Nick Milo Reads My Obsidian Vault - CourtneyR.dev's 2nd Brain

Nick Milo Reads My Obsidian Vault ![rw-book-cover](https://i.ytimg.com/vi/bdjYo-x4PUw/maxresdefault.jpg) 🔗 **Source:** [Nick Milo Reads My Obsidian Vault](https://www.youtube.com/watch?v=bdjYo-x4PUw) > [!note]+ Document Notes > tags: [obsidian, personal-knowledge-management, maps-of-content, ai-agent-workflows, second-brain-methods, linking-your-thinking] > author: Tiago Forte > source: Tiago Forte > reference: https://www.youtube.com/watch?v=bdjYo-x4PUw > date created: 2026-09-10 > > > [!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 Do not quote or attribute any other line to the named author.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: Nick Milo Reads My Obsidian Vault (The Full Session) >  ## 🧭 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) - [🔄 Refactor Appendix](#-refactor-appendix--nick-milo-reads-my-obsidian-vault-the-full-session) ### 🔗 References - [🗒️ Notes — full timestamped walkthrough](#️-notes) - [✅ Action Items — org-mapping and applicability](#-action-items) - [🔄 Refactor Appendix — atoms and permanent candidates](#-refactor-appendix--nick-milo-reads-my-obsidian-vault-the-full-session) ## 📝 Summary - **What happened.** Tiago Forte (*Building a Second Brain*, the PARA method) invites Nick Milo (Linking Your Thinking, maps of content) onto the channel and hands over the screen: Milo does a cold, unrehearsed read of Forte's live Obsidian vault — theme, folder sidebar, graph view, note count — with no prep and no prior look. From there the two spend the session pointing Claude (Co-work) at the vault to build a project dossier, a self-generated map of content, and a gap analysis against Forte's actual marketing plan for his forthcoming book, while narrating the governance rules Milo uses to keep AI from touching his canonical notes. - **Why it matters.** Milo and Forte represent the two dominant, competing schools of personal knowledge management — link-driven emergence (Milo: "ideas emerge from linking") versus folder-driven actionability (Forte: PARA, progressive summarization) — and this is a live compatibility test of the link-driven diagnostic method against a vault that, by Forte's own account, doesn't use PARA at all (t=120). The claims that surface (link count as an importance signal, a numeric vault-size threshold for AI fidelity, a quarantine folder for AI writes) are concrete enough to test against a different vault's actual shape rather than take on faith. - **Key stakeholders.** Solo knowledge workers and authors running a large personal vault against an AI agent with real stakes attached — Forte's vault here is driving marketing for a book already in presales (t=636). Both hosts also sell commercial frameworks built on the claims they're demonstrating (Linking Your Thinking; Building a Second Brain), which is relevant context for how confidently either claim should be taken. - **Risks and tradeoffs.** This is a single vault, a single session, two collaborators who never push back hard on each other, and at least one moment where the agent visibly overindexes on a thin, incomplete note (a short list of "email topics I wanted to write" mistaken for the full plan, t=3333–3391) that Forte has to catch by discernment, not by any system check. None of the method's numeric claims (link-count-as-gravity, the ~500-note fidelity ceiling) are tested against a second vault or a baseline in the video. - **Bottom line.** The durable takeaways are the concrete, testable diagnostics — not the MOC-vs-PARA framing, which the video never actually resolves (Forte's own working vault sidesteps PARA entirely). Worth pressure-testing each diagnostic against a vault built from both traditions at once, at meaningfully larger scale than the ~1,662-note vault shown here. ## 📖 Main Content Forte opens by asking Milo to read his live vault cold — "like a doctor," Forte says, "what is your diagnosis" (t=82). Milo's stated method, narrated as he goes: check the theme first (dark mode, "default," which Milo calls out as unremarkable), then the folder sidebar for a "strong folder structure," then whether an open note has links, because "that would tell me a lot about how the person is linking their thinking" (t=138). Milo immediately notices the vault has no PARA folders — "you don't have parah in this one because it's not like your go-to vault" (t=120) — a detail he treats as informative context rather than a defect, and a detail worth returning to, since PARA is Forte's own framework and this is his vault. Graph view gets a mixed verdict: "they are less helpful than — I mean they're great eye candy but they're not terribly helpful" (t=152), but Milo still reads real structure out of it — one dominant project cluster (the book, *Life in Perspective*), a map-of-content node with many things linking into it, a readwise cluster, and faded "placeholder notes" that already carry backlinks before anyone has written them (t=201–234). Milo frames this last feature — a note accruing links before it exists — as one of Obsidian's most useful properties: by the time you actually create the note, "it could already have like multiple links... you're never starting from a blank page" (t=211). Scale comes up next: Forte's vault has 1,662 notes across four folders; Milo's main vault has roughly 17,000, built almost entirely by hand rather than imported (t=251–276). Milo is explicit that the comparison doesn't matter on its own — "it's fine if you have 20 notes" — what matters is what the notes and links tell an AI (or a person) about what's meaningful. This scale question returns later as a specific numeric claim: past roughly 500 notes, depending on note length, AI reading a vault directly stops working with "100% fidelity" — it's sampling, not reading, and "you might get poor results" without a distillation layer like a map of content (t=1632–1654). Before any AI work starts, Milo states the session's one hard rule: "I don't want AI to touch my notes, my original canonical notes directly" (t=376), because the vault represents "hundreds and hundreds of hours" of manual linking and outlining that an automated import didn't produce. His mitigation isn't access restriction — it's backups: always duplicate the vault before pointing an AI at it, because "if AI can read something, you have to assume... it may be able to write" (t=485–495), and Milo pairs on-device duplicates (a dated Cmd+D copy) with off-device backups (Google Drive, an external drive). He grounds this in what Obsidian actually is — not a proprietary app holding your notes, but a skin over a folder of plain-text markdown files (t=421–454) — which is also why any AI tool can simply be pointed at that folder with no export step, and why leaving Obsidian later costs nothing (t=1100–1118). On connecting AI to the vault, Milo names three options — pointing an agent (Claude Co-work, ChatGPT) straight at a folder, third-party community plugins (Copilot, Smart Connections), or the Obsidian CLI — and states a clear preference for the first: plugins mean depending on outside developers for "all these extra breaking points," and the CLI is powerful but not worth the setup cost without a specific reason (t=698–741, t=1815–1832). Folder-pointing also gives granular privacy control without any new mechanism: "you don't have to point it even at your main Obsidian vault, you could say only the