The Master Prompt 2.0: Live Session with Tiago Forte  🔗 **Source:** [The Master Prompt 2.0: Live Session with Tiago Forte](https://www.youtube.com/watch?v=VfDZnvUCM_Q) > [!note]+ Document Notes > > [!info] Provenance > > **Analysis generated:** 2026-08-27 by Claude (Opus 5) via the `readwise-deep-read` pipeline. > > **Source:** YouTube — https://www.youtube.com/watch?v=VfDZnvUCM_Q > > **Title:** "The Master Prompt 2.0: Live Session with Tiago Forte" · **Transcript:** ~75,600 chars / ~13,950 words · video (90-minute live session, ~500 attendees) > > **Speaker:** **Tiago Forte**, Forte Labs — author of *Building a Second Brain* and the PARA method. > > **Commercial frame, disclosed at the top:** `[STATED]` This is *"session zero"* for a paid cohort running **17 September – 8 October**. His framing: *"instead of giving you the typical sort of launch webinar full of a lot of hype… why don't we just teach the very first session of the program?"* `[EVIDENCE]` Roughly the final 35 minutes are the cohort pitch. **Noted for calibration — the teaching half is substantive and genuinely given away.** > > > > ⚠️ **TRANSCRIPT RECOVERY — how this document exists.** > > `[EVIDENCE]` Reader's scrape of the original save returned the empty sentinel: *"Unfortunately, Youtube does not provide subtitles for this video."* `[EVIDENCE]` `yt-dlp --write-auto-subs` confirmed it — *"There are no subtitles for the requested languages."* `[EVIDENCE]` The transcript was recovered via the **vidIQ** API (step 2 of the skill's YouTube subtitle fallback), then **re-saved into this document with the transcript embedded** (step 4), because the Reader API does not permit updating `html_content` on an existing doc. > > **This document therefore has anchorable text and inline highlights; the original empty save (`01m11k7x15vr46z2b81zk279p7`) has been archived.** Transcript also at `scratchpad/tx_tiago.txt`. > > > > **Register key:** `[STATED]` speaker-stated · `[EVIDENCE]` demonstrated · `[INFERENCE]` reviewer inference · `[RECOMMEND]` reviewer recommendation. > > **Coverage:** Full recovered transcript read. Screen-shared slides, the Claude demo, and the builder document are described as narrated and were not captured. > > # 🏷️ Title: The Master Prompt 2.0: Live Session with Tiago Forte > <!-- 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 - **"Context Maxing"** — `[STATED]` a Brookings meta-study said to coin the term and define *cognitive agency*. `[INFERENCE]` **Not verified.** Named without authors or date; treat as unconfirmed until located. - **Gall's Law** — quoted accurately and attributed correctly - **PARA** (Projects, Areas, Resources, Archives) — Forte's own framework - Tools named: Claude (Co-work / Code), ChatGPT (Codex / agent mode), Notebook LM, Notion, Obsidian, Google Drive - `[STATED]` A master-prompt **builder document** distributed as Markdown, *"version two updated on August 20th… this is really more like version 11"* ## 📝 Summary **The central claim.** `[STATED]` *"There's a new bottleneck in modern productivity… It's no longer your time, your attention, your intelligence… It's not even really the AI's intelligence… The bottleneck has moved. What determines your results is collecting context, curating your context, and providing that context to AI."* **The framework — a five-layer context architecture**, and this is the session's real contribution: | Layer | What it is | Your control | |---|---|---| | 1. **Training data** | Everything the model was trained on | **None** | | 2. **System prompt** | Written by the vendor; sets personality, defaults, prohibitions | **None** | | 3. **Master prompt** | Your personal/professional profile, preferences, rules | **Full — and you must write it** | | 4. **Project knowledge** | Deep context for one domain or project | **Full** | | 5. **The chat window** | Prompts, replies, attachments | Full, **low leverage** | `[STATED]` *"Most of the leverage lies in layers 3 and 4."* And the deliberately provocative corollary: `[STATED]` *"at least 90% of the AI content on social media is about prompting, prompt engineering. It's the most visible layer… but it's also kind of where the least leverage lies. By the time you're typing something into a chat box, it's kind of too late."* **The best single argument in the session**, and it is concrete: `[STATED]` ask a model to summarise *War and Peace* and it will, many ways. Ask *"what is the last sentence in chapter three"* and it cannot — *"it doesn't actually have direct access to its training data… Just because it's in the training data doesn't mean it can directly call on those facts."* `[INFERENCE]` **That is the sharpest argument for note-taking I have seen made in an AI context**, and it is a factual claim about model architecture rather than an appeal to habit. **Bottom line.