentation workflows - Education: Develop curricula on context engineering for developer communities - Developer experience: Build tools for semantic search across codebases and documentation - Developer marketing: Campaign highlighting AI productivity gains through structured knowledge - Community management: Foster discussions on AI-human collaboration best practices #### OSPO - Policy & Governance: Guidelines for AI-assisted code review and documentation - Software Inventory: AI-enhanced SBOM generation using project context - Developer Support: Training on context engineering for internal development teams - Strategic Alignment: KPIs measuring AI productivity gains through knowledge management #### Community - Maintainers: Templates for AI-readable project documentation and decision records - Contributors: Onboarding materials on contributing to AI-enhanced projects - Ecosystem: Cross-project knowledge sharing systems for AI context #### WordPress - Core: AI-assisted code review using WordPress coding standards as context - Documentation: AI-enhanced DevHub content generation using existing documentation patterns - Training: Learning paths on AI-powered WordPress development workflows - Support: AI chatbots trained on WordPress-specific knowledge bases - Design: AI-assisted design system documentation and pattern generation ## 🧪 Hypotheses - Developers with structured knowledge bases will produce higher-quality AI-assisted code than those relying on generic prompting - Teams using context engineering will reduce onboarding time by 40% through AI-enhanced documentation discovery - AI systems trained on project-specific contexts will identify security vulnerabilities more accurately than generic analysis tools - Knowledge workers who document tacit knowledge will demonstrate measurably improved problem-solving outcomes when using AI ### 📊 Methodology - Study design: Randomized controlled trial comparing generic AI prompting vs context-engineered workflows across development teams - Data sources: Code quality metrics, documentation completeness scores, developer productivity measurements - Experiments: A/B testing of AI-assisted onboarding with and without structured knowledge bases ### ❓ Questions - How can we measure the ROI of knowledge documentation efforts in AI-augmented workflows? - What governance frameworks are needed when AI systems access sensitive organizational knowledge? - How do we balance knowledge sharing benefits with security and privacy requirements in AI contexts? ### 🔮 Future Research - Optimal knowledge graph structures for different AI model architectures - Privacy-preserving techniques for context engineering in multi-tenant environments - Automated knowledge extraction from existing codebases and documentation ### 💡 Implications - Practice: Developers need to treat documentation as AI training data, not just human reference material - Policy: Organizations require new governance frameworks for AI access to institutional knowledge - Security: Context engineering surfaces new attack vectors through knowledge injection and extraction - Education: Technical education must include information architecture and knowledge management skills - Funding: Budget allocation should account for knowledge infrastructure as critical AI enablement ### 🧠 Learning Styles Assessment Reading level estimate: Grade 12 (technical content with accessible explanations) #### 👀 Visual learners - Present: Screen recordings showing AI interface comparisons, vault structures - Improve: Knowledge architecture diagrams, context flow visualizations #### 🔊 Aural learners - Present: Video narration with clear explanations and examples - Improve: Podcast version focusing on key concepts and implementation strategies #### 🗣️ Verbal learners - Present: Clear definitions of prompt vs context engineering, explicit frameworks - Improve: Glossary of context engineering terminology, analogy-driven explanations #### 🧩 Kinesthetic learners - Present: Hands-on demonstration of Obsidian and Claude integration - Improve: Step-by-step setup guides, interactive knowledge base templates #### 🧠 Logical learners - Present: Six-category framework, clear before/after comparisons - Improve: Decision trees for choosing context strategies, performance metrics #### 🧑‍🤝‍🧑 Social learners - Present: Community discussion prompts about AI experiences - Improve: Collaborative knowledge base building exercises, peer review systems #### 🧘‍♀️ Solitary learners - Present: Personal reflection on current AI usage patterns - Improve: Self-assessment checklists, private journaling prompts #### 🌿 Naturalistic learners - Present: Evolution metaphor for communication layers - Improve: Ecosystem thinking about knowledge networks, organic growth patterns ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading hierarchy, descriptive video narration - Improve: Transcript with detailed descriptions of visual demonstrations #### Keyboard navigation - Present: Standard video controls and navigation - Improve: Chapter markers for keyboard-only navigation through content #### Contrast