notes that I've manually put into the subfolder" — which is why, Milo says bluntly, "folders are back, baby" (t=875–902), a notable admission for someone whose framework is built around links rather than folders. Before any diagnostic prompt, Milo has the agent read three specific files every session — a "me" file, a vault-map file, and a skill-map file (t=938–987) — which he generalizes into a three-layer model he calls AIOS: the core is the vault itself ("second brain, a knowledge system... everyone has that... before AI was around"); the outer layer is whichever AI tool is currently in fashion ("the least important... the most ephemeral"); and the middle is a translation layer — the three orientation files — that lets any outer-layer tool be swapped without loss (t=1026–1118). Milo frames this as inverting the default mental model most people hold, where the AI tool sits at the center and your data is peripheral context it enriches; here the vault is the center and the AI is interchangeable. He names the stakes explicitly as "digital sovereignty" against "digital feudalism... doing all of our work on somebody else's land" (t=1186–1196), and attributes the opposite default to platform incentives, not user preference: "their incentive is to lock you in if they can do that" (t=1174). Forte then surfaces the mechanism behind the claim that his one existing note with "50 plus links" carries special weight: Milo compares it directly to PageRank — "instead of having thousands, millions of humans over there cataloging, let's infer meaning from the graph, from the connections" (t=1364) — arguing that link density is how a vault tells an AI (or anyone) what's meaningful without needing explicit organization, and that not every vault needs that complexity: "it depends on what the prompt is, like what problem that is trying to be solved" (t=1394). The diagnostic and dossier sequence follows: Milo's go-to opening prompt for an unfamiliar vault is to ask the agent to build a "dossier," which the two then narrow from "write a dossier about me" to a project-scoped version — "read through this entire Obsidian Vault... write a project dossier that summarizes everything that someone might want to know about the project... which was my book, *Life in Perspective*" (t=1441–1488). While the agent runs (several minutes), both hosts describe what they actually do during the wait — Forte checks email and drifts to YouTube; Milo watches the progress indicator and asks himself meta-questions ("is this the right approach... is this a good use of my time," t=1563–1573) — and Milo notes he rarely runs more than about three sessions in parallel (t=1533). Milo then demonstrates a more targeted alternative to pointing an agent at a whole folder: copying the exact system file path to a single core note (often a map of content) and telling the agent to "start with this note... help build a brief around this," which he says is worth the extra step "if I want absolute certainty," to avoid wasted tokens and wasted uncertainty about scope (t=1690–1761). This is also where he introduces the "Sherpa system" — a pair of stacked skills where he first trains the agent on his 15–20 favorite existing maps of content so it can "reverse engineer what makes a good map of content" into a style guide, then runs a "questionnaire skill" against that style guide whenever he starts researching a new topic (t=1862–1935). A conceptual distinction follows that both hosts treat as central: a map of content is not a table of contents. Milo: "I think a map of content is fluid... and an index feels a bit more static" (t=1973); its value is in the act of building it — "organization in this way is ideation" (captioned "ideiation," t=2028) — not in the artifact itself. He's explicit that an AI-built map of content only has "potential meaning" until a person actually works with it: "it's not helpful until you actually have the energy to make it your own" (t=1948). Back in the vault, once the dossier is generated, Milo's first move is governance, not content review: everything the AI creates goes into a dedicated `AIOS/History` folder, timestamped, with "a very clear line of separation between my notes, my thoughts, and then where AI's notes and thoughts go" (t=2202–2261), and he has Forte turn this into a repeatable skill on the spot rather than a one-off copy-paste (t=2354–2398). Only after that separation is established do they ask the agent to backlink the new dossier into the existing vault and, in the same prompt, build a supporting map of content "using all of the advice and recommendations issued by Nick Milo" — a request Milo jokes about ("who's that? that'd be insulting," t=2554) but does not push back on as a method. The session's most substantive test comes when Forte points the agent at his actual marketing and launch plan and asks it to find gaps against the raw vault material — "is there anything I'm missing... that I'm undervaluing or not taking into account" (t=3122–3150). This triggers Forte's sharpest critique of the whole workflow: every AI output he's seen is a summary of a summary, and "the more times you summarize, all the specificity and the signal gets stripped out of it" (t=3193), which makes the artifact a "depreciating asset" unless the agent is explicitly told to cite sources and use verbatim quotes rather than paraphrase (t=3213–3263). Milo's parallel governance rule is the one line he calls out as his only hard callout box: "do not rewrite me" — because when he later asks the agent what he'd previously written, a paraphrase gets returned as if it were his own words (t=3273–3289). The gap-analysis output is genuinely useful but visibly flawed in one place: it overweights a short, incomplete list of "email topics I wanted to write" as though it were the complete plan, because — as Forte diagnoses in real time — "AI always assumes that whatever information it has access to is all the information that exists... it's like a little universe, a closed universe that you've created" (t=3391–3412). Both hosts frame the fix as vigilance rather than a system safeguard: "you bring your own discernment. This is not truth. This is not reality" (t=3353). The session closes on the scattered-source problem this exposes — Forte's information lives across eight to ten platforms despite his stated preference for one source of truth (t=3524–3546) — and on Milo's closing pitch: "we need to teach AI to learn about us" by pointing it at owned notes plus a translation layer, so the underlying knowledge outlives whichever AI tool is fashionable that month (t=3620–3630). ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=0s) Framing the session - Forte invites Milo to do a live, unrehearsed read of his real Obsidian vault, then walk through AI-Obsidian experiments together. - Explicit framing up front: people want to use AI and also "own and control and know where their data is" (t=18) — the video's stated tension. #### ⏱️ [00:01:22](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=82s) Milo's cold-read diagnostic method - Forte's framing: "like a doctor... what is your diagnosis" (t=82). - Milo's stated order of operations: theme first (notes it's the plain dark default, calls it unremarkable), then the folder sidebar for a "strong folder structure," then whether an open note shows links. - Notices immediately the vault has no PARA folders: "you don't have parah in this one because it's not like your go-to vault" (t=120) — treated as context, not a flaw. #### ⏱️ [00:02:32](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=152s) Reading the graph view - Graph view called "great eye candy but... not terribly helpful" (t=152), yet Milo still extracts real structure from it. - Identifies one dominant project cluster (the book *Life in Perspective*), a map-of-content node with many inbound links, and a readwise-import cluster. - Names "placeholder notes" — faded nodes with backlinks pointing at notes that don't exist yet — as the coolest structural feature: a note can accrue five links "two years" before it's created (t=201–234). #### ⏱️ [00:04:11](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=251s) Scale: 1,662 notes vs. 17,000 - Forte's vault: 1,662 notes, four folders. Milo's main vault: ~17,000, built mostly by hand, "not really looking to import a lot" (t=267–276). - Milo explicit that raw count doesn't matter on its own — "it's fine if you have 20 notes" — what matters is what the links and structure communicate. #### ⏱️ [00:06:16](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=376s) The one hard rule: never point AI at canonical notes directly - Milo's stated reason: the vault represents hundreds of hours of manual linking and outlining that an automated import can't replace. - Mitigation is backups, not access limits: "if AI can read something, you have to assume... it may be able to write" (t=485–495); pairs an on-device dated Cmd+D duplicate with off-device backups (Google Drive, a zip drive). - Notes he hasn't personally seen malicious edits but has "heard of those instances" — risk tolerance, not a guarantee, is doing the work here. #### ⏱️ [00:07:01](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=421s) What Obsidian actually is - Plain-text markdown files in a folder; Obsidian is "a skin," not a data holder, unlike Evernote or Notion (t=421–454). - This is why any AI tool can be pointed straight at the folder with no export step, and why switching AI tools later costs nothing (t=1100–1118, referenced back to this point). - Duplicate vault size check: ~120 MB for 1,662 files, which both hosts flag as larger than plain text alone would predict — likely photos/media pulled in via Readwise (t=632–662). #### ⏱️ [00:11:38](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=698s) Three ways to connect AI, and a stated preference - Names three options: point an agent (Claude Co-work, ChatGPT) at a folder; community plugins (Copilot, Smart Connections); the Obsidian CLI. - States a preference for folder-pointing: plugins add dependency on outside developers ("how are they going to maintain... all these extra breaking points," t=741); the CLI is powerful but not worth the setup cost without a specific need (t=1815–1832). - Folder-pointing also gives free-standing privacy granularity — point at a curated subfolder instead of the whole vault (t=875–902) — leading to Milo's line: "folders are back, baby." #### ⏱️ [00:15:38](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=938s) The three orientation files - Before any diagnostic prompt, every new session starts with the agent reading three files: a "me" file, a vault-map file, and a skill-map file (t=938–987). - Doesn't have to be three separate notes — "it needs to hit three buckets" (t=960). - The skill map is where predefined, reusable skills (like the dossier prompt, the Sherpa system) live for the agent to draw on just-in-time (t=997–1019). #### ⏱️ [00:17:06](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1026s) The AIOS three-layer model - Core = the vault/knowledge system (predates AI entirely). Outer = whichever AI tool is currently in use — "the least important... the most ephemeral" (t=1058–1081). Middle = a translation layer, i.e., the three orientation files. - Explicitly frames this as inverting the default mental model, where most people put the AI tool at the center and treat their own data as peripheral enrichment (t=1127–1156). #### ⏱️ [00:19:46](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1186s) Digital sovereignty vs. digital feudalism - Names the stakes directly: "how can we actually own the land that we think on as opposed to digital feudalism when we're doing all of our work on somebody else's land" (t=1186–1196). - Attributes vendor lock-in to platform incentive, not user benefit: "their incentive is to lock you in if they can do that" (t=1174). #### ⏱️ [00:21:07](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1267s) Links as an importance signal — the PageRank analogy - Claim: a note with 50+ links "has high gravity" and an AI (or a human) can use that to decide what's worth reading closely in a vault too large to read in full (t=1316–1356). - Explicit mechanism named: "instead of having thousands, millions of humans over there cataloging, let's infer meaning from the graph, from the connections" — directly compared to Google PageRank (t=1364). - Numeric claim introduced here and returned to later: past roughly 500 notes, full-fidelity AI reading of a vault breaks down without a distillation layer (t=1632–1654). #### ⏱️ [00:23:54](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1434s) The dossier exercise - Milo's go-to opening prompt for an unfamiliar vault: ask for a "dossier" — scan everything, "write a dossier about me" (t=1441–1453). - Narrowed on the spot to project scope: "read through this entire Obsidian Vault, gather the context you need, and write a project dossier" for the book project specifically (t=1470–1488). - While the multi-minute run happens, both hosts describe their actual behavior: Forte checks email and drifts to YouTube; Milo watches the progress indicator and interrogates his own approach (t=1507–1573). #### ⏱️ [00:28:01](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1681s) Precision targeting: exact file paths over whole folders - Demonstrates copying the exact system file path to one core note (often a map of content) and instructing the agent to treat it as the starting point, rather than pointing at a whole folder (t=1690–1761). - Stated reason: avoiding wasted tokens and wasted uncertainty about whether the agent found the right material — "absolute certainty" (t=1744–1753). #### ⏱️ [00:30:32](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1832s) The Sherpa system - Two stacked skills: first, train the agent on 15–20 of Milo's favorite existing maps of content so it can "reverse engineer what makes a good map of content" into a style guide (t=1862–1888). - Second, a "questionnaire skill" that runs against that style guide whenever starting research on a new topic, connecting new material to related existing notes rather than producing a generic summary (t=1906–1935). #### ⏱️ [00:32:42](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1962s) Map of content vs. table of contents - Core distinction: "a map of content is fluid... and an index feels a bit more static" (t=1973). - The value is in the act of making it — "organization in this way is ideation" (captioned "ideiation," likely a caption artifact for "ideation," t=2028) — not in the finished artifact. - Explicit caveat that an AI-built map of content is only "potential meaning" until a person works with it further; otherwise it's "cognitive debt" (t=1935–1948). #### ⏱️ [00:36:42](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=2202s) Quarantining