** `[INFERENCE]` A genuinely useful framework carrying **one branding term that contradicts the practice it names**, one unverified citation doing load-bearing work, and — raised by an audience member rather than the instructor — **the real cost of the whole approach.** ## 📖 Main Content ### "Context maxing" is the wrong name for what he actually recommends `[STATED]` The term is introduced as the alternative to *token maxing*: *"context maxing… is really centering your context. It's putting most of your time and your attention and effort, not in fiddling with the technology, but in curating and cultivating your own context."* `[EVIDENCE]` **Then, answering a question, he recommends the opposite of maximising.** On whether a master prompt can be too long: `[STATED]` *"Yes, you can absolutely have too much… The more details it has to look through to decide which are relevant, the higher the odds it chooses the wrong ones… when you add more and more details, at some point it becomes noise, and noise actually worsens the quality of your results."* `[EVIDENCE]` And on assembling context per task: `[STATED]` *"I'm never giving it this whole stack of folders. That would be crazy. I'm drilling into actually what is the smallest folder… **What is the smallest amount of context that I can give it to solve the task at hand rather than the biggest amount of context that I can give it?**"* `[INFERENCE]` **The practice is right and the name is wrong.** What he describes is **context curation** — a small, stable profile at layer 3 plus modular, task-scoped bundles at layer 4, assembled on demand. *Maxing* names the failure mode he spends the Q&A warning against, and it will be the part people remember. `[INFERENCE]` **The practice also matches a rule already in Courtney's `CLAUDE.md` almost word for word** — *load the smallest useful context: grep/index/paths first, targeted reads second, whole files last.* **Two independent arrivals at the same principle**, which is the third such convergence in this batch after the deterministic-script rule and the verify-delegated-work rule. ### The training-data argument is the best thing in the session `[STATED]` *"If you ask ChatGPT or Claude to summarize the book War and Peace, they can do it… But then ask the AI, what is the last word on page five, or what is the last sentence in chapter three, and it can't tell you. It can't tell you that about any book, even the most well-known books, because it doesn't actually have direct access to its training data."* `[INFERENCE]` **This converts "you should take notes" from an exhortation into a consequence.** The usual argument for note-taking is about memory or discipline. This one is about **retrieval architecture**: a model holds compressed statistical structure, not an indexed corpus, so specifics have to be supplied. `[STATED]` *"If you wanted to write something really novel and interesting and unique about the book War and Peace, you would need to read War and Peace and supply notes… Thus the need for note-taking."* `[RECOMMEND]` **And it is empirically checkable in thirty seconds**, which is the best property an argument like this can have — ask any model for a specific sentence from a famous book and watch what happens. ### The system-prompt observation, and what the master prompt is really for `[STATED]` *"pretty much all the major LLMs have this instruction in their system prompt, this default bias, which is to give quick, easy-to-follow answers… just do the best with the information that you have."* `[STATED]` And the objection: *"a responsible collaborator does not just do the best with the information it has. It asks follow-up questions. It digs deeper. It pushes back. It admits what it knows and what it doesn't know."* `[INFERENCE]` **So part of the master prompt's function is counteracting a vendor default** — which reframes it usefully. It is not only a description of you; **it is a correction applied on top of a product decision you did not make and cannot see.** `[INFERENCE]` **This is a rule Courtney's `CLAUDE.md` already implements**, from two directions: *ask when the requirements are unclear* counteracts the do-your-best default, and *tell that you're guessing — if the reason for dismissing something is "probably fine," that's an inference about runtime from source* counteracts the confident-answer default. `[EVIDENCE]` Forte reports doing the same thing: *"I always wanted to surface what it doesn't know or what would change its answers."* ### Anita's question is the sharpest thing said, and it comes from the audience `[STATED]` *"I sometimes find that the more information I add to the master prompt and the context layers, the more Claude sounds like me — which is sometimes great, but sometimes I do want a genuinely different point of view."* `[INFERENCE]` **This identifies the structural cost of the entire approach, and the instructor did not raise it.