and readability - Present: High contrast in UI demonstrations - Improve: Ensure all text overlays meet WCAG contrast requirements #### Text resizing and reflow - Present: Video content adapts to player sizing - Improve: Responsive transcript layout, adjustable font sizes #### Media and captions - Present: Full transcript provided - Improve: Professional captions with speaker identification, timing markers #### Forms and interactions - Present: No interactive forms in content - Improve: Accessible templates for note-taking systems ### 🌍 DEIB Assessment #### Cultural breadth - Present: Examples from Western academic and business contexts - Improve: Note-taking traditions from different cultures, multilingual examples #### Inclusive language - Present: Gender-neutral language, accessible technical explanations - Improve: Define specialized terms, avoid productivity culture assumptions #### Neurodiversity support - Present: Multiple learning modalities, clear structure - Improve: Content warnings for productivity pressure, alternative organization systems #### Representation - Present: Single female perspective on knowledge work - Improve: Diverse examples of knowledge workers, various professional contexts #### Psychological safety - Present: Non-judgmental tone about current AI usage - Improve: Acknowledge different learning paces, validate various approaches ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Context engineering principles lead to better AI outcomes (demonstrated through examples) - Inductive strength: Personal experience generalized to broader AI usage patterns (needs more data points) - Causal reasoning: Clear causal link between note-taking systems and AI output quality (well-supported) - Analogical strength: Communication evolution analogy effectively illustrates the progression (strong) #### Fallacies - Hasty generalization: Single personal example used to support broad claims about AI effectiveness - False dichotomy: Presents note-taking vs not note-taking as binary choice, ignoring gradations #### Tone of voice - Overall tone: Enthusiastic and encouraging, with concrete examples that build credibility and inspire action ### 🛡️ DARVO Analysis - Deny: No denial patterns present - Attack: No attacks on opposing viewpoints - Reverse victim/offender: No victim/offender dynamics ## 🎯 Conclusion - **Bottom line.** The clearest statement in this batch of *why* a note system matters once agents enter the picture: the field is moving from prompt engineering to **context engineering**, and without notes you have nothing to engineer with. Her six note categories are the most useful artifact here — the three everyone already knows (role, goal and audience, constraints on style and format) plus the three that actually separate good output from generic output: **inputs** (what data the model may use, and what it must not — she scopes hers as the whole vault minus certain tags), **source of truth** (which documents outrank the model's generic knowledge), and **judgment** (how it decides what is good). The framework-distillation trick is the single most actionable thing in the video and takes ten minutes: hand the model ten examples you think are excellent and ten you think are terrible, and ask it to articulate what makes the good ones good and what the bad ones share. You get an explicit framework out of tacit taste, using the one thing LLMs are reliably strong at. Her handling of the "does AI make you stupid" question is also unusually fair — she cites research in both directions rather than only the flattering half, and locates the difference in whether you have externalized context to think *with*. The live A/B is real evidence: the same career question run in an incognito session returns role lists she glazes over, while Claude Code over her vault returns *"creativity and structured thinking are not opposites. They are partners"* — which she says articulated her own position better than she had. What the video never does is check the outputs it finds most impressive. Twice she notes that the model surfaced something she *"never really mentioned"* in her notes and reads that as insight; genuine synthesis and confident confabulation look identical from the inside, and the difference is checkable against the notes. - **Next steps for viewers.** 1. Write the three context-engineering categories most people skip — **inputs**, **source of truth**, and **judgment**. The prompt-engineering three are already handled; these are where output quality changes. 2. Run the distillation trick this week: ten good examples, ten bad, ask for the framework underneath. That converts taste you cannot articulate into something an agent can apply. 3. Scope your inputs explicitly, including exclusions. Her *"use everything in my vault except for these tags"* is the pattern — decide what is off-limits before you point an agent at a personal corpus, not after. 4. When the model surfaces something you never wrote down, treat it as a hypothesis rather than a finding. Search your own notes for it. If it is there, it is synthesis; if it is not, you have learned something about the model rather than about yourself. 