AI output: the AIOS/History folder - Everything the agent creates is saved to a dedicated `AIOS/History` folder, timestamped first so it sorts chronologically, and nowhere else (t=2211–2261). - Stated as a hard boundary: "there's a very clear line of separation between my notes, my thoughts, and then where AI's notes and thoughts go... that's the only place it can write to" (t=2233–2248). - Turned into a standing, repeatable skill on the spot rather than left as a manual copy-paste step each time (t=2354–2398). #### ⏱️ [00:52:42](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=3162s) "A summary of a summary is not good" - Forte's sharpest critique of the whole session's workflow, prompted by noticing AI outputs are consistently drawn from a small fraction of the vault (he cites asking directly and getting "0.5%," t=3162–3180). - Mechanism: "the more times you summarize, all the specificity and the signal gets stripped out of it" (t=3193) — the resulting artifact becomes a "depreciating asset" if never referred back to. - Fix proposed on the spot: instruct the agent to cite sources and use verbatim quotes rather than paraphrase (t=3235–3263), paired with Milo's own hard rule, "do not rewrite me," because paraphrased output can get returned to him later as though it were his own original words (t=3273–3289). #### ⏱️ [00:55:10](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=3310s) Caught overindexing on thin data - The agent's gap analysis overweights a short, admittedly incomplete list of "email topics I wanted to write" as though it captured the whole plan (t=3322–3372). - Forte's real-time diagnosis: "AI always assumes that whatever information it has access to is all the information that exists... it's like a little universe, a closed universe that you've created" (t=3391–3412). - No system safeguard proposed — the fix named is human discernment in the moment: "you bring your own discernment. This is not truth. This is not reality" (t=3353). #### ⏱️ [00:59:49](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=3589s) Close: "teach AI to learn about us" - Closing problem named: Forte's own information is scattered across "eight to ten different platforms" despite wanting one source of truth (t=3524–3546); the fix offered is connecting the agent to those tools (Google Drive, ClickUp) rather than consolidating first. - Milo's closing line, offered as the session's real thesis: "everyone is trying to learn about AI, but really we need to teach AI to learn about us... it becomes futureproof and it doesn't matter what the AI is because we own the core" (t=3620–3630). - Sign-off is self-aware about the format: Forte jokes their shared reputation rests on "not trying to manufacture certainty where it doesn't exist" (t=3667), directly against the caption's likely-garbled "watch how I made a million dollars in 7 minutes" aside. ## 📚 References - [Video: Nick Milo Reads My Obsidian Vault (The Full Session)](https://www.youtube.com/watch?v=bdjYo-x4PUw) - Obsidian (note-taking app referenced throughout as the vault's host — plain-text markdown, no export lock-in) - Readwise (referenced as the source of an automatic-import folder inside the vault, t=189) - Obsidian community plugins named but not linked in the transcript: Copilot, Smart Connections (t=731) - Nick Milo's "Linking Your Thinking" framework and "AIOS"/"Sherpa system" concepts, named and demonstrated but not linked in the transcript - Tiago Forte's *Building a Second Brain* and PARA method — referenced implicitly (Forte's vault is shown explicitly *not* using PARA, t=120) rather than named directly in the transcript ## 🗣️ Paraphrase Forte hands his real, high-stakes vault — the one driving a book already in presales — over to Milo for an unscripted read, the kind Milo says has become "a sixth sense" for him: check the theme, check the folder sidebar for structure, check whether an open note has links, because links reveal how someone actually thinks rather than just what they've filed. From there the two build outward: a project dossier, a self-generated map of content the agent is explicitly told to model on Milo's own stated standards, and finally a gap analysis run against Forte's real marketing plan, which turns up something genuinely useful and something genuinely wrong in the same pass — real blind spots surfaced from source material Forte hadn't reread, alongside an agent quietly treating an admittedly incomplete list as though it were the complete picture. The throughline is governance, not magic. Every technique demonstrated — duplicate before you point AI at anything, quarantine everything AI writes into its own folder, ask for verbatim quotes instead of summaries, tell the model directly not to paraphrase you — exists because the two hosts don't trust the default behavior of an AI agent turned loose on a personal vault. The AIOS framing (vault as core, AI tools as disposable outer layer, three files as translation layer) is offered as the reason any of this is worth doing at all: it keeps the knowledge portable across whichever AI product wins this particular month, rather than binding years of thinking to one vendor's continued goodwill. ## 📰 Newsletter Two of the biggest names in personal knowledge management sat down for an unscripted stress test: Tiago Forte (PARA, Building a Second Brain) hands his real, live, production-stakes Obsidian vault to Nick Milo (Linking Your Thinking, maps of content) and lets him diagnose it cold, then point Claude at it. What's worth your time here isn't the "AI + notes" pitch — it's the concrete governance layer underneath it: never point an agent at your canonical notes without a backup, quarantine everything AI writes into its own folder, and explicitly instruct the model not to paraphrase you, because a summary of a summary strips out the specificity that made your notes worth taking in the first place. The most interesting moment isn't a success — it's Forte catching the agent in real time treating an incomplete list as though it were the whole picture, a reminder that "AI assumes what it has access to is everything" is a live failure mode, not a hypothetical one, however good the demo looks. ## ✅ Action Items #### DevRel - Not applicable to this source — the video is a personal-productivity/PKM demo between two independent creators, with no developer-relations, SDK, or community-program content. #### OSPO - Not applicable to this source — no open-source governance, licensing, security, or SBOM content appears anywhere in the transcript. #### Community - Not applicable to this source — no maintainer, contributor, sponsor, or cross-project ecosystem content; this is a two-person conversation on a personal channel. #### WordPress - Not applicable to this source across all subteams (Core, Core Performance, Core AI, Core Program Team, Mobile, Playground, Community, Training, Marketing, Support, Design, Documentation, Photos, TV, Polyglots, Accessibility, Meta, CLI, Hosting, Tide, Openverse, Themes, Plugins) — WordPress is never mentioned; this is a general Obsidian/AI workflow demo. #### The WPOCC - Not applicable to this source — no fellowship, funding, or governance content; unrelated to The WPOCC's remit. #### FAIR - Not applicable to this source — no federation, provenance, or governance-artifact content. ## 🧪 Hypotheses - H1: An AI agent pointed at a large