I Connected Claude to Everything. Then I Saw the Gap.

I Connected Claude to Everything. Then I Saw the Gap. ![rw-book-cover](https://readwise-assets.s3.amazonaws.com/static/images/article0.00998d930354.png) 🔗 **Source:** [I Connected Claude to Everything. Then I Saw the Gap.](https://www.youtube.com/watch?v=vqu4tNstz4k#transcript) > [!note]+ Document Notes > tags: [pkm, second-brain, ai-connectors, knowledge-management, ai-workflow] > author: Tiago Forte > source: YouTube > reference: https://www.youtube.com/watch?v=vqu4tNstz4k > date created: 2026-06-30 09:00 > > # 🏷️ Title: I Connected Claude to Everything. Then I Saw the Gap. >  ## 🧭 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 - [Video: I Connected Claude to Everything. Then I Saw the Gap.](https://www.youtube.com/watch?v=vqu4tNstz4k) - [Building a Second Brain (Tiago Forte)](https://www.buildingasecondbrain.com/) - [Anthropic — Claude connectors / MCP](https://modelcontextprotocol.io/) ## 📝 Summary - **What happened.** Tiago Forte spent a month connecting Claude to his notes, calendar, email, and cloud files via official connectors, then reported the nuance he says nobody online is talking about. His core metaphor: connectors are a "nervous system" for your AI second brain — the channels that bring context in and push actions out. - **Why it matters.** People assume that once they add a connector (e.g., Notion in ~30 seconds), "Claude now knows my Notion." Forte says that's false. Claude *samples* the source — his tests show it reads roughly 0.5%–5% of what's available and favors recent files. That gap between assumed and actual coverage is the whole point. - **Key stakeholders.** Knowledge workers, PKM/second-brain practitioners, DevRel and training teams onboarding people to AI connectors, and anyone wiring internal docs into an LLM. - **Risks and tradeoffs.** The "illusion of authority" — treating a gut-instinct partial answer as authoritative. Plus security: combining too many connectors in one session invites the "lethal trifecta." - **Bottom line.** Connectors are powerful but not infallible. Human judgment lives in noticing and filling the gap; verify more, not less, when accuracy matters. ## 📖 Main Content - **The nervous-system metaphor.** Forte frames the brain as a black box that can only sense or act on the world through the nervous system. Connectors do the same for an AI second brain: they extend a central repository of intelligence outward, bringing valuable data in and sending observations, recommendations, and sometimes actions back out. The payoff is eliminating hand-curated copy-paste-reformat context work. - **The gap (the thesis).** Adding a connector does not mean the model ingests the source. Claude is *sampling* — doing reconnaissance via targeted searches and reporting on the small slice it assumes is most relevant. Forte's tests put actual coverage at 0.5%–5%, and when he pressed Claude about a Google Drive search it admitted it favored recently touched files, ignoring years of older material. - **The illusion of authority.** The danger isn't that Claude is wrong; it's that a first-pass, gut-instinct answer *looks* authoritative. For accuracy-critical work, that means checking Claude's work is *more* important, not less. - **Four tips + a bonus.** (1) Ask Claude to make and get you to approve a search *plan* before answering. (2) Demand links and citations for everything, including internal docs. (3) Ask for a summary of what it found across the whole search before the single answer, so the result is auditable. (4) Ask it to flag gaps, conflicts, and inconsistencies, especially across multiple connectors. (5 bonus) Unselect connectors you don't need — it saves tokens and reduces the security surface (the "lethal trifecta"). - **Human role.** AI is a "cognitive exoskeleton," an amplifier not a replacement. Judgment lives in interpreting what the data means and noticing the missing piece. - **Uncertainties.** The 0.5%–5% figure is Forte's own testing, not a published Anthropic spec; sampling behavior likely varies by connector, workspace size, and query. He scopes the talk to *official* connectors and defers the "lethal trifecta" and file-system context to a longer membership version and a blog post. ### 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=vqu4tNstz4k?t=0) Setup — the deep dive - People are handing Claude notes, calendar, email, even bank logins. - Forte spent a month on a deep dive; claims more nuance than anyone online covers. #### ⏱️ [00:00:20](https://www.youtube.com/watch?v=vqu4tNstz4k?t=20) The nervous-system metaphor - Brain is a closed black box; only the nervous system/senses move info in or out. - Connectors extend the central repository outward — highways of information both ways. - Eliminates hand-curated copy-paste-reformat context