Claude Cowork Finally Clicked When I Learned This  🔗 **Source:** [Claude Cowork Finally Clicked When I Learned This](https://www.youtube.com/watch?v=szWBwfPX30M) > [!note]+ Document Notes > tags: [claude-cowork, ai-agents, workflow-automation, ai-memory-systems, personal-productivity] > author: Paul J Lipsky > source: Paul J Lipsky (YouTube) > reference: https://www.youtube.com/watch?v=szWBwfPX30M > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Claude Cowork Finally Clicked When I Learned This > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - Paul J Lipsky explains why Claude Cowork (Anthropic's desktop, local-file-aware agent mode) felt confusing at first and what made it click: local file access, local file creation, and a persistent `CLAUDE.md` instructions file plus a `memory.md` file. - The video is a paid sponsorship — Anthropic reached out to sponsor a video Lipsky says he "was going to make anyway" — disclosed openly near the top. - Core argument: Cowork's real advantage isn't file access itself, it's that a folder-level `CLAUDE.md` lets a user write standing instructions once (who they are, how they want Claude to behave, folder structure, output rules, named "skills") so they never have to re-explain context on every task. A companion `memory.md` file lets Claude accumulate learned preferences over time, readable and editable as plain text. - Key stakeholders: individual knowledge workers and small-business operators (teacher example used as the demo case), Anthropic (sponsor, product), and Lipsky himself as a creator monetizing an affiliate/sponsor link. - Risks and tradeoffs: sponsored content has an inherent incentive to undersell friction or overstate the "gamechanger" framing; the demo (grading exams) is a best-case, single-user scenario; no discussion of privacy, data handling of local files, cost, or failure modes when Cowork misreads a `CLAUDE.md` or when memory drifts/becomes stale or contradictory. - Bottom line: the tutorial's practical claim is sound and directly verifiable by any viewer — a folder-scoped `CLAUDE.md` plus `memory.md` reduces repeated context-setting for agentic coding/knowledge tools generally, not just Cowork specifically. The sponsorship disclosure is present and prominent, which mitigates (but doesn't eliminate) the promotional bias. ## 📖 Main Content - **Framing device**: Lipsky structures the video around "three things" that made Cowork click for him: (1) local file *access* without manual uploads, (2) local file *creation* without manual downloads, and (3) the `CLAUDE.md`/`memory.md` persistent-instructions system. The first two are presented as necessary-but-insufficient; the third is presented as the actual differentiator. - **Chat vs. Cowork distinction**: In Claude Chat, files must be attached per conversation, and any generated output (e.g., an Excel file) is created in the chat and must be manually downloaded and relocated. In Cowork, the user points the tool at a folder already on their computer, and Cowork reads from and writes to that folder directly — "with chat, you have to bring the work to Claude. With co-work, it works where you already are." - **The "Cowork Hub" pattern**: Lipsky recommends creating a single top-level folder (his example: "Cowork OS," optionally inside iCloud Drive for backup) that Cowork is pointed at for every task, with subfolders for specific work streams (his demo: a "graded exams" subfolder). This centralizes context so Cowork can find related files without being told where they are each time. - **Worked example**: A teacher's folder of graded exams is handed to Cowork with a single dictated prompt asking for average/median grade, patterns in wrong answers, and an HTML report. Cowork returns both an in-chat summary and a saved HTML file inside the folder — persisting as an artifact the user can revisit, rather than a disposable chat answer. - **`CLAUDE.md` mechanics**: Defined as "a set of standing instructions that co-work reads automatically every single time it starts a task in your folder." It's plain text, human-readable, with no special syntax required. Content categories demonstrated: (a) who the user is and their role, (b) tone/behavior rules ("a helpful colleague, not a corporate bot," keep responses brief, ask clarifying questions, treat grading as a suggestion not a final grade), (c) a folder-structure map so Claude can navigate without being told, (d) output-formatting rules (outputs go in a "reports" folder, always as HTML), (e) named "skills" — reusable step-by-step task instructions invoked by name (e.g., "exam report skill," "parent update skill") rather than loaded all at once, and (f) blanket safety rules (never delete files, ask before large changes, keep summaries short). - **Skills vs. `CLAUDE.md` distinction**: `CLAUDE.md` is what Claude should "always know"; skills are specific workflows loaded only when a task calls for them, avoiding context bloat from loading every