** A master prompt is, by construction, an instrument for making a model reason more like you — Forte agrees: `[STATED]` *"You're almost like giving it your personality, giving it your perspective, giving it your biases, giving it your opinions."* `[INFERENCE]` **Biases included, which is the problem.** The better the master prompt works, the more reliably the model reproduces your framing, and **the harder it becomes to notice what your framing excludes.** That is an echo chamber built deliberately and installed at the deepest layer you control. `[EVIDENCE]` His answer is a set of **AI advisors** with contrasting personalities — a financial advisor tuned risk-averse, a productivity coach tuned prescriptive, adjustable along Big Five traits. `[INFERENCE]` **That is a workaround rather than a resolution, and it is worth being precise about why.** The advisors are also authored by you, so they encode *your model of what disagreement looks like*. **A disagreement you designed is not the same as a disagreement you did not anticipate.** `[RECOMMEND]` The honest version: a master prompt trades *breadth of perspective* for *fit*, that trade is usually correct, and **the mitigation is to keep an unpersonalised context available for the cases where you want to be surprised** — which costs nothing, since a second account or a fresh session provides it. `[INFERENCE]` It also lands directly on this batch's sycophancy thread — the GitHub alt-text and Dreams of Code pairing, where a model reviewing others always found fault and a model reporting its own work always said done. **Personalisation is a third direction of the same failure: a model that agrees with you because you told it how you think.** ### The demo is a real controlled comparison, and he isolated the variable `[EVIDENCE]` **Credit where it is due.** Two paid Claude accounts, same question — *"what are the best ways I could grow my business right now"* — and on the comparison account: `[STATED]` *"this account has nothing turned on. Doesn't have any connectors, doesn't have memory turned on, doesn't have a master prompt."* On the other: `[STATED]` *"it only has access to the master prompt, so I turned off connectors, integrations, I turned off memory."* `[INFERENCE]` **That is a single-variable comparison, deliberately constructed**, and it is more methodological care than this kind of demo usually receives. `[STATED]` He also flags it as unedited: *"this is a real example, this is not massaged, it's not cleaned up."* `[EVIDENCE]` The contrast is specific rather than vibes-based. Without context: *"squeeze more from existing customers,"* *"push harder into B2B enterprise,"* *"expand the certification and facilitator model,"* *"rework the pricing ladder,"* *"build a recurring revenue layer"* — each of which he can name as wrong: a strategy they rejected last year, a thing they already do, an offering it does not know exists. `[INFERENCE]` **The limits are equally clear:** n=1, the question and the evaluation are both his, and the assessment is unblinded. `[EVIDENCE]` The claim it supports is inflated well past what one comparison bears: `[STATED]` *"It's like a hundred times more valuable."* `[RECOMMEND]` **The design is good enough to be worth copying, and the honest version of the homework he assigns is the study.** `[STATED]` *"run it on a real task, a task that you tried before, one that's failed."* **Done blind, across several people, that is a real experiment** — and nothing about it requires his cohort. ### Two things that undercut his own maximalism `[EVIDENCE]` **Gall's Law**, quoted accurately and treated as the organising principle: `[STATED]` *"A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works."* He would put it on a billboard. `[EVIDENCE]` **And then**, minutes later: `[STATED]` *"some of you are maximalists… you want to make the ultimate, totally thorough master prompt. And if that's you, I have a maximalist version of this master prompt builder that has like 100 questions. If you really want to go ham and spend probably six to eight hours answering all those."* `[INFERENCE]` **A hundred-question intake producing a document designed from scratch is precisely what Gall's Law says will not work**, and it is offered as a premium option four minutes after the billboard line. The MVP path — 1,000–2,000 words, then a monthly amendment cycle — is the one consistent with his own stated principle. `[EVIDENCE]` **The monthly update ritual is the good half and carries an unusually honest number:** `[STATED]` ask the model to review connectors, integrations and local files and *"propose a set of amendments… I only accept maybe a quarter of the ones that it recommends."