The Reason AI Gives You Useless Answers  🔗 **Source:** [The Reason AI Gives You Useless Answers](https://www.youtube.com/watch?v=o2S1ZbmVAAY#transcript) > [!note]+ Document Notes > tags: [personal-knowledge-management, context-engineering, ai-workflows, prompting, second-brain, delegation] > author: Tiago Forte > source: Tiago Forte (YouTube) > reference: https://www.youtube.com/watch?v=o2S1ZbmVAAY > date created: 2026-06-17 14:10 > > # 🏷️ Title: The Reason AI Gives You Useless Answers (Fix This First) > <!-- 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) ### 🔗 References - [Tiago Forte — Building a Second Brain](https://www.buildingasecondbrain.com) - [The PARA Method](https://fortelabs.com/blog/para/) - [Forte Labs — AI Second Brain program](https://fortelabs.com) - [Obsidian](https://obsidian.md) ## 📝 Summary - **What happened.** Tiago Forte argues that the reason AI returns generic, surface-level answers usually isn't a bad prompt — it's missing *context*. He names the skill "personal context management" (PCM), defines its three layers (persistent, project, perishable), and demonstrates them live in Claude Cowork by progressively feeding context into the same book-launch request and watching the output go from vanilla to executable. - **Why it matters.** It reframes the popular "prompt engineering" obsession as the wrong lever. The bottleneck moves from how you phrase a request to how well you've curated and staged the surrounding information — a shift that maps cleanly onto note-taking and knowledge-management practice rather than technical skill. - **Key stakeholders.** Knowledge workers using AI assistants; people already running PARA/second-brain systems; anyone storing notes in tools like Obsidian; and Forte Labs (whose "AI Second Brain" program this video funnels toward). - **Risks and tradeoffs.** As you manage context better, you do less of the actual work yourself — the constraint shifts from output to clarity of intention. Forte also pushes back hard on the industry term "context engineering," which is partly a definitional/branding argument, not a purely empirical one. - **Bottom line.** Don't fix your prompt first — fix your context. Stage three layers (persistent → project → perishable), and treat the practice as delegation to a system. The identity you grow into is "Context Architect." ## 📖 Main Content - **The misdiagnosis.** You gave the AI good information and wrote a good prompt, but the answer came back generic. Forte's claim: the failure is in the *context* you did or didn't give, not the prompt wording. This is the framing the whole video hangs on. - **PCM defined.** Forte has taught personal knowledge management (PKM) for a decade — "a system for preserving and then resurfacing your best knowledge." AI changes the audience for your notes: they're no longer just for you, they're inputs to a system that acts on your behalf. He renames the evolved practice **personal context management (PCM)** — cultivating the right context so AI can do meaningful work for you. - **The three layers (the "three Ps").** - *Persistent context* — stable, durable facts about who you are (role, values, communication preferences) that inform every interaction. Lives in your "master prompt," stored in custom instructions / personal preferences, so you don't repeat yourself each conversation. - *Project context* — what you're working on right now: product details, decisions, research, stakeholders, deadlines, objectives. In PARA, this is a single project folder. Forte argues PARA users have been "preparing for PCM without even knowing it." - *Perishable context* — the precise, minimal slice the AI needs for one specific request: the latest nuance and subtlety that lives only in your head. - **The live demo.** In Claude Cowork with memory, connectors, skills, and preferences all turned off (a "blank slate"), Forte asks for a launch/marketing plan for his book *Life in Perspective* (out November 3). With no context the output is "completely generic." He then layers in (1) his master prompt as custom instructions, (2) the project folder plus his Obsidian vault (where the book's raw material lives), and (3) a perishable brain-dump (his son was born May 11, take it easy a few weeks, cut the podcast idea, prioritize highest-leverage activities). The plan moves from one-size-fits-all to "so dialed in" it's executable. - **Pushback on "context engineering."** Forte concedes the industry calls this "context engineering" but objects to "engineering" — it implies a technical skill. He argues it's really arranging ordinary information (text, images, files, folders) into "loosely curated collections" and pointing AI at them: "this isn't engineering, it's just managing information." Caveat: accessible ≠ easy; you have to think at the level of the broader system. - **Three identity shifts.