possible procedure on every task. - **`memory.md` mechanics**: A second plain-text file where Cowork writes down things it learns about the user's preferences over time (his example: "remember that I like my reports sorted by class period"). Unlike Claude Chat's memory, which Lipsky characterizes as a "black box," `memory.md` is inspectable and editable text that persists across sessions (next week, next month). - **Setup mechanism**: Rather than hand-writing these files, Lipsky provides a prompt (linked in a public Google Doc, no email gate) that has Claude interview the user question-by-question and generate both files automatically inside the chosen folder. - **Closing synthesis**: Lipsky reframes the three points as: Cowork works inside local folders; it creates files there (no upload/download cycle); and with `CLAUDE.md`/`memory.md` in place, the user never re-explains context, and instead precisely controls how the agent behaves. His claim: "it stops being a chatbot you talk to and becomes an agent you've tailored to do exactly what you want." - **Uncertainties/contested claims not addressed in the video**: How Cowork handles conflicting instructions between `CLAUDE.md` and in-task prompts; token/context-window costs of large folders or bloated memory files over time; multi-user or team folder scenarios; what happens if a `CLAUDE.md` file becomes outdated or self-contradictory; data privacy implications of granting broad local folder access to a cloud-connected agent. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=szWBwfPX30M?t=0) Intro and sponsorship disclosure - Lipsky states Cowork's power didn't click until he learned the three things covered in the video. - Explicit sponsorship disclosure: Anthropic reached out to sponsor a video he says he "was going to make anyway." - Quote: "I do want to let you know though that Anthropic reached out to me recently about sponsoring this video." #### ⏱️ [00:00:53](https://www.youtube.com/watch?v=szWBwfPX30M?t=53) Chat vs. Cowork - Claude Chat requires manually attaching files per conversation and manually downloading/relocating any generated output. - Cowork reads directly from a chosen folder and writes output back into it. - Quote: "With chat, you have to bring the work to Claude. With co-work, it works where you already are." #### ⏱️ [00:02:10](https://www.youtube.com/watch?v=szWBwfPX30M?t=130) Building a "Cowork Hub" - Recommends one central folder always granted to Cowork (example: "Cowork OS," placed in iCloud Drive for backup). - Walkthrough of the folder picker: choose folder → new folder → name it → create → open → "always allow." #### ⏱️ [00:03:27](https://www.youtube.com/watch?v=szWBwfPX30M?t=207) Worked example: grading exams - Subfolder ("graded exams") created inside the hub; scanned exam files dropped in. - Dictated prompt asks for average grade, median grade, and recurring wrong-answer patterns, plus an HTML report. - Result: average grade ~72%, range 30–100, breakdown of common struggle points and per-student scores, saved as a persistent HTML file in the folder. #### ⏱️ [00:05:38](https://www.youtube.com/watch?v=szWBwfPX30M?t=338) Introducing CLAUDE.md - States that most people understand file access/creation but miss why `CLAUDE.md` matters. - Defines it as instructions read automatically at the start of every task in that folder. - Quote: "A claude.md file is a set of standing instructions that co-work reads automatically every single time it starts a task in your folder." - Emphasizes it's plain text with "nothing magical," "no code," "no special settings or format." #### ⏱️ [00:07:41](https://www.youtube.com/watch?v=szWBwfPX30M?t=461) Walking through a real CLAUDE.md - Shown sections: user identity/role (a high-school history teacher), behavior rules (helpful colleague tone, brevity, clarifying questions, grading as suggestion only), folder-structure map, output rules (reports folder, HTML format), named skills (exam report skill, parent update skill), and blanket safety rules (never delete files, ask before big changes, keep summaries short). - Distinguishes `CLAUDE.md` (always-loaded context) from skills (loaded only on demand) to avoid overloading context. - Quote: "The instructions live in the folder, not in my prompt." #### ⏱️ [00:10:49](https://www.youtube.com/watch?v=szWBwfPX30M?t=649) memory.md and the "click" moment - Introduces `memory.md` as where Cowork records what it learns about the user's preferences over time (example: report sort order by class period). - Contrasts with Claude Chat memory, calling it a "black box," versus `memory.md` being inspectable, editable plain text. - Quote: "Between these two files, your assistant actually gets better the more you use it." #### ⏱️ [00:12:13](https://www.youtube.com/watch?v=szWBwfPX30M?t=733) Setup walkthrough - Points to a free, ungated Google Doc (linked in the video description) containing a prompt that has Claude interview the user to auto-generate