personal vault will, in practice, read and weight a note's inbound-link count as a proxy for its importance, consistent with Milo's "gravity" claim (t=1316). - H2: Full-fidelity AI reading of a vault degrades past a specific note-count threshold (Milo states ~500, t=1632), and degradation is measurable as completeness/accuracy loss in agent-generated summaries as vault size scales up. - H3: Pointing an agent at one precisely-targeted core note (a map of content or similar) via an exact file path produces more accurate, lower-token output than pointing it at the containing folder or whole vault (t=1690–1761). - H4: Explicitly instructing an agent "do not rewrite me — use verbatim quotes" measurably reduces unwanted paraphrase in later retrieval tasks compared to no such instruction (t=3235–3289). - H5: Given a small, admittedly incomplete data subset, an agent will treat it as representative of the whole rather than flagging its own uncertainty about coverage, unless explicitly prompted to check for gaps (t=3391–3412). - H6: A vault's organizational style (folder-first/PARA vs. link-first/MOC) has a measurable effect on the quality of AI-generated context, independent of vault size — testable because Forte's own working vault in this video uses neither PARA folders nor a dense pre-existing MOC layer, yet still produces a usable dossier (t=120, t=1470–1488). ### 📊 Methodology - For H1: instrument an agent's actual file-access pattern against a vault with known link-count distribution; compare which notes get opened/quoted versus each note's inbound-link count, controlling for note length and recency. - For H2: run the identical "write a dossier" prompt against vaults sampled at increasing note counts (e.g., 100, 500, 1,000, 2,500, 5,000) from the same underlying corpus, scoring completeness against a human-annotated ground truth of what the vault actually contains. - For H3: A/B compare agent output (accuracy, token cost, time) from (a) a single exact file path to a core note vs. (b) the containing folder vs. (c) the whole vault, holding the prompt and underlying material constant. - For H4: run a retrieval/summarization task with and without the "do not rewrite me, verbatim quotes only" instruction, scoring output against source text for unattributed paraphrase. - For H5: deliberately supply an agent with a partial data subset labeled as such, and measure how often it flags incompleteness unprompted versus only after being asked directly what percentage of source material it reviewed (the video's own diagnostic move, "what percentage of the notes did you actually look at," t=3162–3180). - For H6: build two matched test vaults from the same source notes — one organized PARA-first with minimal links, one organized MOC-first with minimal folders — and compare agent dossier quality, using the same rubric, on identical underlying content. ### ❓ Questions - Does the "link count = high gravity" heuristic (t=1316) hold when links are auto-imported (Readwise highlights, footnotes) rather than manually made one at a time, as Milo's own vault mostly is (t=267–276) but Forte's Readwise-fed vault partly is not (t=189)? - Is the ~500-note fidelity ceiling (t=1632) a fixed number or does it scale with average note length and model context window — and does it move as context windows grow? - Forte's working vault for this book visibly skips PARA (t=120). Is that specific to project-dedicated vaults, or does Forte's main/archival vault also depart from PARA in practice — the video never says? - What would happen if the agent were pointed at a genuinely blind, unlabeled folder split — half PARA-organized, half MOC-organized — and asked to produce equivalent output, to actually test the folder-vs-links question rather than assume one demo settles it? - Does "do not rewrite me" (t=3273) hold up across model versions and providers, or is it a fragile instruction that degrades as the underlying model changes — something neither host tests on camera? ### 🔮 Future Research - No controlled comparison exists in the video between a PARA-organized and a MOC-organized vault on AI-context quality; the claim that links matter more than folders is demonstrated once, on one vault, by the person whose framework depends on links mattering. - No test of the 500-note fidelity threshold across vault sizes — it's stated as a known number, not measured on camera. - No check of whether link-provenance (manually made vs. auto-imported) changes the reliability of link-count-as-importance; Forte's vault mixes both and this was never disambiguated. - No adversarial test of the "AI assumes what it has is everything" failure mode beyond the one instance caught by chance (t=3391) — worth testing deliberately with known-incomplete inputs to see how often it's missed rather than caught. - No test of whether the AIOS/History quarantine folder actually prevents unwanted writes to canonical notes, versus simply being where the agent happens to write when given no other instruction — the boundary is never stress-tested (e.g., asking the agent directly to edit a canonical note to see if it refuses). ### 💡 Implications - **Practice.** This is the most directly testable video in recent memory against Courtney's own vault, because her vault already runs the exact hybrid this video treats as two separate camps: PARA-style top-level folders (`1. Projects/`, `2. Areas/`, `3. Resources/`) *and* 23 Milo-style MOCs in `3. Resources/MOCs/`, on top of a Zettelkasten Literature/Permanent/Fleeting split with roughly 1,362 atomic Permanent notes and ~2,400 notes total. Milo's cold-read diagnostic (t=82–138) applied to that vault would register PARA folders, MOC membership, *and* link density simultaneously — three importance signals that could disagree — a disambiguation Milo never had to make here, because Forte's demo vault visibly lacks PARA (t=120). Worth running the actual experiment: point an agent at the vault the way Milo points at Forte's, and check whether its sense of "what matters" tracks folder location, MOC membership, or raw inbound-link count. - **Scale check.** Milo's stated ~500-note fidelity ceiling (t=1632) is a specific, falsifiable number. Courtney's vault is roughly 5x past it. If the claim holds even loosely, any single "read the whole vault" prompt against her corpus is already sampling, not reading — worth confirming directly against whatever the synthesis-pass and vault-hygiene tooling currently assumes about full-vault reads, rather than taking it on faith either way. - **Governance.** Milo's AI-output quarantine — nothing an agent writes lands anywhere but a dedicated `AIOS/History` folder (t=2202–2261) — is a stricter, generalized version of the staging pattern this very deep-read pipeline already uses structurally (writing to `notes6/` before any vault write, never touching the vault directly). Worth naming as a shared principle across both systems and checking whether any part of the pipeline currently writes to the vault without an equivalent staging step, rather than assuming the analogy already covers every path. - **Security.** Both hosts state the write-access risk as fact, not hypothesis: "if AI can read something, you have to assume... it may be able to write" (t=485–495). Their mitigation is backup cadence, not access restriction — worth checking Courtney's own agent permissions and backup schedule against that specific bar (automatic, on-device *and* off-device) rather than assuming notes6-staging alone is sufficient coverage. - **Discernment as the actual safeguard.