** - *Curating context* — spend your time finding the most detailed, specific, nuanced, contrarian, up-to-date context; act like a scientist or explorer seeking sources outside the AI's training data. - *How you spend your time* — the better you manage context, the less execution you do yourself. The bottleneck moves "up a level or two" from output to clarity of intention. PCM is **delegation to a system instead of a person**, and the constraint is identical: the quality of your handoff. - *Getting clearer about what you want* — AI "homogenizes" and "flattens" knowledge; the way to stand out is to externalize your judgment, taste, values, and opinions "so the AI can borrow it." The deepest reframe: PCM isn't curating data, it's curating *yourself*. - **The identity.** "PCM is the practice. Context Architect is who you become when you embrace it." It's not a new methodology but a new lens on what you already do — the notes, projects, and prompts. The video closes by pitching the AI Second Brain program. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=0) The misdiagnosis: it's the context, not the prompt - You did everything right and still got a generic answer; the problem is the context you're "not giving." - Promises three things: a name for the skill, the three layers, and the identity that comes with it. #### ⏱️ [00:00:34](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=34) From PKM to PCM - Ten years teaching personal knowledge management — "a system for preserving and then resurfacing your best knowledge." - AI changed *who's reading* your notes: they're now inputs to a system acting on your behalf. New name: personal context management (PCM). #### ⏱️ [00:01:08](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=68) Layer one — persistent context - Stable, long-term details: role, values, communication preferences. - Lives in the "master prompt" stored in custom instructions / personal preferences, loaded every interaction. #### ⏱️ [00:01:40](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=100) Layer two — project context - Active, scoped info: product details, decisions, research, stakeholders, deadlines, objectives. - In PARA, all of this lives in a single project folder — "you've already been preparing for PCM without even knowing it." #### ⏱️ [00:02:12](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=132) Layer three — perishable context - The precise minimal slice for one specific request; nuance/subtlety that only lives in your head. - "Persistent, project, perishable. Three Ps, three layers." #### ⏱️ [00:02:45](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=165) The blank-slate demo setup - Claude Cowork with memory off, connectors disconnected, no skills, empty preferences. - Task: launch/marketing plan for *Life in Perspective* (out November 3). #### ⏱️ [00:03:49](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=229) No-context baseline - AI asks "very basic vanilla questions" because it knows nothing about the book, publisher, or plans. - Result: "not bad" but "completely generic" — same plan for any author/book. #### ⏱️ [00:04:21](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=261) Adding layer one (master prompt) - Pastes the master prompt into the "instructions for Claude" (formerly personal/user preferences). - Includes name, age, location, family, interests, professional context, title, company, what they sell. #### ⏱️ [00:05:29](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=329) Persistent context changes the follow-ups - AI now asks about named team members (Julia, Jen); knows about the YouTube channel. - Surfaces a holistic insight: newborn timing could land content creation in a "newborn haze." "Like talking to a friend or a close colleague who has a holistic picture of your life." #### ⏱️ [00:06:34](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=394) Adding layer two (project folder + Obsidian vault) - Gives the whole *Life in Perspective* project folder, then his Obsidian vault — "most of the like specific raw material that went into the book actually lives in Obsidian." #### ⏱️ [00:07:38](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=458) Project context produces near-executable output - AI finds the specific launch day (November 3) and that YouTube is "53% of our audience growth"; reads the manuscript; pulls his origin story. - "Very, very close to something that I could really start executing on." #### ⏱️ [00:08:43](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=523) Adding layer three (perishable brain dump) - Son born May 11, a few weeks early; take it easy a few weeks then go full swing; kill the podcast idea; focus on highest-leverage / biggest-bang-for-buck activities. - AI tightens phases, swaps in a workshop bonus, cuts the podcast, "ruthlessly prioritizing." #### ⏱️ [00:10:55](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=655) Pushback on "context engineering" - Industry term is "context