both files. - Asks viewers for a thumbs-up in exchange for the free resource — a soft engagement ask, not a paywall. #### ⏱️ [00:13:35](https://www.youtube.com/watch?v=szWBwfPX30M?t=815) Closing synthesis and call to action - Recaps the three points: works in local folders, creates files there, and with `CLAUDE.md`/`memory.md` the user never re-explains context. - Quote: "It stops being a chatbot you talk to and becomes an agent you've tailored to do exactly what you want, the way you want it done." - Notes Cowork requires a paid Claude plan; link at top of description. ## 📚 References - [Claude Cowork (referral link used in video)](http://clau.de/paul-j-lipsky) - [Claude Cowork setup prompt (Google Doc, per video description)](https://docs.google.com/document/d/1fi1pSlkBjZzvZA9r2oa-Nwud_cjz48E8xYv5S05zVdc/edit?usp=sharing) - [Paul J Lipsky — YouTube channel](https://www.youtube.com/channel/UCmeU2DYiVy80wMBGZzEWnbw) - Anthropic (sponsor of this video; maker of Claude and Claude Cowork) ## 🗣️ Paraphrase - Cowork feels like a bigger leap than it first appears because the obvious differences — reading files straight from a folder and writing output back into that folder — are just table stakes. They remove the upload/download busywork of chat, but by themselves they don't explain why people who've tried Cowork describe it as transformative. - The actual unlock is a boring-looking text file. Put a `CLAUDE.md` in the folder you always point Cowork at, and every future task starts already knowing who you are, how you like things formatted, where things belong, and which named "skills" to run for recurring jobs. Pair it with a `memory.md` that Cowork updates as it learns your preferences, and the tool stops resetting to zero every session — you configure it once, in your own words, and it keeps working the way you specified. ## 📰 Newsletter - If you've been sitting on Claude Cowork wondering what the fuss is about, the answer isn't the file access — it's a plain-text `CLAUDE.md` file sitting in your working folder. Set one up (who you are, how you want output formatted, a map of your subfolders, a couple of named "skills" for repeat tasks) and Cowork stops needing a fresh briefing every single time. - Pair it with a `memory.md` companion file and the assistant actually accumulates your preferences over time — readable, editable, no black box. - Worth a look for anyone building personal or small-team AI workflows: WordPress folks managing recurring client reports, community managers running the same triage steps weekly, or anyone tired of re-pasting the same context into a chat window. - Heads-up: the source video is sponsored by Anthropic, so treat the enthusiasm level accordingly — the underlying `CLAUDE.md`/`memory.md` pattern is genuinely useful and easy to verify yourself in about ten minutes. ## ✅ Action Items #### DevRel - Developer advocacy: consider a short internal demo showing how a `CLAUDE.md`/`memory.md` pair reduces onboarding friction for any agentic coding tool, not just Cowork — this is a transferable pattern for docs and templates. - Developer experience: if maintaining developer-facing quickstarts, a "starter `CLAUDE.md`" template with named-skill placeholders could shortcut new-user setup the way Lipsky's Google Doc prompt does. #### OSPO - Not applicable — video contains no open-source governance, licensing, or compliance content to act on. #### Community - Community: worth sharing the `CLAUDE.md`/`memory.md` pattern in any community Slack/Discord focused on AI-assisted personal productivity, since it's tool-agnostic advice dressed up as a Cowork tutorial. #### WordPress - Not applicable — no WordPress-specific content in this video. #### The WPCC - Not applicable — no WPCC-relevant governance or funding content. ### FAIR - Not applicable — no FAIR/provenance/governance content; this is a consumer productivity tutorial. ## 🧪 Hypotheses - H1: Users who set up a folder-scoped `CLAUDE.md` before starting agentic-tool tasks report needing fewer clarifying turns per task than users who don't. - H2: A `memory.md`-style editable memory file produces higher user trust in an AI assistant's "remembered" preferences than an opaque, non-inspectable memory system, holding accuracy constant. - H3: Centralizing all agent-accessible files in a single top-level "hub" folder reduces task setup time compared to ad hoc, scattered folder access, for repeat workflows. - H4: Named "skills" (task-specific instruction sets invoked on demand) produce more consistent output formatting across repeated tasks than embedding all instructions directly in a single always-loaded context file. - H5: Viewers who received this content as sponsored are, on average, no less likely to find the technical claims (file access, `CLAUDE.md` behavior, memory persistence) accurate than viewers of an unsponsored equivalent — i.e., the sponsorship affects framing/enthusiasm more than factual content. ### 📊 Methodology - For