** The one moment the video catches a real failure (t=3391–3412, the agent overweighting an incomplete email-topics list) is caught by a human noticing, not by any system check either host has built. That's a real limit worth carrying forward: neither framework, on its own terms, claims to catch this reliably — it's presented as a habit ("you bring your own discernment," t=3353), which doesn't scale the way a system check would. - **Funding models.** Not applicable directly, though both hosts sell commercial frameworks built on the claims demonstrated here (Linking Your Thinking; Building a Second Brain) — relevant context for how much weight to put on either framework's untested numeric claims, addressed further under Logical Validity below. ### 🧠 Learning Styles Assessment - **Reading level estimate:** Flesch-Kincaid Grade Level ≈ 7–8. Basis: unscripted, second-person conversational register; frequent short backchannel affirmations ("Interesting," "100%," "Mhm") standing in for connective prose; long paratactic run-on sentences typical of live spoken narration rather than edited writing. #### 👀 Visual learners - **Present:** live on-screen Obsidian graph view, folder sidebar, and Claude Co-work interface throughout; a referenced graphic explaining the AIOS three-layer model (t=1019). - **Improve:** a static after-the-fact diagram of the AIOS three layers and the dossier → backlink → MOC workflow would help viewers who tune out during the live click-through segments. #### 🔊 Aural learners - **Present:** entire video is a narrated, unscripted two-person conversation with minimal reliance on unread on-screen text. - **Improve:** a standalone 3–4 minute audio recap of the AIOS model and the four governance rules (backup first, quarantine folder, verbatim quotes, don't rewrite me) without the screen-share dependency. #### 🗣️ Verbal learners - **Present:** clear working definitions offered live for map of content vs. table of contents (t=1973), and the PageRank analogy for link-weighted importance (t=1364). - **Improve:** an explicit glossary distinguishing "map of content," "dossier," "vault map," and "skill map" — four related-but-distinct terms used somewhat interchangeably across the session. #### 🧩 Kinesthetic learners - **Present:** exact reusable prompts spoken on camera for the dossier, the backlink-and-MOC request, and the marketing gap-analysis request. - **Improve:** a downloadable checklist of the four governance rules and the three-orientation-file setup so a viewer doesn't have to transcribe the spoken prompts by hand. #### 🧠 Logical learners - **Present:** the explicit three-layer AIOS structure and the folder-vs-plugin-vs-CLI comparison laid out as a clean decision point (t=698–741). - **Improve:** a decision table mapping vault size to recommended approach (whole-vault prompt vs. targeted file path vs. MOC-first distillation) would make the branching logic Milo narrates explicit rather than spoken. #### 🧑‍🤝‍🧑 Social learners - **Present:** the entire format is a peer-to-peer dialogue with visible disagreement-adjacent moments (e.g., Forte questioning whether the agent is overindexing, t=3329) rather than a solo lecture. - **Improve:** a prompt inviting viewers to run the same cold-read diagnostic on their own vault and compare notes would extend the format into an actual social loop. #### 🧘‍♀️ Solitary learners - **Present:** the governance rules (backup, quarantine folder, don't-rewrite-me) are all individually adoptable without needing a second person or a live demo. - **Improve:** a self-check worksheet ("does my vault have a strong folder structure? how many links does my most-connected note have? do I have an AI-output quarantine folder?") would let a solitary viewer self-diagnose the way Milo diagnoses Forte's vault. #### 🌿 Naturalistic learners - **Present:** organic metaphors run throughout — "gardener" vs. "architect" organizing styles (t=2052), "cartographer of my ideas" (t=2005), placeholder notes as things that "accrue" over time (t=211). - **Improve:** none needed; the metaphor set is already a natural fit for this content. ### 📊 Accessibility Assessment *Assessed from the transcript only — the video itself was not visually reviewed, so on-screen contrast, focus order, and caption presence/accuracy are not directly observed.* #### Screen reader semantics - **Present:** the session has clear topical segmentation (cold read → connection methods → orientation files → dossier → gap analysis → close) that would translate to good heading structure if repurposed as a written page. - **Improve:** if published as an article, map each phase to its own `<h2>` landmark (WCAG 2.2 SC 1.3.1 Info and Relationships; SC 2.4.6 Headings and Labels). #### Keyboard navigation - **Not observed** — the video shows live mouse/click interaction with Obsidian and Claude Co-work throughout; no keyboard-only accessibility of the demoed software itself is assessable from a transcript (would map to SC 2.1.1, 2.4.3 for any companion interactive material). #### Contrast and readability - **Not observed** — on-screen text contrast in the dark-theme Obsidian and terminal views is not visually verified from the transcript; flag for direct review before treating as compliant (WCAG 2.2 SC 1.4.3, SC 1.4.11). #### Text resizing and reflow - **Not observed** — no reflow-relevant UI is described in the transcript itself (SC 1.4.4, SC 1.4.10 would apply to a companion written page rather than the video). #### Media and captions - **Present:** a full timestamped transcript exists (the source of this analysis), satisfying the base text-alternative requirement (WCAG 2.2 SC 1.2.2 Captions). The transcript is clearly auto-generated, not human-corrected — evidenced by garbled proper nouns and terms ("parah" for what is almost certainly "PARA," "ideiation" for "ideation," "Diego" where context makes "Tiago" far more likely, t=3651, and "Frank" as a joking made-up AI model name that the caption renders without any signal it's a bit, t=1068–1081). - **Improve:** confirm on-platform captions are auto-generated vs. creator-reviewed before publishing any derivative that quotes them as authoritative (SC 1.2.4); the video's own screen-shared UI (folder names, note titles) is never described aloud independently of what's visible, so a non-sighted viewer would lose specifics that a sighted viewer gets from the screen alone. #### Forms and interactions - **Not applicable** — no forms or structured interactive input appear in the video itself; the "interactions" shown are prompt entry into a chat interface, not a form in the accessibility-standards sense. #### Standards reference - Gaps above map to WCAG 2.2 AA: 1.2.2, 1.2.4, 1.3.1, 1.4.3, 1.4.4, 1.4.10, 1.4.11, 2.1.1, 2.4.3, 2.4.6. Most items require direct visual/audio review of the published video to confirm or close; this assessment is transcript-only. ### 🌍 DEIB Assessment #### Cultural breadth - **Present:** none beyond the two hosts' own frameworks — no cited thinkers, traditions, or non-Western examples appear in this transcript, unlike PKM-adjacent videos that pull in named philosophers or historical figures. - **Improve:** the video would benefit from naming where either framework's core claims (link-weighted importance, progressive summarization) draw