work. #### ⏱️ [00:01:30](https://www.youtube.com/watch?v=vqu4tNstz4k?t=90) Not infallible + human judgment - The AI nervous system can get overloaded, overwhelmed, and forget, just like ours. - Judgment lives in interpretation and in noticing/filling gaps. - "AI is a cognitive exoskeleton… an amplifier of you, not a replacement of you." #### ⏱️ [00:02:40](https://www.youtube.com/watch?v=vqu4tNstz4k?t=160) Scope — official connectors - Official = built by Anthropic or by the external platform. Third-party/open-source/custom exist but are out of scope. - Full 1-hour version is in the membership. #### ⏱️ [00:03:10](https://www.youtube.com/watch?v=vqu4tNstz4k?t=190) The deeper truth — sampling - Adding a Notion connector takes ~30 seconds; people then assume "Claude knows my Notion." Not true. - Workspaces are too big for the context window; Claude samples via targeted searches. - Tests show it reads ~0.5%–5% of available content; potentially major missing pieces. - Google Drive test: it admitted favoring recently modified files, ignoring years-old files. #### ⏱️ [00:04:40](https://www.youtube.com/watch?v=vqu4tNstz4k?t=280) The illusion of authority - A gut-instinct candidate answer is not the authoritative final answer. - Accuracy-critical tasks → check Claude's work *more*, not less. #### ⏱️ [00:05:20](https://www.youtube.com/watch?v=vqu4tNstz4k?t=320) Tips 1–4 - Tip 1: have Claude make and get approval for a search *plan* first. - Tip 2: require links/citations for everything, internal docs included. - Tip 3: ask for a broad summary of findings before the single answer (auditability). - Tip 4: ask it to flag gaps, conflicts, inconsistencies — especially across connectors. #### ⏱️ [00:07:30](https://www.youtube.com/watch?v=vqu4tNstz4k?t=450) Bonus tip 5 — security - Plus button → connectors → unselect any not needed for the task. - Reasons: saves tokens, and reduces the security surface (the "lethal trifecta," expanded in the blog post). - Teaser: files/folders on your computer are the next context frontier. ## 📚 References - [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) — open standard behind connectors. - [Building a Second Brain](https://www.buildingasecondbrain.com/) — Forte's PKM methodology and book. - [Simon Willison — "the lethal trifecta"](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/) — private data + untrusted content + exfiltration in one session. - [Anthropic — Claude connectors / directory](https://www.anthropic.com/) — official first-party and platform-built connectors. ## 🗣️ Paraphrase - Connectors act like a nervous system for an AI second brain: they wire a central store of intelligence to your external tools so it can sense context and act, replacing tedious manual copy-paste. But the nervous system isn't perfect — it overloads and forgets — so human judgment still owns interpretation and gap-spotting. - Adding a connector doesn't teach the model your whole workspace; it lets the model take small, recency-biased samples (Forte measures 0.5%–5%) and report a confident-looking slice. That confidence is the trap, so plan the search, demand citations, request a broad auditable summary, ask it to surface conflicts, and disable unused connectors for tokens and security. ## 📰 Newsletter Ever added a Notion or Google Drive connector and thought, "great, now the AI *knows* my stuff"? Tiago Forte spent a month testing that assumption and found the AI reads maybe 0.5%–5% of what's there — favoring whatever you touched last week. For WordPress folks wiring Claude into docs, Trac, or a Notion knowledge base, the lesson travels: your AI second brain is *sampling*, not *reading*. Make it show its work — a search plan you approve, citations for internal docs too, a broad "here's what I found" summary before the one answer, and a flag for any conflicts. Then trim your connector list: fewer connected sources means fewer tokens burned and a smaller "lethal trifecta" attack surface. Treat AI as a cognitive exoskeleton, not an oracle — verify more when it matters, not less. ## ✅ Action Items #### DevRel - Developer advocacy: build a "connectors reality check" demo showing the 0.5%–5% sampling gap live against a real workspace. - Education: add a workshop module on the four-tip verification checklist (plan → cite → summarize → flag conflicts). - Developer experience: publish a copy-paste "master prompt" that forces a search plan and citations before answers. - Developer marketing: a short-form clip on "your AI samples, it doesn't read" as a myth-buster. - Community management: N/A — no moderation angle in this source. - Internal/external/cross-project: align training and docs teams on one canonical connector-verification prompt. #### Personal knowledge work - Add the five-step checklist to the Obsidian PARA "AI-assisted retrieval" note; wire it into