* `[INFERENCE]` **A 25% acceptance rate, volunteered, is the most calibrated statistic in the session** — it tells you the tool is useful and unreliable, which is the correct characterisation of automated suggestion, and it is exactly the "human keeps judgment, machine does volume" split this batch has now found in five separate domains. ### The connectors warning has no mechanism — and this batch has the mechanism `[STATED]` *"The second you add a connector, you are taking on a much higher level of risk… They're very powerful, very helpful. They're also like a security and privacy disaster."* `[STATED]` And his framing, which is the right one: *"What is the new capability? And then how do I counterbalance the risks that it introduces with some kind of boundary?"* `[INFERENCE]` **He names the question correctly and offers no implementation** — the boundaries suggested are *"rules and guidelines"* in the master prompt, which is to say **instructions to a model, not enforcement.** An instruction can be ignored, overridden by injected content, or simply lost in a long context. `[INFERENCE]` **The Kent C. Dodds session elsewhere in this batch has the missing half.** There, credentials are never visible to the agent, and the sandbox **intercepts every outbound fetch** to check the secret against an approved-domain list — *"not only can the agent not see it, but it also cannot send it to anywhere but approved domains."* **That is enforcement rather than instruction**, and it is the difference between asking a model to behave and making misbehaviour impossible. `[RECOMMEND]` **Read the two together.** Forte states the risk framing clearly for a general audience; Dodds supplies the architecture. Neither source knows about the other, and the combination is more useful than either. ### The model-tier claim matches the batch's data `[STATED]` *"This is a recurring theme: when you have good context, you don't need the best models. You don't need the bleeding edge frontier when you have good context, which helps you save time, helps you save money and tokens."* He recommends Sonnet for running the builder. `[INFERENCE]` **This is the DX finding from this batch, arrived at independently.** Across 460 companies and 190,000 engineers, staff engineers used *"significantly less tokens"* than juniors on the same use case — **judgment showing up as efficiency.** Forte's version relocates the same effect from the person to the context: **better context lowers the model tier you need**, because the model has less to infer. `[INFERENCE]` It also matches the standing model-and-effort policy of *diagnose failures in order: context problem → effort problem → model problem; never escalate model or effort to compensate for vague prompts or missing context.* **Three sources, one principle.** ## 🗒️ Notes - **The harness recommendation is unambiguous and correct:** `[STATED]` *"don't use the standard chat. That's the most important thing. Use Claude Co-work or Claude Code instead of the standard Claude chat, and use OpenAI Codex or… agent mode."* `[STATED]` Because harnesses bring *"file system access… plugins… scheduled tasks… sub-agents… computer use."* - **On Gemini:** `[STATED]` *"was in the past but has had quite a surprising fall from grace over the past year."* `[INFERENCE]` Asserted without evidence, and stated as settled. - **"Hardly anyone is doing this — if you have any version of a master prompt, I swear you're in the top one percent."** `[INFERENCE]` Unsourced adoption claim, hedged by *"I swear"*, which is not a source. Plausible and unmeasured. - **The 1,000–2,000 word target** and the stability rule are good, checkable constraints: `[STATED]` *"stick to things that stay true for months, not weeks… If something is specific to one project that is about to end, that shouldn't go in your master prompt."* - **Twelve sections named**, including several most people would not think to include — income sources and financial position, core values, **intentions for AI usage** (*"You might have an intention to not use AI in writing"*), a personal SWOT, and *"what does quality mean to you."* - **The AI-usage-intentions section is the most interesting and gets one line.** `[INFERENCE]` A standing instruction about *where you do not want the tool involved* is a boundary the model can enforce on your behalf rather than one you re-state every session — and it is the section least likely to occur to someone writing their first master prompt. - **The builder is a self-contained prompt**, distributed as Markdown, run in a fresh session, producing an interview of roughly 60–90 minutes. `[INFERENCE]` **Using a model to elicit the context it will later consume is a neat inversion** and the session's most reusable artifact design. - **Version control for the master prompt is raised and deferred** to the paid cohort. `[INFERENCE]` A plain-text document under git solves it, and that is not said. - **PARA is presented as total coverage:** `[STATED]` *"There is not one thing in my life, personal up here and work here, that is not reflected in a folder."