engineering"; Forte rejects "engineering" — it's "just managing information, managing content." - Caveat: accessible doesn't mean easy — you must think in terms of the broader system. #### ⏱️ [00:12:02](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=722) Shift one — curating context - Find the most detailed, specific, nuanced, contrarian, up-to-date context; mindset of a scientist or explorer seeking sources outside the AI's training data. #### ⏱️ [00:13:08](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=788) Shift two — you do less of the work - More context management → less execution by you; bottleneck moves from output to clarity of intention. - "PCM... is delegation, but to a system instead of to a person. And the constraint is exactly the same, the quality of your handoff." #### ⏱️ [00:13:41](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=821) Shift three — get clearer about what you want - AI "homogenizes" and "flattens" knowledge; you stand out by externalizing judgment, taste, values, opinions "so the AI can borrow it." - "What you're really doing in personal context management... isn't curating data, it's curating yourself." #### ⏱️ [00:14:48](https://www.youtube.com/watch?v=o2S1ZbmVAAY&t=888) The identity and the pitch - "PCM is the practice. Context Architect is who you become when you embrace it." Closes with the AI Second Brain program. ## 📚 References - [Building a Second Brain (book + site)](https://www.buildingasecondbrain.com) — Forte's PKM framework PCM extends. - [The PARA Method](https://fortelabs.com/blog/para/) — the project-folder structure Forte maps to layer-two context. - [Forte Labs / AI Second Brain](https://fortelabs.com) — the program the video promotes. - [Obsidian](https://obsidian.md) — where Forte stores the book's raw research material. - [Claude (Cowork)](https://claude.ai) — the assistant used in the live demo. ## 🗣️ Paraphrase - Generic AI answers usually trace to missing context, not a bad prompt. - The mature version of personal knowledge management is personal context management: curating context so AI can act for you. - Context has three layers — persistent (who you are, in your master prompt), project (the current work, a PARA folder), and perishable (the last-mile nuance only in your head). - A live demo shows the same book-launch request improving from generic to executable as each layer is added. - Forte objects to calling this "context engineering" because it's really information management, not a technical skill — though it's accessible, not easy. - Three identity shifts follow: curate context like a scientist, do less execution and more delegation, and externalize your judgment so the AI can borrow it. - Ultimately you're curating yourself; the identity you grow into is "Context Architect." ## 📰 Newsletter **Stop fixing your prompt. Fix your context first.** Tiago Forte's latest is a tidy reframe for anyone frustrated that AI keeps handing back beige, one-size-fits-all answers. His diagnosis: it's almost never your prompt — it's the context you didn't stage. He renames the discipline *personal context management* (the grown-up sequel to Building a Second Brain) and splits it into three layers worth stealing: **persistent** (a "master prompt" of who you are, parked in your custom instructions), **project** (a single PARA folder of the current work), and **perishable** (the last-mile nuance still rattling around your head). The proof is a live Claude demo where the same "plan my book launch" request goes from vanilla to genuinely executable as each layer drops in — including the AI reading his Obsidian vault to find that YouTube drives 53% of his audience growth. For folks already running a Readwise → Obsidian pipeline, this lands close to home: your notes were always for future-you, and now they're inputs an assistant can act on. The spicy take is Forte's refusal of the term "context engineering" — he insists it's "just managing information," not a technical skill (accessible, he grants, but not easy). The deepest line: you're not curating data, you're curating *yourself* — externalizing your taste and judgment so the AI can borrow it. The identity on offer: Context Architect. Worth a watch, with the usual caveat that the runway ends at his AI Second Brain program. ## ✅ Action Items #### DevRel - Developer advocacy: build a "three layers of context" demo (persistent/project/perishable) for an AI-assistant audience. - Education: workshop on writing a reusable "master prompt" for custom instructions. - Developer experience: ship a master-prompt template + a PARA-to-context-folder mapping guide. - Developer marketing: clip the "curating yourself, not data" line as evergreen content. - Community management: host office hours on staging context before prompting. - Internal/external/cross-project: standardize a team "master prompt" convention for shared assistants. #### OSPO - Policy & Governance: guidance on what personal/company data belongs in persistent vs. perishable context. - License Compliance: Not applicable to this source. - Security Management: review PII exposure when feeding vaults/folders to AI assistants. - Software Inventory: catalog approved local-first stores (Obsidian, file folders) for context staging. - Developer Support: consultation on connecting note vaults to assistants safely. - Community Engagement: Not applicable to this source. - Strategic Alignment: track "context readiness" as an internal AI-adoption metric. #### Community - Maintainers: Not applicable to this source. - Contributors: a "master prompt" onboarding artifact for newcomers using AI tools. - Sponsors: Not applicable to this source. - Ecosystem: share context-staging patterns across knowledge-worker communities. #### WordPress - Core: Not applicable to this source. - Core Performance: Not applicable to this source. - Core AI: map "persistent/project/perishable" context to any WP AI-assistant feature design. - **Core Program Team:** Not applicable to this source. - Mobile: Not applicable to this source. - Playground: Not applicable to this source. - Community: a "context for AI" session for content-focused WordCamps. - Training: learning path on PCM/master-prompt authoring for non-technical users. - Marketing: Not applicable to this source. - Support: Not applicable to this source. - Design: Not applicable to this source. - Documentation: examples-first guide to writing a master prompt. - Photos / TV / Openverse: Not applicable to this source. - Polyglots: glossary for "personal context management," "master prompt," "perishable context." - Accessibility: caption/transcribe the demo for non-visual learners (see Accessibility Assessment). - Meta: Not applicable to this source. - CLI: Not applicable to this source. - Hosting: Not applicable to this source. - Tide / Themes / Plugins: Not applicable to this source. #### The WPCC - Programs: Not applicable to this source. - Projects: documentation sprint on PCM for content teams. - Funding experiments: Not applicable to this source. ### FAIR - Technical independence: local-first vaults (Obsidian/folders) as a privacy-preserving context store — document it. - Policies and covenants: transparency about what personal data is fed into AI context layers. - Productized governance artifacts: a "context data-handling" checklist for AI-assisted work. ## 🧪 Hypotheses - H1: Adding structured persistent + project context to a request reduces clarifying-question rounds versus prompt tweaking alone. - H2: PARA users reach "executable" AI output faster than users without an organized project structure. - H3: A persistent "master prompt" improves cross-task output consistency more than per-task prompt engineering. - H4: Perishable context contributes a disproportionate share of final output quality relative to its small token size. - H5: Users who externalize explicit judgment ("taste") produce less homogenized AI output than users who don't. - H6: Time spent on execution falls as users mature their context-management practice (delegation effect). ### 📊 Methodology - H1/H2/H4: controlled prompt experiments — same task across context conditions (none, +persistent, +project, +perishable), measuring clarifying-question count, edit distance to a usable plan, and rater quality scores. - H3/H5: within-subject study comparing master-prompt vs. ad-hoc prompting across multiple unrelated tasks; blind raters score consistency and distinctiveness. - H6: longitudinal diary study tracking self-reported execution vs. context-setup time as PCM practice matures. ### ❓ Questions - Where exactly is the line between "context engineering" (technical) and "context management" (information work) — is the distinction real or rhetorical? - How much perishable context is optimal before diminishing or negative returns (context bloat)? - Does heavy persistent context risk over-anchoring the AI to stale preferences? - How transferable is the three-layer model across assistants with different memory/connector designs? ### 🔮 Future Research - Token-economy analysis of where quality gains actually come from across the three layers. - Best practices for versioning and pruning a "master prompt" over time. - Privacy/security patterns for exposing personal vaults to assistants. - Whether externalized "judgment" measurably reduces output homogenization at scale. ### 💡 Implications - **Practice:** Stage context before prompting; build a reusable master prompt and a per-project folder. - **Policy:** Organizations need rules for what data lives in persistent vs. perishable layers (privacy, retention). - **Security:** Feeding whole vaults/folders to assistants raises PII and access-scope questions. - **Education:** Curricula should teach context curation and judgment-externalization, not just prompt phrasing. - **Funding/tools:** Demand