on or diverge from established information-science work (e.g., citation-graph analysis, structured note-taking traditions predating digital tools) rather than presenting them as self-originated. #### Inclusive language - **Present:** consistent second-person, non-gendered address between the two hosts throughout; no idiomatic language likely to translate poorly. - **Improve:** none needed; the register is already plain and direct. #### Neurodiversity support - **Present:** the three-orientation-files pattern and the four governance rules are repeated and reinforced multiple times across the session, aiding retention. - **Improve:** the conversation runs as one long unbroken back-and-forth with no spoken chaptering cues ("that's the end of the diagnostic phase," etc.) — on-screen or spoken segment markers would help viewers who rely on explicit structure to track a long-form, unscripted conversation. #### Representation - **Present:** two named individuals converse throughout; neither states pronouns in the transcript, so this analysis uses they/them for both rather than inferring from their names. - **Improve:** both speakers represent the same narrow niche (US-based, English-language, commercial PKM-coaching content) — a genuinely diverse vault-reading demo (different note-taking traditions, non-English vaults, non-commercial use cases) would test whether the diagnostic method generalizes beyond this specific community. #### Psychological safety - **Present:** low-stakes, collegial tone; Forte's real, high-stakes production vault is treated respectfully, and self-deprecating jokes ("you already know I'm basic," t=92) keep the tone light without dismissing genuine mistakes (the overindexing catch, t=3391, is named plainly rather than glossed over). - **Improve:** none needed for this content; no sensitive or high-stakes material requiring a content warning. ### ✅ Logical Validity #### Strong reasoning - **Mechanistic analogy, correctly scoped (t=1364):** Milo's comparison of link-count-weighted note importance to PageRank is a real, established technique (using in-link density as a relevance/authority signal), not just a rhetorical flourish — it explains *why* the "gravity" claim would work, not just that it does. - **Causal reasoning on information loss (t=3193):** Forte's claim that repeated AI summarization strips specificity and signal is a plausible, mechanistically stated, and testable claim about lossy compression — it's the load-bearing critique of the whole session's workflow and holds up on its own terms. #### Fallacies - **Hasty generalization — single-vault demonstration treated as a general diagnostic method.** Location: throughout, but sharpest at the numeric claims (link-count-as-gravity, t=1316; the ~500-note fidelity ceiling, t=1632). Why it's weak: both are stated as established facts of how AI reads vaults, demonstrated exactly once, on one vault, by the person whose own framework depends on links mattering more than folders. Fix: flag both as "this is what I've observed" rather than settled fact, and test against a second vault built on different principles — precisely the kind of test this video never runs. - **Appeal to (self-)authority embedded in the prompt itself.** Location: t=2540–2554, the literal instruction to the AI: "use all of the advice and recommendations issued by Nick Milo" for what makes a good map of content, delivered on Milo's own appearance on the video. Why it's weak: it substitutes a named authority for an external, demonstrated criterion of quality — the AI is told whose taste to defer to rather than given an independently verifiable standard. Milo's own joke response ("who's that? that'd be insulting") acknowledges the moment's self-referential humor without addressing the substance. - **Confirmation-loop risk (adjacent to circular reasoning, not a classical named fallacy).** Location: near-constant short affirmations between the two hosts ("Interesting," "100%," "So important," repeated throughout) with only one clearly named moment of internal pushback (the overindexing catch, t=3391). Why it's weak: two collaborators mutually promoting adjacent commercial frameworks are structurally unlikely to challenge each other's core claims on camera; the demo's smoothness is partly a function of that dynamic, not purely of the method's robustness. #### Tone of voice - Collegial, enthusiastic, first-person demonstrative register between two peers; frequent short backchannel affirmations creating a conversational rather than lecture-style tone; self-aware and mildly self-deprecating about the genre's own hype tendencies, made explicit in the closing line about "not trying to manufacture certainty where it doesn't exist" (t=3667). ### 🛡️ DARVO Analysis - **Deny:** Not applicable to this source — no accusation exists to deny. There is no conflict, wrongdoing, or harm narrative anywhere in the transcript; this is a collaborative technical demo between two peers. - **Attack:** Not applicable to this source — no critic or accuser is named or attacked by either host. - **Reverse victim/offender:** Not applicable to this source — no victim/offender framing appears anywhere in the transcript. - **What the actual rhetoric is doing instead:** since DARVO's framework doesn't fit a source with no accusation in it, the more relevant read is the mutual-promotion dynamic named above under Logical Validity — two hosts with adjacent commercial frameworks (Linking Your Thinking; Building a Second Brain) demonstrating compatibility and validating each other's claims in real time, with exactly one visible moment of internal challenge (the overindexing catch, t=3391) and no instance of either host contesting the other's stated method. That's a collegiality/incentive-alignment observation, not an accusation of bad faith — both hosts are transparent about who they are and what they sell, and the closing line explicitly names the risk of overclaiming rather than hiding it. ## 🎯 Conclusion The session's real value sits below the "watch two PKM experts talk AI" framing: it's a working checklist of governance moves for anyone pointing an AI agent at years of personal notes — back up before you point AI at anything, quarantine everything the AI writes into its own folder, ask for verbatim quotes instead of summaries, and tell the model directly not to paraphrase you. The MOC-vs-PARA framing the video sets up at the start (Milo's method meeting Forte's un-PARA'd vault) never gets a real resolution — Forte's own working vault simply skips PARA for this project, which is itself a more interesting data point than anything either host says about it directly. The most exportable, testable claims — link count as an importance signal, a numeric vault-size ceiling for full-fidelity AI reading, and the observed failure mode of an agent assuming partial data is complete — are all falsifiable against a real vault. Practical next step: run Milo's exact cold-read sequence (theme, folder structure, open-note links) against a vault that runs both PARA folders and a dense MOC layer at once, and see which signal the diagnostic actually tracks when both are present. ## 🔄 Refactor Appendix — Nick Milo Reads My Obsidian Vault (The Full Session) > Density: 5 Literature atoms (H4), 3 Permanent candidates (H4). > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > Atomicity: one declarative sentence per atom. ### 