H1/H3: A/B task-completion study with matched users given the same multi-file analysis task, one group with a pre-built `CLAUDE.md` and hub folder, one without; measure prompt turns to completion and time-to-first-usable-output. - For H2: Survey-based comparison of trust/satisfaction scores between users of editable-memory agent tools versus opaque-memory chat tools, controlling for task type and session frequency. - For H4: Compare output consistency (rubric-scored formatting/tone adherence) across tasks run with an all-in-one instructions file versus a `CLAUDE.md` + separate named-skills structure. - For H5: Independent fact-check of the video's specific technical claims (upload requirements, file creation location, `CLAUDE.md` read timing) against Anthropic's own Cowork documentation, scored for accuracy regardless of sponsorship status. ### ❓ Questions - What happens when `CLAUDE.md` instructions conflict with an in-task prompt — which one does Cowork prioritize? - How does Cowork handle a `CLAUDE.md` or `memory.md` file that has grown large or self-contradictory over months of use? - Is there a practical file-count or folder-size ceiling before Cowork's folder scanning becomes slow or unreliable? - What are the data-privacy and access-control implications of granting a cloud AI product persistent access to a local folder, especially for shared/team folders? - Does Cowork behave differently across operating systems (the video only demonstrates macOS)? ### 🔮 Future Research - Comparative study of `CLAUDE.md`-style standing-instruction files across competing agentic tools (e.g., other desktop AI agents) to see whether the pattern generalizes or is Cowork-specific. - Longitudinal tracking of `memory.md` file drift/staleness over months of real use, and what maintenance cadence keeps it accurate. - Investigation of team/multi-user folder scenarios where multiple people share one `CLAUDE.md` — whose preferences take precedence, and how conflicts get resolved. ### 💡 Implications - Practice: knowledge workers adopting agentic AI tools should treat "standing instructions + inspectable memory" as a setup step, not an afterthought — the pattern in this video is directly reusable outside Cowork (e.g., in coding-agent CLAUDE.md conventions Courtney's own projects already use). - Policy: as agent tools gain persistent local folder access, organizations may need lightweight policies on what can live in an agent-accessible hub folder (PII, credentials) — this video doesn't address that risk surface at all. - Security: any workflow that grants broad folder access plus file-creation rights to a cloud-connected agent introduces a new attack surface (e.g., a compromised or manipulated `CLAUDE.md` altering agent behavior silently) that the video doesn't discuss. - Education: the "interview-driven setup prompt" approach (having the AI ask you questions to build its own config) is a reusable onboarding pattern for non-technical users of any configurable AI tool. - Funding models: the transparent sponsorship disclosure model here (up-front, plain statement, still delivers a video the creator says he'd have made anyway) is a reasonable practice worth citing when discussing creator-sponsorship norms in AI content. ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid Grade approximately 6-7 (conversational spoken transcript; short-to-medium sentences, mostly one- and two-syllable words, occasional technical terms like "context" and "folder structure" that nudge it slightly above pure casual speech). Estimate based on manual sampling of sentence length (~15-20 words average) and syllable density (mostly monosyllabic/disyllabic vocabulary) rather than a computed score, since only the plain-text transcript was available. #### 👀 Visual learners - Present: the video itself (not captured in transcript) demonstrably shows on-screen folder navigation, the `CLAUDE.md` file opened in Preview, and the generated HTML report — but none of that visual layer is available from transcript text alone. - Improve: an accompanying screenshot sequence or annotated diagram of the Cowork Hub folder structure would make the visual walkthrough usable independent of watching the video. #### 🔊 Aural learners - Present: the entire piece is narrated, conversational, first-person spoken explanation — strong native fit for aural learners. - Improve: none needed; format already serves this learning style directly. #### 🗣️ Verbal learners - Present: the "bring the work to Claude" vs. "it works where you already are" analogy, and the chat/Cowork comparison, are clear verbal framings. - Improve: a short glossary of terms (`CLAUDE.md`, skill, memory.md, hub folder) would reinforce the verbal explanations already present. #### 🧩 Kinesthetic learners - Present: the video includes a literal step-by-step setup (choose folder → new folder → name → create → open → always allow) and a copy-paste prompt for hands-on replication. - Improve: a