grows for local-first stores and assistants that ingest folders/vaults cleanly. ### 🧠 Learning Styles Assessment - Reading level estimate: spoken-monologue transcript; ~Flesch–Kincaid Grade 8–10 (conversational, common vocabulary; light jargon: "master prompt," "perishable context," "vibe"). Flesch Reading Ease ~60–70. #### 👀 Visual learners - Present: a live screen demo of Claude Cowork showing before/after outputs. - Improve: a three-layer diagram (persistent → project → perishable), a before/after split of the two plans, and a labeled screenshot of the master prompt. #### 🔊 Aural learners - Present: strong — clear narrated delivery; it's a spoken video. - Improve: a 2–3 minute audio recap and chapter markers. #### 🗣️ Verbal learners - Present: memorable coinages and analogies ("three Ps," "delegation to a system," "curating yourself," "newborn haze"). - Improve: an inline glossary (PCM, master prompt, perishable context). #### 🧩 Kinesthetic learners - Present: the demo is a reproducible step-by-step you can mirror. - Improve: a fill-in master-prompt worksheet and a "stage your three layers" checklist. #### 🧠 Logical learners - Present: clean cause→effect (more context → more specific output) shown incrementally. - Improve: a decision tree for what context belongs in which layer; a token-vs-quality chart. #### 🧑‍🤝‍🧑 Social learners - Present: framed as "talking to a friend or close colleague." - Improve: discussion prompts for teams building shared master prompts. #### 🧘‍♀️ Solitary learners - Present: reflective, self-directed framing ("curating yourself"). - Improve: journaling prompts ("what judgment of mine is still implicit?"). #### 🌿 Naturalistic learners - Present: minimal — light "perishable/persistent" lifecycle metaphor. - Improve: tie context lifecycle (perishable expires, persistent endures) to ecosystem/seasonal analogies. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: YouTube page provides standard heading/landmark structure. - Improve: ensure the description carries a text outline of the three layers for non-visual users. #### Keyboard navigation - Present: YouTube player is keyboard-operable with visible focus. - Improve: confirm custom controls/end cards are reachable and labeled. #### Contrast and readability - Present: standard YouTube UI generally meets contrast norms. - Improve: verify on-screen demo text (custom instructions, plan output) hits 4.5:1; it can be small. #### Text resizing and reflow - Present: responsive player and transcript panel. - Improve: confirm captions remain legible and unclipped at 200% zoom. #### Media and captions - Present: auto-captions exist (this transcript derives from them) — but they contain errors and music tags. - Improve: publish a human-corrected transcript and a downloadable version; avoid relying on auto-captions alone. #### Forms and interactions - Present: comment/subscribe controls are standard. - Improve: ensure any linked landing-page forms have labels and inline validation. #### Standards reference - Maps to WCAG 2.2 AA: 1.2.2 Captions (auto-captions present but error-prone), 1.4.3 Contrast (verify demo text), 1.4.10 Reflow, 2.1.1 Keyboard, 2.4.6 Headings/labels, 1.4.5 Images of Text (on-screen plan text). ### 🌍 DEIB Assessment #### Cultural breadth - Present: examples are personal/individual (book launch, family) — broadly relatable but US/knowledge-worker-centric. - Improve: show PCM applied in non-Western or team/enterprise contexts and lower-resource toolchains. #### Inclusive language - Present: plain, friendly, mostly jargon-light; coined terms are defined as introduced. - Improve: define "master prompt," "perishable context," and "vibe" inline for newcomers. #### Neurodiversity support - Present: clear sequential structure (three named layers, three named shifts) aids predictability. - Improve: add on-screen signposts/timestamps and chunked recaps for the longer demo stretch. #### Representation - Present: single presenter; one personal scenario. - Improve: include varied user personas (different jobs, tools, budgets) demonstrating the same method. #### Psychological safety - Present: encouraging, low-pressure framing; "you've already been preparing for PCM without even knowing it." - Improve: acknowledge anxiety about "doing less of the work yourself" and data-privacy concerns when sharing vaults. ### ✅ Logical Validity #### Strong reasoning - Causal demonstration (00:03:49–00:10:55): each added context layer visibly improves the same request — a clean, observable cause→effect held constant on one task. - Analogical (00:13:08): "PCM is delegation, but to a system instead of to a person" — apt and clarifying; the "quality of your handoff" maps well to managing people. - Definitional/conceptual: the persistent/project/perishable taxonomy is mutually distinct and exhaustive enough to be operational. - Inductive (00:13:41): the "AI homogenizes/flattens" claim supports