5. Take the "sand castles" test seriously — if your ideas do not compound because nothing is written down, the note system is the prerequisite, not the optimization. 6. Note the token cost she reports in passing: a single vault-wide question consumed **39,000 tokens**. Vault-scale context is not free, and the knowledge-map pattern from her later second-brain video exists partly to address that. ## 🔄 Refactor Appendix — Why you should take notes if you use AI ### 📚 Literature Atoms (H3) #### [H1] Context engineering needs six categories, and three are usually missing > [!quote] ID:rw_ftj_h1 > "the next three things are more covered in context engineering and also notes that we want to have which are about the inputs the source of truth as well as judgment" **Context & use**: - ~04:01 — the three prompt-engineering categories (role, goal and audience, constraints) plus the three that matter more. - **Inputs** = what data is in scope, including exclusions. **Source of truth** = what outranks generic knowledge. **Judgment** = how good and bad get decided. The most useful checklist in the video. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=FtJYJF8QodM&t=241)  #### [H2] Distill a framework from paired examples of good and bad > [!quote] ID:rw_ftj_h2 > "if you have 10 examples of things that you think are great and 10 examples of things that you think are terrible, get the LLM to create that framework for you" **Context & use**: - ~07:17 — the most actionable technique in the video, and it takes about ten minutes. - Converts tacit taste into an explicit framework using the thing models are genuinely reliable at: distilling patterns from supplied material. The output is reusable *consciously* next time. **Tags**: #xref #articleseed #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=FtJYJF8QodM&t=437)  #### [H3] Whether AI dulls thinking depends on whether you externalized context first > [!quote] ID:rw_ftj_h3 > "what is holding us back from being more engaged in the thinking process is have we downloaded the context out from our brain and into the LLM" **Context & use**: - ~08:02 — she cites research in both directions: outsourcing thinking degrades output and engagement, *and* deeper engagement is possible depending on use. - Unusually fair treatment of a question most creators answer only in the flattering direction. Her proposed mechanism is specific and testable. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=FtJYJF8QodM&t=482)  #### [H4] Undocumented ideas do not compound > [!quote] ID:rw_ftj_h4 > "if you are working on something and you want to build on top of it, you want the ideas to compound and not just build sand castles then you have to document it" **Context & use**: - ~25:17 — the closing argument, and the answer to "why bother taking notes at all." - The externalization is what makes an idea available to build on later, by you or by an agent. Explicit beats remembered. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=FtJYJF8QodM&t=1517)  ### 🧠 Permanent Candidates (H4) #### [P1] Extract your framework from examples you have already judged **Claim** (one sentence, Courtney's voice): You often cannot state your quality criteria directly, but you can recognize good and bad instances, so hand over the instances and have the criteria distilled from them. **Evidence pointer**: [[#H2]] **Scope & conditions**: Holds wherever judgment is real but tacit — writing quality, code review standards, design taste, hiring signals. Requires enough already-judged examples on both sides; ten and ten is a workable floor. Fails where your judgments are inconsistent, in which case the distillation surfaces the inconsistency, which is itself useful. The output needs review before use: a distilled framework is a hypothesis about your own taste, not a transcript of it. **Links** (≥1 with annotation): - [[Tacit knowledge]] — because extends - [[Encode your judgment as frameworks]] — because extends the same taste-into-artifact move - [[Rubric design]] — because applies-in evaluation  #### [P2] Output that was never in your notes is a hypothesis, not a discovery **Claim** (one sentence, Courtney's voice): When a model surfaces something you never wrote down, genuine synthesis and confident confabulation are indistinguishable from the inside, so check before you believe it. **Evidence pointer**: [[#H3]] **Evidence**: > [!quote] ID:rw_ftj_p2 > "I never really mentioned this in my notes, so it's kind of crazy that it articulated this" **Scope & conditions**: Holds for any synthesis over a corpus you own — vault analysis, codebase summarization, meeting-note rollups. Sharpest when the output flatters you or matches what you hoped to find, which is exactly when checking feels unnecessary. Verification is usually cheap: search the corpus for the claim. If it is there, the model synthesized; if it is not, you have learned about the model, not about yourself. **Links** (≥1 with annotation): - [[Plausible and interesting is not the same as