* `[INFERENCE]` A strong claim about a personal system, offered as a model without evidence that the completeness is achievable by others. - **The cohort framing is unusually honest for a sales event:** `[STATED]` *"this cohort is not in any way like 'I have the knowledge, I am the instructor'… I almost see the cohort as like this big, crowdfunded, real-time R&D conference."* `[INFERENCE]` Whether it works that way is unknowable from here; **framing a paid course as collective research rather than transmission is at least the right ambition.** - **The Q&A is run by upvote**, with the audience voting questions up before he answers. `[INFERENCE]` It is why Anita's question — the best one asked — reached him at all. ## 📚 References `[EVIDENCE]` **One citation does load-bearing work and I could not verify it.** The **Brookings "Context Maxing"** paper is described as a meta-study coining the term and defining *cognitive agency* as *"the capacity for people to think and act with generative AI in ways that support control, efficacy, and mastery."* `[STATED]` *"This paper, you guys should really go check it out."* **No authors, no date, no link captured.** `[INFERENCE]` **This matters more than a normal missing citation**, because the paper is doing the session's heaviest lifting: it supplies the two-paths framing (passive use degrades cognitive agency; context-maxing enhances it), and that framing is the argument for the entire product. `[RECOMMEND]` **Locate it before repeating the claim.** If it exists and says what he says, it is the most important source in this batch for the delegation-atrophy question. If it does not, the session's strategic argument rests on an anecdote. `[EVIDENCE]` **Gall's Law is quoted accurately and attributed properly** — the one clean citation in the session. PARA is his own and identified as such. ## 🗣️ Paraphrase Forte argues the bottleneck has moved. It used to be your time and attention; more recently it was the model's capability. Now the models are good enough for almost everything, so what determines your results is the context you supply. He describes context as a five-layer structure. The training data and the vendor's system prompt you cannot touch. Your personal preferences — what he calls the master prompt — you can, and must write yourself. Project knowledge sits below that: narrow and deep, per domain. And the chat window is the top layer, which is where almost all public discussion of AI lives and where the least leverage is. By the time you're typing, four layers have already shaped the answer. The argument for note-taking is architectural. Ask a model to summarise *War and Peace* and it will. Ask for the last sentence of chapter three and it can't — it holds compressed structure, not the text. So specifics have to be supplied. He also notes that the major system prompts instruct models to give quick answers and do their best with what they have, which is reasonable product design and a real flaw. Part of the master prompt's job is countering that: ask follow-up questions, push back, say what you don't know. His demo is two Claude accounts, same question, connectors and memory off in both, master prompt on in one. Without it, generic advice — squeeze existing customers, push into enterprise — several of which he'd already tried and rejected. With it, specifics about his book launch, his email list, his team's bandwidth, and a list of what the model still didn't know. Asked whether a master prompt can be too long, he says yes — more detail means more chance of selecting the wrong detail, and past a point it's noise. He keeps his small and puts depth in per-project folders, drilling to the smallest relevant one. Asked by an attendee whether all this makes the model sound too much like her, he agrees, and proposes a set of AI advisors with deliberately contrasting personalities. He cites Gall's Law — complex systems designed from scratch never work — then offers a hundred-question maximalist builder for anyone who wants one. ## 📰 Newsletter **The sharpest argument for note-taking I've seen made in an AI context, and it's a fact about architecture rather than a pep talk.** Ask a model to summarise *War and Peace* — fine, a hundred different ways. Ask for the last sentence of chapter three and it can't, *"about any book, even the most well-known ones, because it doesn't actually have direct access to its training data."* It holds compressed structure, not an indexed corpus. **So specifics have to be supplied, and you're the only one who can supply yours.