the Readwise→Obsidian pipeline as a pre-answer gate. - Audit active Claude connectors monthly; disable any not needed per session (token + security hygiene). - When querying the vault via any AI connector, always request citations back to note paths and flag missing coverage. #### Training (WordPress) - Lesson plan: "Verifying AI answers over internal docs" using Forte's citation-and-summary pattern. #### OSPO - N/A — no open-source governance, licensing, or SBOM content in this source. #### Community / The WPCC / FAIR - N/A — no community-program, funding, or federation content in this source. ## 🧪 Hypotheses - H1: For workspaces above a size threshold, connector answer accuracy drops as sampling coverage falls below a fixed percentage band (~0.5%–5%). - H2: Requiring a pre-approved search plan increases coverage breadth versus asking for a direct answer. - H3: Recency bias means older-but-relevant documents are systematically under-retrieved regardless of query specificity. - H4: Requiring citations reduces confident-but-wrong answers (fewer unsupported assertions surviving to the final answer). - H5: Reducing active connectors per session lowers both token spend and successful prompt-injection/exfiltration attempts. - H6: A broad "summary of findings before the single answer" step raises user-detected error rates (better auditability). ### 📊 Methodology - Forte's approach was a month-long self-experiment: connect official connectors across notes/calendar/email/Drive, then run "a series of tests" probing what Claude actually read, including directly asking Claude what percentage of files it examined. - To test H1–H3: seeded workspaces of varying sizes and document ages; measure retrieved-vs-available coverage and relevance, controlling for query type. - To test H4/H6: A/B prompts (direct answer vs plan+cite+summary) scored by independent graders for unsupported claims and caught errors. - To test H5: red-team sessions varying the number of active connectors, measuring injection success and token usage. ### ❓ Questions - What exactly drives the sampling percentage — workspace size, connector API limits, context budget, or query phrasing? - Does the recency bias come from the connector's search ranking or from Claude's own selection? - How much does the plan-approval step actually widen coverage in practice? - Which connectors leak the most in a "lethal trifecta" configuration? ### 🔮 Future Research - Reproduce the 0.5%–5% figure across Notion, Drive, Gmail, and Slack connectors with published methodology. - Build an open benchmark for "retrieval coverage" of AI connectors over private corpora. - Study whether citation-forcing prompts measurably cut hallucination over internal docs. ### 💡 Implications - **Practice:** Treat connector answers as leads, not verdicts; bake verification into any AI-over-docs workflow. - **Policy/security:** Minimize active connectors per session; document the lethal-trifecta risk in org AI guidelines. - **Education:** Onboarding for AI knowledge tools must teach the sampling model, not the "it reads everything" myth. - **Funding/tooling:** Demand for retrieval-coverage observability and citation-first UX in connector products. ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid Grade ~8–9 (conversational, short clauses, everyday vocabulary with a few technical terms like "context window" and "connectors"). #### 👀 Visual learners - Present: spoken metaphor only; no on-screen diagrams in transcript (a QR code is referenced). - Improve: add a nervous-system schematic and a "assumed vs actual coverage" bar chart. #### 🔊 Aural learners - Present: full narrated video; clear spoken pacing. - Improve: publish a 2–3 minute audio summary of the five-step checklist. #### 🗣️ Verbal learners - Present: strong analogies ("nervous system," "cognitive exoskeleton," "reconnaissance"). - Improve: a short glossary (connector, context window, sampling, lethal trifecta). #### 🧩 Kinesthetic learners - Present: actionable tips users can try immediately. - Improve: a downloadable checklist and a hands-on lab against a sample workspace. #### 🧠 Logical learners - Present: cause→effect chain (sampling → illusion of authority → verify more). - Improve: a decision tree for "when to trust vs verify a connector answer." #### 🧑‍🤝‍🧑 Social learners - Present: none in-transcript. - Improve: discussion prompts and a shared prompt library. #### 🧘‍♀️ Solitary learners - Present: self-paced video format. - Improve: reflection prompts on your own connector setup. #### 🌿 Naturalistic learners - Present: biological analogy (nervous system) ties to living systems. - Improve: extend the ecosystem analogy to data lifecycle/decay. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: N/A for video; applies to the companion blog post. - Improve: ensure the blog uses real headings and descriptive link text. #### Keyboard navigation - Present: standard YouTube player controls. - Improve: ensure the QR/description links are keyboard-reachable on the post. #### Contrast and readability - Present: unknown from transcript. - Improve: verify 4.5:1 for on-screen captions/QR overlay text. #### Text resizing and reflow - Present: N/A (video). - Improve: responsive blog post at 200% zoom. #### Media and captions - Present: auto-transcript exists (this source); "[music]" cues noted. - Improve: publish a clean human-corrected caption track and a downloadable transcript; the QR code needs a plain-text link alternative. #### Forms and interactions - Present: N/A. - Improve: label the checklist opt-in form and provide non-QR access. #### Standards reference - Map to WCAG 2.2 AA: 1.2.2 Captions, 1.2.5 Audio Description, 1.4.5 Images of Text (avoid QR-only paths), 2.4.4 Link Purpose, 1.4.3 Contrast. ### 🌍 DEIB Assessment #### Cultural breadth - Present: neutral, universal examples (notes, calendar, email). - Improve: examples from non-English workspaces and varied tools. #### Inclusive language - Present: plain, gender-neutral language. - Improve: define jargon ("lethal trifecta," "context window") inline. #### Neurodiversity support - Present: chunked tips, consistent numbering. - Improve: add headings/timestamps and an icon+text checklist. #### Representation - Present: single-presenter format; limited perspective diversity. - Improve: cite other practitioners and security researchers. #### Psychological safety - Present: reassuring framing ("I don't want to leave you with this pessimistic view"). - Improve: acknowledge users without paid membership access to the fuller version. ### ✅ Logical Validity #### Strong reasoning - Analogical (00:20): the nervous-system analogy is apt — it maps closed-system + I/O channels cleanly onto model + connectors. Clarify by naming where the analogy breaks (nerves don't "sample" or hallucinate). - Inductive/empirical (03:10): the 0.5%–5% claim is grounded in his own repeated tests and a direct admission from Claude about recency bias — reasonable inductive support, though sample and method are undisclosed. - Causal (04:40): sampling → illusion of authority → "verify more, not less" is a valid causal chain with an actionable conclusion. #### Fallacies - Appeal to authority / anecdote: "this really is the right answer" (about the 0.5%–5% range) asserts confidence without shared methodology. Fix: publish the test protocol and per-connector numbers. - Hasty generalization risk: generalizing from his workspaces to "most" workspaces. Fix: bound the claim by workspace size and connector type. - Teaser/appeal-to-scarcity: gating the "full nuance" behind membership and a QR checklist mildly pressures action. Fix: give the core checklist free. #### Tone of voice - Measured, reassuring, educational — "honest opinion," "I don't want to leave you with this pessimistic view." Confident but hedged ("I'm willing to bet"). ### 🛡️ DARVO Analysis - N/A — this is an explanatory/educational monologue with no accusation, defense, or victim/offender dynamic. No denial, attack, or reversal present. ## 🎯 Conclusion - Bottom line: connectors give your AI second brain a nervous system, but it *samples* your data (≈0.5%–5%, recency-biased) rather than reading it — and the real risk is mistaking a confident partial answer for an authoritative one. - Next steps for readers: adopt the five-step checklist (plan → cite → summarize → flag conflicts → disable unused connectors), verify more on accuracy-critical tasks, and mind the lethal-trifecta security surface. AI amplifies your judgment; it doesn't replace it. ## 🔄 Refactor Appendix — I Connected Claude to Everything. Then I Saw the Gap. > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. > > Konik claim-header rules: present tense, declarative, ≤10 words when > possible, searchable, unique, mirrors annotation focus if present. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Atoms (H3) #### [H1] Connectors are a nervous system for AI > [!quote] ID:rw_c1h1 > Adding connectors [music] to your claw is kind of like giving your AI second brain [music] a nervous system. **Context & use**: - ~00:20 — the framing metaphor for the whole video. - Names the mental model PKM practitioners need: connectors are I/O channels, not knowledge transfer. **Tags**: #xref **Highlight**: [view highlight](readwise://)   #### [H2] Connectors sample 0.5%–5% of a source > [!quote] ID:rw_c1h2 > the content that it's actually reading through that connector is somewhere between 0.5% and 5% of everything that's available. **Context & use**: - ~03:10 — the empirical core of "the gap." - Kills the "the AI now knows my Notion" assumption; drives the verify-more conclusion. **Tags**: #xref #articleseed **Highlight**: [view highlight](readwise://)   #### [H3] The danger is the illusion of authority > [!quote] ID:rw_c1h3 > The real danger here is the illusion of authority. **Context & use**: - ~04:40 — reframes the risk from "wrong answer" to "confident partial answer." - Justifies checking Claude's work more, not less, on accuracy-critical tasks. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://)   ### 🧠 Permanent Candidates (H4) #### [P1] AI retrieval samples; humans own the gap **Claim** (one sentence, Courtney's voice): An AI connector returns a recency-biased sample of a source, so the human's job is to notice and fill the coverage gap it hides. **Evidence pointer**: [[#H2]] > [!quote] > You have to remember that AI is a cognitive exoskeleton. [music] It is an amplifier of you, not a replacement of you. **Scope & conditions**: Holds for large workspaces where the corpus exceeds the context window. Breaks for tiny sources that fit whole, where sampling ≈ full read. **Links**: - [[Building a Second Brain]] — because extends - [[The lethal trifecta]] — because contrasts on {risk dimension: coverage vs security} - [[Readwise to Obsidian retrieval workflow]] — because applies-in {vault Q&A}   ### ✏️ Fleeting Queue - [F1] Draft the canonical "connector verification" master prompt (plan → cite → summarize → flag) — from this video #fleeting #inbox - [F2] Open question: does recency bias originate in the connector's ranking or Claude's selection? — 03:10 #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Add the five-step connector-verification checklist to the vault AI-retrieval note — #things - [ ] Audit and trim active Claude connectors this month for token + security hygiene — #things  📄 **Full text:** [[Full Document Contents/Articles/I Connected Claude to Everything. Then I Saw the Gap..md|I Connected Claude to Everything. Then I Saw the Gap. by Tiago Forte]] ## Highlights - Adding connectors to your claw is kind of like giving your AI second brain a nervous system. Think about the human brain. It's really kind of this black box. It has no built-in way of getting any sort of sensation or information or communication in or out. It's completely closed off. The only way that it can communicate or interact or sense anything in the external environment, including from the rest of the body, is through the nervous system or through the senses. ([View Highlight](https://read.readwise.io/read/01kwdg1qnhw7g6yvspn8177c6t)) - Note: The framing metaphor: connectors are a "nervous system" for your AI second brain — the channels that bring external data in and send observations/actions back out. - Tags: second-brain, ai-connectors, atom/literature/h1 - It's the judgment to notice when there's a gap or missing piece and to fill it in. It's in the nuances and the details and the subtlety that AI still often misses. You have to remember that AI is a cognitive exoskeleton. It is an amplifier of you, not a replacement of you. There is a whole world of connectors out there. ([View Highlight](https://read.readwise.io/read/01kwdg200y79ba210e8jxwmb7q)) - Note: Zettel: "AI is a cognitive exoskeleton... an amplifier of you, not a replacement of you." The scarce human role is the judgment to notice the gap and fill it. Echoes Newport and the bottleneck-is-judgment thread. - Tags: pkm, second-brain, atom/zettel/z1 - It is sort of doing a reconnaissance. It's doing a series of targeted searches and then pinpointing what it thinks, what it assumes is the most relevant little pieces and reporting back its answer based on those. I've run a series of tests with Claude and found that in most cases, the content that it's actually reading through that connector is somewhere between 0.5% and 5% of everything that's available. ([View Highlight](https://read.readwise.io/read/01kwdg1tv1rrh6yvsda595fqtk)) - Note: The gap, quantified: a connector doesn't read the source — it *samples* it ("reconnaissance"). Forte's tests: Claude actually reads "somewhere between 0.5% and 5% of everything available," so major pieces can be silently missing. - Tags: atom/literature/h1, ai-workflow, ai-connectors - It was looking at files that I had touched or modified in recent days or weeks or at most months. But I have files in there going back many, many years. The real danger here is the illusion of authority. It's Claude's just first kind of almost gut instinct for what it thinks might be a good answer, not the end all beall totally authoritative final answer. ([View Highlight](https://read.readwise.io/read/01kwdg1xe4swc6m0nxsat783jm)) - Note: "The real danger here is the illusion of authority" — a sampled, recency-biased answer (recent files only) arrives sounding final. The more confident it sounds, the more important it is to check its work. - Tags: knowledge-management, ai-workflow, atom/literature/h1