true]] — because extends - [[Confirmation loop]] — because extends the same unverified-acceptance failure - [[Grounding must survive every handoff]] — because applies-in verifying attributed claims `proof_state: working`  ### ✏️ Fleeting Queue - [F1] Run the 10-good/10-bad distillation on my own writing samples and see what framework comes out. Cheapest high-value item in this batch. #fleeting #inbox - [F2] Write the three missing context categories for my agents — inputs (with exclusions), source of truth, judgment. #fleeting #inbox - [F3] Scope agent vault access explicitly: which tags or folders are off-limits? Decide before granting, not after. #fleeting #inbox - [F4] 39,000 tokens for one vault-wide question — measure what my own vault costs per query before making this a habit. #fleeting #inbox - [F5] Find the actual studies she alludes to on both sides of the "does AI dull thinking" question. She cites them vaguely and the balance is worth sourcing properly. #fleeting #inbox - [F6] Post idea: "synthesis or confabulation?" — when a model tells you something about yourself that isn't in your notes, how do you tell? #fleeting #articleseed ### → Things 3 (max 2 actions) - [ ] Run the 10-good/10-bad framework distillation on my writing and save the result as a framework file — writing #things - [ ] Define inputs, source-of-truth, and judgment files for my agent context folder — tooling #things --- ## Full Transcript [00:00] If you use AI and you don't have a [00:03] note-taking system, you have to start [00:05] one now. And the reason lies in the [00:09] clear trend that we're moving away from [00:11] prompt engineering and the increasing [00:14] importance of context engineering and [00:16] having used these tools for two plus [00:18] years I can really say that having the [00:22] right context having the right notes to [00:24] be able to set up what you actually want [00:27] and what you want these LLMs whether [00:29] it's chitty claude Gemini whoever to be [00:33] able to draw upon the quality quality of [00:36] the output is going to be different as [00:39] well as the quality of thinking you can [00:41] put into it. Everyone's complaining that [00:44] you know we have to outsource thinking. [00:46] It doesn't have to be that way. But if [00:48] you don't have notes then yeah there's [00:51] no choice. I've talked about this in [00:53] another video that you can explore here [00:55] which is it is so important to turn [00:57] tacit knowledge knowledge that just [00:59] exists in our head and we cannot [01:01] articulate into explicit knowledge so [01:04] that we can share with other people [01:05] which is the source of why note [01:08] takingaking came into place right before [01:11] level one was just us in our brain we [01:13] have things swirling in it and because [01:16] we kind of know what they are we say [01:18] okay that's fine right don't need to [01:20] write anything now until the second [01:23] iteration we developed writing systems [01:25] and we realized okay there are things [01:27] that we need to unload in order for her [01:30] brains to do more. So that's why we [01:33] started to translate this tacet [01:35] knowledge into explicit knowledge [01:36] because then we can work with other [01:38] people right so as we're talking about [01:40] on this channel it really is very [01:43] important to recognize that [01:45] communication is not about performance [01:48] of course for for some people who are I [01:50] don't know motivational speakers or [01:52] whatever I guess the performance side is [01:54] also important but for people who are [01:56] working on ideas knowledge workers right [02:00] why we need to communicate clearly is [02:02] not to look good but to get the tacid [02:05] knowledge into explicit knowledge and to [02:07] share with other people so they can [02:09] understand and we can take action [02:11] together with as little friction as [02:13] possible right and vice versa when [02:16] they're saying something we can process [02:18] it and understand it which is why [02:20] frameworks are really important but that [02:23] is the core of communication and now [02:25] we're adding on another layer of it's [02:27] not just us and other people but there's [02:29] also an AI that has access to other [02:32] kinds of information so we can put it [02:34] together all move forward and having a [02:38] note takingaking system that has the [02:40] important information that the LLMs can [02:43] use makes everything better. it then you [02:47] can use it to work on the things that [02:50] take you a lot of time. And I'll show [02:51] you an example in just a moment. But [02:53] just want to [02:56] have a note-taking system and have [02:58] information downloaded from your brain [03:01] onto something that LLM can read and [03:04] then use is superpower to be able to [03:06] summarize, to be able to spot patterns, [03:09] all of those good stuff. Okay, so that's [03:12] part number one. Now what do we actually [03:15] put into the notes then? What kind of [03:17] notes do we need to have? Well, luckily [03:19] there are just six categories we have to [03:22] think about and I'm drawing straight [03:24] from context engineering because uh to [03:28] be able to communicate with AI, we might [03:31] as well use its rules and its way of [03:34] thinking. So we mentioned three already [03:37] that mostly are covered in prompt [03:39] engineering which is we have to give it [03:41] a role and then give it some general [03:45] idea about the goal and the audience uh [03:49] that is going to use that output. Then [03:51] also we want to give it some constraints [03:53] around style or format and these kinds [03:56] of things. Most people know this part [03:57] already. Now the next three things are [04:01] more covered in context engineering and [04:04] also notes that we want to have which [04:06] are about the inputs the source of truth [04:10] as well as judgment. Okay. So by input I [04:14] mean for the LLM you know what's the [04:18] data just in general that it can use. So [04:20] for example for me everything in my [04:23] vault in my obsidian vault it can use or [04:26] maybe it's okay maybe there are some [04:28] personal things or sensitive things or [04:30] confidential things I don't want it to [04:32] use then I can also say okay use [04:34] everything in my vault except for these [04:36] tags right you want to be able to tell [04:38] it explicitly what's the input it can [04:42] use number two is the source of truth so [04:45] what kind of documents outrank it [04:48] generic information Right? So for [04:50] example, when you have a do not use, it [04:53] means yes, you have these things in your [04:56] general information, but I I don't want [04:58] you to use them. That's a source of [05:00] truth. So that it knows, okay, I'm not [05:02] going to rely on those. I'm going to f [05:04] focus specifically on these um these [05:07] documents. And the third one is [05:10] judgment. So how does LLM judge what is [05:15] good, what is bad? Right? This is [05:19] especially where we need to take task of [05:21] knowledge into explicit knowledge and [05:23] luckily there are frameworks to do this [05:25] and frameworks help us better articulate [05:27] what's good and what's bad. So that's [05:30] the type of notes that we're going to [05:32] take. A quick note on frameworks. I [05:34] think if you've used AI for a while then [05:37] you probably noticed it has several [05:40] frameworks that it keeps on using right [05:42] for example for formatting it uses the [05:45] heading options. So that as it goes [05:48] through it organizes things by headings [05:51] or it likes to say a lot like it is not [05:55] this, it is that. Right? A great little [05:57] via negativa there. But the problem with [06:00] its frameworks is it keeps on using the [06:03] same ones and it keeps on using very [06:05] generic ones when it produces output. [06:10] But to be able for you to judge what's [06:12] good or what's bad, a lot of the time a [06:15] simple way of doing that is to say, [06:17] "Okay, write me a sales letter, right?" [06:19] And mostly the [06:22] LLM might use ADA, like something that's [06:25] very basic in copywriting, but if you [06:28] have a different framework you would [06:29] like it to focus on, then just share [06:32] that one, right? And this way you are [06:35] inherently building what's good into uh [06:39] your prompt and your context so that you [06:42] don't have to worry about how does a [06:43] judge I don't know like intuitively I [06:46] look at it as pretty good right instead [06:49] what you can do is give it a framework [06:51] or if really you don't have any [06:53] frameworks but you have examples of [06:56] what's good and what's bad it would be [06:58] great to have you know the good and the [07:00] bad so that you can get your LLM to [07:04] figure out what is the what is the [07:06] framework underneath. And the good thing [07:08] about LLMs is they're great at [07:11] distilling, right? As long as you give [07:13] it information, it's pretty good at [07:15] distilling. So, what you can do is if [07:17] you have 10 examples of things that you [07:18] think are great and 10 examples of [07:20] things that you think are terrible, get [07:22] the LLM to create that framework for [07:25] you, right? help me articulate what [07:28] makes these examples great and what are [07:31] the patterns of the bad ones. Right? [07:33] Then you've got your own framework that [07:34] we you can use next time consciously. So [07:38] that is a great way of injecting [07:40] framework in a way that helps you get [07:42] the outcome that you want. All right, [07:44] I'll give you an example from my actual [07:46] obsidian vaults so that you can see how [07:49] not only the quality of the output [07:51] changes uh also the quality of the [07:54] thinking changes and I really believe [07:58] that yes working with AI can absolutely [08:02] make us dumb and studies have shown [08:04] right if we just outsource our thinking [08:07] then yes as we go on quality of the [08:10] output drops also we are disengaged from [08:13] the thinking experience. But that's not [08:15] the only way to use LLM. And studies [08:18] have also found depending on how you use [08:22] it, right, you can deeply engage and [08:24] actually think deeper and access things [08:26] that you didn't have the mental capacity [08:29] to do before. And really what is holding [08:32] us back from being more engaged in the [08:35] thinking process is have we downloaded [08:38] the context out from our brain and into [08:41] the LLM. So when we have a conversation, [08:43] it's intelligent one. It's one that's [08:45] based on the information that we wanted [08:48] to work with, right? So if you feel [08:50] like, okay, I'm just getting dumber by [08:52] the day using CHBT or whoever, then you [08:56] have to also ask yourself, am I also [09:00] just consuming and not processing it, [09:03] outputting documents that I can feed my [09:06] LLM? Because if you're not doing that [09:08] process right then there is not much [09:11] context to share right because it's [09:13] difficult you not every single time [09:15] you're there you want to start typing [09:17] out the context it's very timeconuming [09:20] there's a lot of friction which is why [09:22] taking notes as a habit that help you [09:24] just drag and drop the context into the [09:27] LLM and your conversation with it right [09:29] so if you feel like okay this is what [09:32] I'm missing then let me just show you [09:34] how significantly things can change. For [09:38] the example, I'm going to use a problem [09:40] a lot of us on this channel face, which [09:42] is I'm multi-passionate. I'm interested [09:44] in a bunch of stuff, but I feel so [09:47] scattered all the time. All I know is I [09:49] don't want that singular track career [09:52] or, you know, life goals that other [09:54] people have. But at the same time, I [09:56] don't know how do I focus in order to [09:58] get to where I want to go. Like, there [10:01] must be something that can tie [10:02] everything together. So let's talk about [10:05] that and let me show you the difference [10:08] between talking to I'm going to use [10:10] TAGBT here in incognito versus using [10:16] your own context. So I'm going to share [10:18] my obsidian vault in a moment but let's [10:22] give this a try. Okay. So what we can do [10:25] with chai is I'm interested in [10:29] creativity and frameworks. [10:33] So the choosing contra contradicttory [10:39] help me see how to [10:45] combine these two interests together [10:49] into a career. That's that makes sense. [10:53] Let's see what it says. I haven't really [10:55] shared anything, [10:58] but let's see what it says. [11:01] Uh, blah blah blah blah blah blah. Where [11:06] is the [11:10] design thinking, [11:12] product strategy? Okay. Creative [11:16] direction with systems, [11:19] creative director, brand strategist. I [11:21] mean, I guess, but you know, for [11:24] example, the career I'm thinking about [11:26] is not like these roles, but of course, [11:30] I didn't tell it. So, okay, makes sense. [11:32] It's saying something like this. [11:35] Yeah. Okay. So, I'm kind of in this [11:38] category, the framework builder, thought [11:40] leader kind of [11:43] uh learning design. I mean, these are [11:44] things I'm interested in. Um, but just [11:48] seeing this list doesn't really help me [11:51] say, okay, so how do I know which ones [11:53] for me? And I guess they all sound good, [11:55] but how do I prioritize them? [11:59] Uh, let me ask it. How do I prioritize [12:03] this? [12:07] The emoji. [12:09] Okay. Rule of thumb. Prioritize the [12:12] intersection, not the options. Okay, I [12:13] like that. [12:18] Does this compound? I don't know. Like [12:20] already I feel not so energized looking [12:24] at [12:26] this list. I mean, I get it. It is very [12:30] generic like this is best practice [12:34] so to speak but at the same time you [12:36] know is it could also just be I glaze [12:38] over everything. So you know just as a [12:41] demonstration I think it's going to give [12:44] some good ideas. I mean in the end I do [12:48] mostly [12:50] three and four [12:53] and a little bit of five as well but um [12:59] not so easy to get from this list to [13:01] finding what I want to do. Okay so that [13:04] is the you know lack of context example. [13:09] Now we can see with claw code what does [13:12] it look like when I actually share my [13:16] thoughts unfiltered straight in my [13:18] Obsidian vault. Let me show you. All [13:21] right. I'm using clawed code here. [13:26] It looks pretty intimidating, but it's [13:29] actually not that difficult to set up. [13:31] Anyway, um I can share the instruction [13:36] with you maybe in a different video, but [13:39] the best place is actually Claude has a [13:41] website. I'll link it right here that [13:44] you can go there and it's quite easy to [13:45] set up. All right. So, what I'm going to [13:48] do is go go into my obsidian vaults [13:55] and help me understand what exactly am I [14:00] interested in with creativity [14:04] and framework