** Checkable in thirty seconds, which is the best property this kind of claim can have. **The framework is a five-layer context pyramid** — training data, vendor system prompt, your master prompt, project knowledge, the chat window — with the deliberately provocative conclusion that **prompting is the lowest-leverage layer**: *"at least 90% of the AI content on social media is about prompting… By the time you're typing something into a chat box, it's kind of too late."* **But "context maxing" is the wrong name for what he actually does.** Asked whether a master prompt can be too long, he says yes: *"the more details it has to look through to decide which are relevant, the higher the odds it chooses the wrong ones… at some point it becomes noise."* And on assembling per task: **"What is the smallest amount of context I can give it to solve the task at hand rather than the biggest?"** That's **curation**, not maxing — and *maxing* names the failure mode he spends the Q&A warning against. **The best moment comes from the audience, not the instructor.** An attendee named Anita: *"the more information I add… the more Claude sounds like me, which is sometimes great, but sometimes I do want a genuinely different point of view."* **That's the structural cost of the whole approach** — a master prompt is an instrument for making a model reason like you, biases included, and the better it works the harder it gets to notice what your framing excludes. His answer is AI advisors with contrasting personalities, which is a workaround: **a disagreement you designed isn't the same as one you didn't anticipate.** The free mitigation is keeping an unpersonalised session around for when you want to be surprised. **Credit on the demo:** two accounts, same question, connectors and memory off in both, master prompt the only difference. **That's a genuinely isolated variable**, which is more care than these demos usually get — though n=1, and *"a hundred times more valuable"* is not what one comparison shows. **Two things to check.** He cites a **Brookings "Context Maxing" paper** for the claim that passive AI use degrades cognitive agency while curated use enhances it — no authors, no date, and it's carrying the session's entire strategic argument. **If it exists and says that, it's the most important source I've read this week on the judgment-erosion question.** And he quotes **Gall's Law** — *"a complex system designed from scratch never works"* — then offers a hundred-question maximalist builder four minutes later. ## ✅ Action Items #### DevRel - **Developer advocacy:** `[RECOMMEND]` The **War and Peace test** is the best teaching device here — it converts "take notes" from advice into a consequence of retrieval architecture, and an audience can verify it live. - **Education:** `[RECOMMEND]` Teach the **five-layer model**, and teach that **layer 5 is the low-leverage one.** `[INFERENCE]` Most AI education is entirely layer 5, which makes the framework's main claim also a critique of the field it sits in. - **Developer experience:** `[INFERENCE]` The **builder-as-self-contained-prompt** pattern is worth stealing generally: a document that runs an interview and emits a configured artifact, with no setup. - **Developer marketing / Community management / Cross-project:** Not applicable to this source. #### OSPO - **Security Management:** `[RECOMMEND]` His connector warning is right and unimplemented — *"rules and guidelines"* in a prompt is **instruction, not enforcement.** Pair it with the egress-control architecture from the Dodds session in this batch, which is the enforcement half. - **Policy & Governance:** `[INFERENCE]` The **AI-usage-intentions** idea generalises to an organisational artifact: a written statement of *where the tool is not to be used*, enforceable at the assistant level rather than restated per task. - **All other OSPO items:** Not applicable to this source. #### Community - **Maintainers / Contributors / Sponsors / Ecosystem:** Not applicable to this source. #### WordPress - **All teams:** Not applicable to this source. `[INFERENCE]` A personal-productivity session with no WordPress surface; enumerating teams would manufacture relevance. #### The WPOCC - **Programs / Projects / Funding experiments:** Not applicable to this source. ### FAIR - Not applicable to this source in any respect. ### Personal practice — where this actually lands `[RECOMMEND]` Four items, ordered by how much they change anything: - **The `CLAUDE.md` file *is* a master prompt, and it already implements the corrections he recommends.** `[INFERENCE]` *Ask when the requirements are unclear* counteracts the vendor's do-your-best default; *tell that you're guessing* counteracts the confident-answer default; *load the smallest useful context* is his layer-4 curation rule. **The gap is not the corrections — it is the profile.** His twelve sections include several that are absent here: working style, communication preferences, tool preferences, personal SWOT, **what quality means to you**, and explicit **AI-usage intentions**. - **The stability rule is the useful filter for what to add:** `[STATED]` *"stick to things that stay true for months, not weeks."