downloadable checklist mirroring those exact steps would let kinesthetic learners follow along without pausing/rewinding video. #### 🧠 Logical learners - Present: the "three things" framework and the clean two-file (`CLAUDE.md` / `memory.md`) model are structured, logically sequenced explanations. - Improve: a simple flowchart of "task starts → Cowork reads CLAUDE.md → loads relevant skill (if named) → checks memory.md → executes" would make the causal sequence explicit rather than implied. #### 🧑‍🤝‍🧑 Social learners - Present: minimal — the video is a solo tutorial with no discussion prompts, comments engagement beyond a thumbs-up ask. - Improve: could invite viewers to share their own `CLAUDE.md` examples in comments, turning a one-way tutorial into a community knowledge exchange. #### 🧘‍♀️ Solitary learners - Present: the linked Google Doc setup prompt supports self-paced, independent setup without needing to watch further. - Improve: a self-check list ("did your CLAUDE.md end up with these five sections?") would help solitary learners verify their own setup without feedback from others. #### 🌿 Naturalistic learners - Not applicable to this source — no sustainability, environmental, or ecosystem-analogy content present in the video. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: cannot be assessed from transcript alone; depends on the actual video's caption/description structure, which is outside this text's visibility. - Improve: ensure YouTube auto-captions are corrected/reviewed for accuracy (product-specific terms like "Cowork" and "claude.md" are common auto-caption failure points, evident in this very transcript's inconsistent capitalization and phrasing, e.g. "Claw.md" and "Cowork" rendered as "Co-work"). #### Keyboard navigation - Not applicable to this source — the video demonstrates the creator's own local file navigation (Mac Finder/Preview), not a web interface under this analysis's control. #### Contrast and readability - Not applicable to this source — visual contrast of on-screen UI cannot be assessed from a transcript-only source. #### Text resizing and reflow - Not applicable to this source — no reflowable text interface in scope. #### Media and captions - Present: a transcript exists (this document is built from it), meaning the video is at least auto-captioned. - Improve: the auto-generated transcript shows clear errors ("Claude Co-work" transcribed inconsistently, "clawed" for "Claude," "coowwork" for "Cowork," "MD" spoken but likely meant "claude.md"/"memory.md") — a creator-provided, corrected caption file or downloadable transcript would materially improve accessibility and searchability. #### Forms and interactions - Not applicable to this source — no forms or interactive elements in the video content itself. #### Standards reference - Gaps map to WCAG 2.2 AA 1.2.2 (Captions, Prerecorded) in that auto-captions alone, with visible transcription errors on key product terms, fall short of the "accurate" bar implied by that success criterion; a human-reviewed transcript would close the gap. ### 🌍 DEIB Assessment #### Cultural breadth - Not applicable to this source — the video uses a single American teacher persona as its only example; no multi-region or multilingual content present to evaluate. #### Inclusive language - Present: language throughout is plain, jargon is defined on first use ("skill," `CLAUDE.md`, `memory.md`), and the tone is casual/accessible rather than corporate-technical. - Improve: could explicitly note that the demonstrated example (Mac, iCloud Drive, Finder) assumes a specific OS/ecosystem; a line acknowledging Windows/Linux equivalents would broaden applicability. #### Neurodiversity support - Present: the video is chunked into clearly signposted chapters (per its own description timestamps: Intro, Chat vs. Cowork, Cowork Hub, Files, CLAUDE.md, Instructions, Memory, Set Up, Thoughts), which supports predictable, navigable structure. - Improve: an on-screen recap card at each chapter transition (visible in video, not verifiable from transcript) would reinforce the chunking already present in the structure. #### Representation - Not applicable to this source — a single presenter and a single hypothetical persona (a teacher) are used; no varied identities or roles are otherwise depicted or discussed. #### Psychological safety - Present: repeated emphasis that "grading is always a suggestion, not a final grade, because the teacher makes the final call" models a safe, human-in-the-loop framing for AI-assisted decision-making. - Improve: could generalize this principle explicitly beyond the teacher example (e.g., stating it as a general rule for any high-stakes Cowork use case) rather than leaving it implicit in one example. ### ✅ Logical Validity #### Strong reasoning - The chat-vs-Cowork comparison (upload/download friction vs. direct folder access) is a clear, falsifiable, and easily-verified causal claim — any viewer with Cowork access can