thinking. [14:07] Help me understand [14:10] what is that intersection. [14:13] Okay, I'm going to find the length. One [14:16] second. [14:20] See, it'll take some time. So, [14:23] >> a few moments later, [14:25] >> what does a find? I am fundamentally [14:28] interested in building the case that [14:30] creativity and structured thinking are [14:31] not opposites. They are partners. My [14:35] notes consistently reject the false [14:37] dichotomy of logical and creative [14:38] thinkers. Okay, it just articulated that [14:43] better than me cuz I was thinking about [14:44] this. I was like, okay, yeah, I know [14:45] it's a paradox that I'm interested in, [14:48] but actually I've said this many times [14:50] in my notes and the key ideas or [14:52] constraints enable creativity and that [14:55] creativity is a refined refined mental [14:57] models and I really believe it's a proxy [15:00] for intelligence. So, it's captured that [15:03] word note. Uh framework thinking is [15:07] creative. I haven't really talked about [15:09] this on the channel that much because I [15:11] feel like oh you know are people [15:13] interested is in this. I really think [15:16] creativity is at the top of Maslo [15:18] hierarchy but I know a lot of us you [15:22] know analytical people we're like [15:25] creativity who's that for I think we're [15:28] not quite understanding the word and I [15:30] have a problem with the word itself but [15:32] you know it's really the polymathy [15:35] connection right creative thinkers are [15:39] often highly logical right like Einstein [15:42] was a great violinist [15:45] They did some study where most of the [15:47] Nobel laureates had a very creative part [15:50] of their life that they delic dedicate a [15:52] good chunk of their time and energy to [15:55] and cross-disciplinary thinking I really [15:58] is a core at what I do. So all right [16:02] what I seem to be building towards is a [16:05] unified view where frameworks are the [16:07] tools that break down complexity [16:09] creativity reassembling pieces into new [16:12] holes is interesting. I [16:15] interesting. I like this analysis and [16:18] synthesis contrast. I didn't put that [16:22] thought in anywhere explicitly in my [16:26] vault. So, just having it go through, [16:29] well, I went through 39,000 tokens, but [16:33] having it going through all my notes and [16:36] being able to just extract this is gold, [16:39] right? My incognito conversation is [16:42] never going to get me to this level. [16:45] Uh would I like to dig deeper into any [16:48] specific notes? So, um can you list down [16:53] what parts of creativity [16:57] I'm actually interested in? If I had a [17:02] research uh question of what would that [17:06] be? Give me a few options. [17:10] Let's see what it'll say. I'll let it [17:12] think for a bit. [17:13] >> A few moments later. [17:15] >> Okay, here we go. Uh, based on my notes, [17:19] logic, creativity, unity, rea. Yeah. So, [17:23] you know, rejecting the left brain, [17:25] right brain. Well, I mean, I don't [17:26] reject it, but I do think that both work [17:31] together, and we all have a left brain [17:34] and right brain. Um, [17:37] okay. Constraint as creative fuel [17:42] structure input. I feel like this is [17:43] already pretty overdone. So [17:46] that's probably not yet I think is the [17:50] the so what of that and being able to [17:53] explain this for those of us who are [17:55] analytical and to understand that okay [17:57] constraint the the the so what of having [18:00] constraint is then we can be creative. [18:02] Uh, okay. [18:06] Intense learning, refining models and [18:08] creative output. Oh, interesting. Again, [18:10] I did not put this specifically in, but [18:14] now that I see it articulated this way, [18:17] it's giving me some ideas. This is, you [18:20] know, this is the context that can get [18:23] you what you are working on. Like, I'm [18:27] like a kid in the candy store. Even [18:29] though I these are all my notes, but I [18:32] wasn't going to go through each note one [18:34] by one and come up with this, right? [18:35] What I can say is, okay, here's [18:37] everything I'm interested in that I've [18:38] documented in my own words. Now, can you [18:41] summarize it, synthesize it for me? And [18:44] that's a LLM strength. So, [18:49] this is I'm I'm just a very happy girl [18:51] right now. Okay. Cross-disciplinary [18:53] thinking. I think I will what I will do [18:55] next is push it further. Okay. Actually [18:58] I'll do it with you here. Okay. So based [19:00] on these, [19:02] can you suggest [19:05] some sources resources [19:10] to go deeper on [19:15] books [19:17] uh or author's schools [19:21] of thought so that I can build this out? [19:26] also consider my other interests [19:32] um [19:34] put it in my vault so that [19:43] I'm exploring the intersections of my [19:48] interests. Okay, let's see how it goes. [19:51] >> A few moments later. All right. So, five [19:54] minutes later, here is what it shared. I [19:58] already recognize [20:00] a lot