📚 Literature Atoms (H4) #### [H1] Links in an open note reveal how someone thinks > [!quote] > I would try to see if there were links. That would tell me a lot about how the person is linking their thinking. **Context & use**: - [00:02:18](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=138s) — part of Milo's stated cold-read diagnostic sequence (theme, folder sidebar, then open-note links). - Load-bearing for the whole diagnostic method: it's the one check Milo names as revealing thinking rather than just filing. **Tags**: #xref #aphorisms  #### [H2] Never point AI directly at canonical notes > [!quote] > I don't want AI to touch my notes, my original canonical notes directly. **Context & use**: - [00:06:16](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=376s) — stated as the one hard rule before any AI work begins in the session. - Grounds every governance move that follows: backups, the AIOS/History quarantine folder, "do not rewrite me." **Tags**: #xref  #### [H3] A heavily linked note has more "gravity" for AI > [!quote] > that one that had 50 plus links to it, it knows that is has high gravity **Context & use**: - [00:21:56](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=1316s) — caption grammar as captured ("is has" rather than "it has"), likely a transcription artifact rather than the speaker's actual phrasing. - Directly compared moments later to Google's PageRank (t=1364): link density as an inferred-importance signal instead of manual cataloging. **Tags**: #xref  #### [H4] A map of content is ideation, not mere organization > [!quote] > in this mapping way, it's not organization for organization's sake. organization in this way is ideiation **Context & use**: - [00:33:48](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=2028s) — "ideiation" is the caption's rendering; the surrounding sentence and every other use in the session make "ideation" the clear intended word. - Directly distinguishes a map of content from a table of contents (t=1973): the value is in the act of building it, not the finished artifact. **Tags**: #xref #aphorisms  #### [H5] A summary of a summary strips out the signal > [!quote] > a summary of a summary is not good. The more times you summarize all the specificity and the signal gets stripped out of it. **Context & use**: - [00:53:13](https://www.youtube.com/watch?v=bdjYo-x4PUw&t=3193s) — Forte's sharpest critique of the AI-generated artifacts produced earlier in the same session. - Directly motivates the fix proposed immediately after: cite sources, use verbatim quotes, don't let the agent paraphrase. **Tags**: #xref #aphorisms  ### 🧠 Permanent Candidates (H4) #### [P1] Link-weighted reading only works when links are deliberate **Claim** (Courtney's voice): An AI agent reading my vault will treat my most-linked Permanent notes as the important ones, so that ranking is only trustworthy where I made the links myself and not where they arrived via automated import. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for manually built links (the case Milo demonstrates, and the case for most of my 1,362 Permanent notes); weakens for vaults or sections where links arrive automatically (bulk Readwise imports, auto-generated backlinks) rather than through deliberate synthesis. **Links** (≥1 with annotation): - [[MOC design principles]] — because extends: both depend on the underlying links reflecting real synthesis work, not import volume. - [[Progressive summarization]] — because contrasts on: progressive summarization signals importance through manual highlighting layers; this claim signals it through link topology instead.  #### [P2] Building a map of content is a thinking act, not a filing act **Claim** (Courtney's voice): The value of assembling a map of content comes from the act of relating notes to each other while I build it, not from having the finished document available afterward. **Evidence pointer**: [[#H4]] **Scope & conditions**: Holds when I (or an agent, with my review) build the MOC from notes I've actually internalized; weakens sharply for an MOC an agent drafts wholesale from a folder I haven't read yet, which Milo himself flags as only "potential meaning" until someone works with it further. **Links** (≥1 with annotation): - [[PARA method]] — because contrasts on: PARA organizes by actionability at the folder level; this claim locates the value one layer up, in the act of relating ideas regardless of which folder they live in — and Forte's own working vault in this video skips PARA entirely for this project, which is evidence the two layers can be decoupled. - [[Zettelkasten atomicity]] — because applies-in: a note only earns a place in an MOC once its claim is specific enough to relate meaningfully to others, the same discipline atomicity enforces at the note level.  #### [P3] Quarantine everything an AI agent writes into its own folder **Claim** (Courtney's voice): Any AI agent given access to my vault should only ever write into a dedicated, clearly separate location, never directly into my canonical Literature or Permanent notes. **Evidence pointer**: [[#H2]] **Scope & conditions**: Applies regardless of vault size or agent trustworthiness — the failure mode it guards against (an accidental edit or deletion of canonical material) doesn't require malicious intent, only write access plus an assumption of read-only behavior that isn't guaranteed. **Links** (≥1 with annotation): - [[readwise-deep-read pipeline]] — because extends: the notes6/-style staging directory this very deep-read task writes to, before any vault write, is the same governance pattern Milo demonstrates as AIOS/History — worth confirming no part of the pipeline currently writes to the vault without an equivalent staging step.  ### ✏️ Fleeting Queue - [F1] Test whether Milo's "50+ links = high gravity" claim actually predicts which of my Permanent notes an agent treats as load-bearing when pointed at a real subfolder of the vault. — https://www.youtube.com/watch?v=bdjYo-x4PUw #xref - [F2] Check whether my 23 existing MOCs would read to an outside auditor (Milo-style cold read) as fluid ideation, or as static indexes that have drifted toward PARA-style filing over time. — #xref ### → Things 3 (max 2 actions) - [ ] Run Milo's three-orientation-file pattern (a me-file, a vault-map file, a skill-map file) against my vault and compare it to what the readwise-deep-read pipeline currently hands a fresh agent session — context #pkm-experiment #things - [ ] Confirm every write-capable step in the deep-read pipeline stages through notes6/ (or equivalent) before touching the vault, matching Milo's AIOS/History quarantine pattern, rather than assuming it already does — context #vault-hygiene #things  📄 **Full text:** [[Full Document Contents/Articles/Nick Milo Reads My Obsidian Vault.md|Nick Milo Reads My Obsidian Vault by Tiago Forte]] ## Highlights - gets beyond let's say around 500 notes and it depends on how long the notes are then again it's a situation where um what's being indexed is not with 100% ([View Highlight](https://read.readwise.io/read/01m270d1fxfdbefwkmqdc23gab)) - Note: Milo's fidelity threshold for AI reading a vault without a distillation layer. Courtney's vault is ~1,362 Permanent notes and ~2,400 total — roughly 5x past it. Stated as fact but demonstrated once, by the person whose framework depends on it. - Tags: xref, personal-knowledge-management