replicate it in minutes, which is a strength for a product-explainer format. - The "CLAUDE.md always known, skills loaded on demand" architecture claim is explained with a specific mechanism (avoiding loading all skills at once) rather than asserted vaguely — this is a reasonably strong inductive/mechanistic explanation, though unverified against Anthropic's actual technical documentation within this video alone. #### Fallacies - Appeal to personal testimony without independent verification: "I've been using Co-work for months now and it has been such a gamechanger" (~00:00:32) is an anecdotal claim presented as a general truth about the product; the fix is to pair it with independent user reports or Anthropic's own usage data rather than relying on one creator's experience. - Mild false-dichotomy framing in the closing synthesis — "it stops being a chatbot you talk to and becomes an agent" implies a clean binary between "chatbot" and "tailored agent" when in practice the difference is one of configuration depth, not category; the fix would be to frame it as a spectrum (more/less configured) rather than a hard transformation. #### Tone of voice - Enthusiastic, first-person, tutorial-style, consistently encouraging ("that's the real gamechanger," "this is the part that made Co-work finally click for me") — tone matches a sponsored product-explainer genre and is transparent about that framing rather than concealing it. ### 🛡️ DARVO Analysis - Not applicable to this source — DARVO (Deny, Attack, Reverse Victim/Offender) is a framework for analyzing responses to accusations of wrongdoing. This video contains no accusation, denial, attack on an accuser, or victim/offender reversal; it is a straightforward product tutorial with an upfront sponsorship disclosure, not a defensive or reputational-crisis communication. ## 🎯 Conclusion - The bottom line holds up under scrutiny: Cowork's headline features (local file access, local file creation) are real but incremental conveniences, while the actual behavior change comes from a folder-scoped `CLAUDE.md` (standing instructions, always read first) paired with an editable `memory.md` (accumulated, inspectable preference memory). That two-file pattern is a genuinely reusable idea for anyone configuring any agentic AI tool, not just Cowork. - Practical next step for a reader: before adopting Cowork (or any similar local-folder agent tool), draft a `CLAUDE.md` covering identity/role, tone rules, folder-structure map, output-format rules, and 2-3 named skills for recurring tasks — then let the tool's memory file accumulate the rest, reviewing it periodically rather than trusting it silently. - Second next step: treat the sponsorship disclosure as a cue to verify the more enthusiastic claims (e.g., "gamechanger," "months of use") against independent sources or one's own trial, rather than as a reason to distrust the technical walkthrough, which is specific and independently testable. ## 🔄 Refactor Appendix — Claude Cowork Finally Clicked When I Learned This > 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] Cowork works where you already are, not vice versa > [!quote] ID:rw_szwbwfpx30m_1 > With chat, you have to bring the work to Claude. With co-work, it works where you already are. **Context & use**: - [00:00:53](https://www.youtube.com/watch?v=szWBwfPX30M?t=53), during the chat-vs-Cowork comparison. - Crisp encapsulation of the core file-handling difference between chat-based and folder-scoped agent tools; useful as a one-line explainer when introducing agentic tools to non-technical users. **Tags**: #xref **Highlight**: [view highlight](readwise://open?url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DszWBwfPX30M) <!-- atomicity: 1 idea ✓ --> #### [H2] CLAUDE.md is standing instructions read at every task start > [!quote] ID:rw_szwbwfpx30m_2 > A claude.md file is a set of standing instructions that co-work reads automatically every single time it starts a task in your folder. **Context & use**: - [00:05:38](https://www.youtube.com/watch?v=szWBwfPX30M?t=338), introducing the concept. - The definitional core of the whole video's thesis — directly transferable to any coding-agent CLAUDE.md convention (relevant to Courtney's own project CLAUDE.md practices already in use). **Tags**: #xref #addendum **Highlight**: [view highlight](readwise://open?url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DszWBwfPX30M) <!-- atomicity: 1 idea ✓ --> #### [H3] Instructions live in the folder, not the prompt > [!quote] ID:rw_szwbwfpx30m_3 > The instructions live in the folder, not in my prompt. **Context & use**: - [00:07:41](https://www.youtube.com/watch?v=szWBwfPX30M?t=461), closing the walkthrough of an example CLAUDE.md. - A memorable, quotable framing of "context as infrastructure" rather than "context as repeated conversation input" — a reusable aphorism for explaining agent configuration to others. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://open?url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DszWBwfPX30M) <!-- atomicity: 1 idea ✓ --> #### [H4] memory.md makes the assistant improve with use, transparently > [!quote] ID:rw_szwbwfpx30m_4 > Between these two files, your assistant actually gets better the more you use it. **Context & use**: - [00:10:49](https://www.youtube.com/watch?v=szWBwfPX30M?t=649), contrasting editable memory.md with opaque chat memory. - Names the mechanism (inspectable, editable memory file) behind a claim that's otherwise easy to dismiss as marketing — worth cross-referencing against how other agent tools implement persistent memory. **Tags**: #xref **Highlight**: [view highlight](readwise://open?url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DszWBwfPX30M) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Standing-instructions files replace repeated context-setting **Claim** (one sentence, Courtney's voice): A folder-scoped, always-loaded instructions file (CLAUDE.md-style) eliminates the need to re-explain identity, tone, and workflow to an AI agent on every task. **Evidence pointer**: [[#H2]] > [!quote] > It stops being a chatbot you talk to and becomes an agent you've tailored to do exactly what you want **Scope & conditions**: Holds for any agentic tool that supports folder-scoped or project-scoped always-loaded context files (this practice already underlies Courtney's own project CLAUDE.md conventions). Breaks down if the instructions file grows stale, self-contradictory, or isn't reviewed — the tool has no mechanism shown here for flagging drift. **Links** (≥1 with annotation): - [[CLAUDE.md conventions]] — because applies-in Courtney's own coding-project CLAUDE.md practice, already a standing pattern across her repos. <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Compare CLAUDE.md/memory.md pattern against other agentic tools' persistent-context mechanisms (e.g., project rules files in other coding agents) to see how generalizable the pattern is. — https://www.youtube.com/watch?v=szWBwfPX30M #xref - [F2] Consider whether Courtney's own project CLAUDE.md files would benefit from an explicit "skills" subsection (named, on-demand task procedures) the way the video's example folder does. — internal workflow question #articleseed ### → Things 3 (max 2 actions) - [ ] Review whether any active project CLAUDE.md files would benefit from adding named "skills" subsections for recurring tasks, per this video's pattern — courtney workflow #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Claude Cowork Finally Clicked When I Learned This.md|Claude Cowork Finally Clicked When I Learned This by YouTube]] ## Highlights - can think of it like this. With chat, you have to bring the work to Claude. With co-work, it works where you already are. Let's start to set this up and I'll show you an example of this in action before we get to the third and probably the most important thing about Claude co-work that finally made it click for me. We'll start by opening up the Claude desktop app and switching from chat over to co-work. Here in the prompt box is where you can start any task. And you'll ([View Highlight](https://read.readwise.io/read/01kxh9hpbq0e5yjajzq22918k4)) - Tags: atom/literature/h1 - supposed to act. This single file is the most important thing to understand about Claude Co-work because once you get this, you'll understand what makes Co-work so powerful. So, what exactly is this file? A claude.md file is a set of standing instructions that Co-work reads automatically every single time it starts a task in your folder. So, as long as you name your file claude.md, ([View Highlight](https://read.readwise.io/read/01kxh9ht7rjb2gym32k4zx3t9k)) - Tags: atom/literature/h2 - If this file had been set up, I wouldn't have needed to explain any of that in my prompt. I could literally just say, "Analyze the new exams." And Claude would know where to look, what kind of report to build, and how to format it, and where to save it. The instructions live in the folder, not in my prompt. This is the part that made Co-work finally click for me. Out of the box, every new Co-work task starts fresh. Without a system in place, you're re-explaining who you are, what you're working on, what your workflows are, and ([View Highlight](https://read.readwise.io/read/01kxh9hywfqf62j1532few25ps)) - Tags: atom/literature/h3 - to repeat yourself ever again. But more than that, you get to decide exactly how Claude works. The workflows, the memory, the way it shows up, and everything you want it to know about you all in one place. At that point, it stops being a chatbot you talk to and becomes an agent you've tailored to do exactly what you want the way you want it done. That is the big shift. If you want to try co-work for yourself, it's available on any paid Claude plan. The link is at the top of ([View Highlight](https://read.readwise.io/read/01kxh9j3wftnhg4q3ea2zw2hm9)) - Tags: atom/literature/h4