of [20:04] these books, [20:07] but [20:13] uh I mean, it's good to know that [20:15] they're aligned. I wonder if it saw my [20:18] book list. I assume it did, but [20:23] I like that it broke down some of the [20:26] authors to follow and you know I already [20:28] follow them in this case. [20:37] Okay. [20:39] Process philosophy. Oh, I didn't know [20:41] Whitehead. I do have Whitehead's book on [20:44] creativity, but [20:46] interesting process philosophy. I have [20:48] to look into it. Um [20:51] I feel like the recommendations now kind [20:54] of um you know it's averaging down to [20:58] what's very common. So I have to [21:00] probably push a little bit here more. [21:04] Maybe in claw code maybe I will go to [21:07] elicit and look at more academic [21:10] research or use research mode and get [21:14] some of these as well. But uh here is [21:17] interesting. [21:24] But these these are good. I mean [21:27] definitely know some of these [21:29] the recommendations. I've read them. But [21:33] um all right that's good to know. I [21:37] would now come back to the question of [21:40] the career. Right. Okay. So knowing [21:44] you know about my interests and [21:48] questions for creativity and frameworks. [21:54] Help me uh design my career [21:59] path. I want something that is [22:03] not that is practical [22:07] and [22:09] pushes forward [22:12] our understanding of how to do knowledge [22:18] work meaningfully [22:22] and get paid for it. [22:25] Okay, let's say something like this and [22:28] see what kind of recommendation it would [22:30] come up with. [22:32] >> A few moments later. [22:34] >> Okay, here we go. Um, it's mentioned. [22:36] Okay, so first of all, the positioning. [22:39] It knows what I'm doing. [22:43] I'm glad this uh showed up as NA. Great. [22:47] Um, the gap you could fill productivity [22:51] influencer. [22:52] There is a missing voice in the [22:54] knowledge work conversation tactics [22:56] without theory. Academic theory without [22:59] practice. [23:00] Yeah, these are great. The gap is [23:02] someone who can articulate how knowledge [23:04] work actually works [23:07] and translate it into teachable [23:09] applicable methods. Honestly, this is [23:12] how I feel. I mean, I never really [23:15] mentioned this in my notes, so it's kind [23:18] of crazy that uh it articulated this [23:27] building intellectual property for [23:29] writing. I am writing a book currently. [23:32] I don't know about the book to speak [23:34] into corporate workshop to advisory type [23:37] revenue stream. Yes, I know these are [23:40] awesome people do it. Okay. [23:44] The educator builder. [23:46] Uh, okay. [23:50] Okay. [23:54] The applied researcher work inside AI [23:57] companies. I don't really want to do [23:59] that. Hybrid phase one. Continue content [24:03] but shift towards original frameworks [24:05] not just explaining others. Then [24:09] productize. I don't know. I think this [24:12] framework is very boring and actually I [24:16] need to find some good examples of [24:19] people I find interesting and actually [24:21] let me do that. I find [24:25] okay before I continue I just realized [24:28] now you're just indulging me in my me [24:32] figuring out my future plans. So, if [24:36] you're not interested in that, honestly, [24:38] you can stop right here because uh I'm [24:41] just going to continue doing this uh for [24:44] the next little bit, but the point is [24:46] bring me back for having a note-taking [24:48] system, right? What really continue to [24:51] surprise me is how good LLMs are at [24:54] taking information and being able to [24:56] pull out some really great ideas and [24:59] being able to identify, especially with [25:01] connected notes, the core notes that [25:04] matter to you. So, what it's able to do [25:08] is without me needing to re review all [25:10] of my notes, it's able to pick out [25:12] things like this. So if you are working [25:17] on something and you want to build on [25:19] top of it, you want the ideas to [25:21] compound and not just build sand castles [25:24] then you have to document it. It has to [25:27] be explicit so that AI can work with it [25:30] and help you see the things that you as [25:32] a human being have you know natural [25:35] weaknesses like digesting a lot of [25:38] information in a short amount of time, [25:40] right? So being able to engage in that [25:42] and have your curiosity and your [25:45] questioning mind and your judgment mind [25:47] and your editorial mind come instead of [25:50] just say hey you know give me something [25:53] and then now oh this is not that great [25:54] right so let me move on and have that [25:56] task sit in its current state is going [26:00] to be so much more different than if you [26:02] had these context so you can go deeper. [26:04] So with that in I hope you enjoyed this [26:07] video. If you have any questions, let me [26:09] know down below. I will link um the [26:12] cloud code setup information down there [26:15] as well. All right, I'll see you in the [26:17] next video.