# I Connected Claude to Everything. Then I Saw the Gap. > [!question] Lingering Questions > - What exactly drives the sampling percentage — workspace size, connector API limits, context budget, or query phrasing? > - Does the recency bias Forte found come from the connector's own search ranking or from Claude's selection layer on top of it? > - Which connectors leak the most in a "lethal trifecta" configuration (private data + untrusted content + exfiltration path in one session)? > [!tip] Article Ideas > - Publish the canonical "connector verification" master prompt (plan → cite → summarize → flag) as a standalone piece. #articleseed > - Add the five-step checklist to the vault's own AI-assisted-retrieval note as a pre-answer gate. #addendum > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Sampling a Fraction of a Source Produces False Authority Over the Whole]] — the existing Permanent note for this exact source; this Literature note is the missing intake record for it. > - [[3. Resources/Permanent Notes/A Trusted Intermediary Makes a False Claim Durable]] — a confident partial answer from a connector is exactly the trusted-intermediary failure mode that note describes. > - [[3. Resources/MOCs/Personal Knowledge Management MOC]] > [!quote] Vocabulary > - **Lethal trifecta** (Simon Willison's term, cited in the video) — private data + untrusted content + an exfiltration path present in the same AI session; the security risk of stacking too many active connectors at once. > - **Cognitive exoskeleton** — Forte's framing for AI's proper role: an amplifier of the user's own judgment, not a replacement for it. ## Metadata - Source: [I Connected Claude to Everything. Then I Saw the Gap.](3.%20Resources/Readwise/Articles/I%20Connected%20Claude%20to%20Everything.%20Then%20I%20Saw%20the%20Gap..md) - Author: Tiago Forte - Category: Article (YouTube transcript) ## Author's thesis Adding an official Claude connector does not mean the model reads your whole workspace — it means the model can *sample* it. Forte's month-long self-experiment measured actual coverage at roughly 0.5%–5% of available content, recency-biased toward whatever was touched most recently. The danger isn't that the sampled answer is wrong; it's that a partial, recency-biased sample gets delivered in the same confident register as a complete answer — "the illusion of authority." The fix is procedural: force a search plan, demand citations, request a pre-answer coverage summary, ask the model to flag gaps and conflicts, and disable unused connectors. ## Key arguments - Connectors are a nervous system, not a memory transplant — they extend what a model can sense and act on, but the model still has to choose what to look at inside a workspace too large for its context window. - Measured coverage sits at 0.5%–5% of available content in Forte's tests, and skews toward recently modified files regardless of relevance. - The real risk is presentational, not factual: a confident-sounding partial answer is mistaken for an authoritative complete one, which inverts the normal instinct to verify less as confidence rises. ## Claim atoms ### Connectors are a nervous system for AI > [!quote] ID:rw_c1h1 > Adding connectors to your Claude is kind of like giving your AI second brain a nervous system. - **Where**: ~00:20, the framing metaphor for the whole video. - **Why**: Names the mental model PKM practitioners need — connectors are I/O channels, not knowledge transfer. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### 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. - **Where**: ~03:10, the empirical core of "the gap." - **Why**: Kills the "the AI now knows my Notion" assumption; drives the verify-more conclusion. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> ### The danger is the illusion of authority > [!quote] ID:rw_c1h3 > The real danger here is the illusion of authority. - **Where**: ~04:40, reframing the risk from "wrong answer" to "confident partial answer." - **Why**: Justifies checking Claude's work more, not less, on accuracy-critical tasks. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> ## Permanent candidates - [[3. Resources/Permanent Notes/Sampling a Fraction of a Source Produces False Authority Over the Whole]] — already exists for this exact source, covering all three atoms above. Not duplicated here. ## Processing checklist - [ ] Draft the canonical "connector verification" master prompt (plan → cite → summarize → flag) referenced in the video. - [ ] Audit and trim active Claude connectors for token and security hygiene, per the video's bonus tip. - [x] Connect into the relevant MOC — [[3. Resources/MOCs/Personal Knowledge Management MOC]] ## Links - [[3. Resources/MOCs/Personal Knowledge Management MOC]] — the connector-sampling gap is a direct PKM-tooling concern for how the vault itself gets queried by AI.