the "externalize judgment" prescription, though it's asserted more than evidenced. #### Fallacies - **Anecdotal evidence / sample of one** — the entire proof is a single self-run demo. Location: 00:03:49–00:10:55. Weakness: no control for prompt order, cherry-picking, or the demonstrator's incentive to show improvement. Fix: show multiple tasks/users and a blind comparison. - **Definitional fallacy / "no true Scotsman" on "engineering"** — "this isn't engineering, it's just managing information." Quote (00:10:55): "This isn't engineering, it's just managing information, managing content." Weakness: redefines a term to win the point rather than engaging the technical reality of context windows/retrieval. Fix: argue the substantive claim (accessibility) without legislating the word. - **Appeal to consequences / motivated framing** — "the only way you can stand out is by being more clear" (00:13:41) plus the closing program pitch. Weakness: a confident universal claim that conveniently routes to a paid product. Fix: frame as a hypothesis and separate the claim from the upsell. - **Loaded analogy (mild)** — "curating yourself" elevates an information task to an identity transformation. Location: 00:13:41–00:14:48. Weakness: rhetorically inflates scope beyond what the demo shows. Fix: keep the identity framing aspirational and clearly distinct from the demonstrated mechanics. #### Tone of voice - Calm, teacherly, optimistic-but-grounded; lightly self-deprecating (the newborn aside); persuasive with a clear funnel toward the program. Confident coinage ("Context Architect") balanced by the "accessible, not easy" caveat. ### 🛡️ DARVO Analysis - Deny: Not applicable — no accusation is being deflected; the "it's not your prompt" framing reassigns cause but doesn't deny wrongdoing. - Attack: Not applicable — no attack on critics; the pushback on "context engineering" targets a *term*, not people. - Reverse victim/offender: Not applicable. - Overall: DARVO is not applicable to this source. It's an instructional/marketing monologue with no accuser, defendant, or grievance dynamic. ## 🎯 Conclusion - **Bottom line.** Generic AI output is usually a context problem, not a prompt problem. Stage three layers — persistent (master prompt in custom instructions), project (a PARA folder), perishable (last-mile nuance) — and the same request goes from vanilla to executable. The deeper move is treating this as delegation to a system and externalizing your own judgment so the AI can borrow it. - **Next steps for readers.** (1) Write a reusable "master prompt" and store it in your assistant's custom instructions. (2) Point the assistant at one well-organized project folder (PARA) and your note vault. (3) Before a request, do a 60-second perishable brain-dump. (4) Spend a session making your implicit taste and judgment explicit. (5) Treat weak output as a context-staging gap, not a prompting failure. ## 🔄 Refactor Appendix — The Reason AI Gives You Useless Answers (Fix This First) > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults applied. Resonance filter on. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Atoms (H3) #### [H1] Generic AI answers come from missing context, not bad prompts > [!quote] ID:rw_pending > "The problem isn't in your prompt. It's in the context you're giving, or more specifically, not giving it." **Context & use**: - Opening thesis, [00:00:00]. - Reframes the prompt-engineering obsession; the lever is context staging, not wording. **Tags**: #xref #aphorisms #share **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H2] Context has three layers: persistent, project, perishable > [!quote] ID:rw_pending > "Persistent, project, perishable. Three Ps, three layers. Together, they're what we mean when we say context." **Context & use**: - [00:02:12]; persistent = master prompt, project = PARA folder, perishable = last-mile nuance. - Operational taxonomy that maps directly onto an existing PARA/Obsidian setup. **Tags**: #xref #addendum **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H3] Personal context management is delegation to a system > [!quote] ID:rw_pending > "PCM, personal context management, is delegation, but to a system instead of to a person. And the constraint is exactly the same, the quality of your handoff." **Context & use**: - [00:13:08]; bottleneck moves from output to clarity of intention. - Lets you import management/delegation skill into AI work wholesale. **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> #### [H4] You curate yourself, externalizing judgment so AI can borrow it > [!quote] ID:rw_pending > "You have to externalize your judgment so the AI can borrow it. What you're really doing in personal context management, I believe, isn't curating data, it's curating yourself." **Context & use**: - [00:13:41]; counter to AI's "homogenizing/flattening" effect. - The distinctiveness lever: taste, values, opinions made explicit. **Tags**: #xref #articleseed **Highlight**: [view highlight](readwise://pending) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] AI output quality is bounded by staged context, not prompt phrasing **Claim** (one sentence, Courtney's voice): The quality of an AI's answer is determined primarily by how well you've staged the surrounding context across stable, project, and last-mile layers, so context curation — not prompt wording — is the real lever. **Evidence pointer**: [[#H1]], [[#H2]] **Scope & conditions**: Holds for open-ended knowledge work where the model needs facts it can't infer. Breaks for self-contained tasks where the prompt fully specifies the request, or where added context exceeds useful window size and degrades focus. **Links**: - [[Building a Second Brain — PARA]] — because extends note-taking into AI inputs - [[Prompt engineering]] — because contrasts on which lever drives output quality - [[Readwise to Obsidian PARA workflow]] — because applies-in staging project/persistent context <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Draft a reusable "master prompt" and paste it into custom instructions — https://www.youtube.com/watch?v=o2S1ZbmVAAY #fleeting #inbox - [F2] Open question: how much perishable context before context-bloat hurts output? #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Write a personal master prompt (persistent layer) and store it in Claude custom instructions — #things - [ ] Map one active PARA project folder + Obsidian notes as a reusable project-context bundle for AI — #things <!-- deep-read:sections end --> 📄 **Full text:** ![[Full Document Contents/Articles/The Reason AI Gives You Useless Answers.md|The Reason AI Gives You Useless Answers by Tiago Forte]] ## Highlights - You've done everything [music] right. You gave the AI all the information it needed. You wrote what you thought was a good prompt, and the answer [music] came back generic, surface-level. The problem isn't in your prompt. It's in the context you're giving, or more specifically, [music] not giving it. In this video, I'm going to give you a name for that [music] skill, the three layers it operates on, and the identity that comes with practicing [music] it well. For 10 years now, I've been teaching ([View Highlight](https://read.readwise.io/read/01kvbc4m5snbw01mbx0f9avdf1)) - Note: Thesis: context, not the prompt, is the lever. - Tags: atom/literature/h1, context-engineering - even knowing it. [music] Layer three is perishable context. This is the precise minimal slice that the AI needs [music] to do its job. It's just for this specific request, what little bits of context and details [music] and nuance and subtlety is going to help it get to a useful result. [music] Persistent, project, perishable. Three Ps, three layers. Together, they're what we mean when we [music] say context. In the AI second brain, we actually ([View Highlight](https://read.readwise.io/read/01kvbc4qbfbgw5wvhmkyrw99ny)) - Note: The 3-layer context taxonomy (persistent/project/perishable) — maps to PARA. - Tags: atom/literature/h2, atom/zettel/z1 - management, is delegation, but to a system [music] instead of to a person. And the constraint is exactly the same, the quality of your handoff. That's what determines the quality of the result. The third mindset shift is you have to get clearer about what you really want. This one is kind of the deepest, it's [music] the most profound. See, AI has this effect of homogenizing everything it comes into contact [music] with. It flattens human knowledge. The only way you can stand out [music] is by being more clear. [snorts] So, you really have ([View Highlight](https://read.readwise.io/read/01kvbc4v2hxajaernamvx9m077)) - Note: PCM = delegation to a system; same constraint = quality of your handoff. - Tags: atom/literature/h3 - to think about what are the unique inputs that you can provide, your values, your taste, your perspective, your opinions. >> [music] >> You have to get that implicit judgment and make it explicit. You have to externalize your judgment [music] so the AI can borrow it. What you're really doing in personal context management, I believe, isn't curating data, it's curating yourself. PCM is the practice. >> [music] >> Context Architect is who you become when you embrace it. PCM isn't a new methodology to learn. It's a new way of ([View Highlight](https://read.readwise.io/read/01kvbc4ypmqtp4rvj6e4rb8hkn)) - Note: Curating yourself, not data — the anti-homogenization lever. - Tags: second-brain, atom/literature/h3 ## Links - [[Context Not Constraints Is What Unhobbles an Agent]] — existing Permanent captures the core diagnosis that missing situation, not weak phrasing, limits useful output. - [[Intent Written on the Outside Compounds While Re-Derived Intent Is Debt]] — applies-in: persistent, project, and perishable context externalize judgment so it can be reused.