* `[INFERENCE]` Project-specific facts belong in project context, not standing instructions — which is a clean test for whether something belongs in `CLAUDE.md` at all. - **The monthly amendment ritual is worth adopting and the acceptance rate is the honest part.** `[STATED]` Ask the model to propose amendments from available context; *"I only accept maybe a quarter."* `[INFERENCE]` A quarterly prune is already scheduled in `CLAUDE.md`; **a monthly propose-and-review is the lighter-weight complement**, and 25% is a realistic expectation to hold going in. - **Keep an unpersonalised session available.** `[INFERENCE]` **This is the mitigation for Anita's problem and it costs nothing.** The better a master prompt works, the more consistently a model reflects your framing back — which is the point, and is exactly wrong when you want to be told something you have not already thought. ## 🧪 Hypotheses 1. **H1 — A stable personal profile beats prompt refinement.** Holding the model constant, adding a well-formed master prompt improves task-outcome ratings more than iterative refinement of the prompt itself. 2. **H2 — Context quality substitutes for model tier.** With good context supplied, a mid-tier model matches a frontier model on the same tasks at materially lower cost. 3. **H3 — Master prompts exhibit an inverted-U against length.** Output quality rises then falls as master-prompt length increases, with the peak well below available context capacity. 4. **H4 — Personalisation narrows perspective.** Responses generated with a master prompt show measurably less variance in framing than unpersonalised responses to the same question. 5. **H5 — Task-scoped context beats comprehensive context.** Supplying the smallest relevant context bundle produces better results than supplying a larger superset. 6. **H6 — Curated use protects cognitive agency; passive use degrades it.** The Brookings claim, if the paper exists. ## 📊 Methodology - **H1 — his own homework, done properly.** `[STATED]` *"run it on a real task, a task that you tried before, one that's failed."* **Make it a real study:** several participants, matched tasks, outputs rated blind by judges who do not know which condition produced which. **Confound:** writing a master prompt requires reflecting on your own goals, which may improve your task framing independently of what the model receives — **that is a genuinely interesting alternative explanation and needs a condition that separates them.** - **H2 — directly runnable and immediately useful.** Same tasks, same context, vary model tier; rate outcomes blind; record cost. `[INFERENCE]` **A positive result changes routing decisions and saves money**, which makes this the most practically valuable test here. - **H3 — tests the inverted-U he asserts.** Generate master prompts at controlled lengths from the same underlying profile; measure output quality against length. **Confound:** longer prompts contain more information *and* more distraction; **hold information constant and vary only verbosity** to separate the two. - **H4 — the one that tests Anita's observation**, and it is measurable. Same question, many runs, with and without a master prompt; code responses for framing diversity. `[INFERENCE]` **A confirmed narrowing effect would justify the keep-an-unpersonalised-session rule as a standing practice rather than a hedge.** - **H5:** Vary context-bundle scope from minimal to comprehensive on identical tasks. **Confound:** "relevant" is defined by the experimenter, so the minimal condition can be constructed to win — **fix the selection rule in advance and have someone else apply it.** - **H6 — first, find the paper.** `[RECOMMEND]` **This is the highest-value follow-up in the batch.** Six sources here have claimed judgment erodes with delegation and none supplied a mechanism or a remedy. `[INFERENCE]` If this meta-study exists and identifies *curation versus passivity* as the variable determining direction, it is the missing piece — **and it is a citation-chase, not a study.** ## ❓ Questions - **Does the Brookings "Context Maxing" paper exist, and does it say what he says?** `[INFERENCE]` The session's strategic argument depends entirely on it, and it is the one thing here I could not check. - **How much does a master prompt narrow perspective, and does the advisor workaround actually restore variance** — or does it produce disagreement shaped by the same author? - Where is the length inflection point? He asserts one and gives no number. - `[INFERENCE]` If a master prompt is *"the AI's manual for how to work with you,"* **what happens when it is wrong about you** — outdated, aspirational, or self-flattering? He says *"describe your current life and work as it is today, not what you wish it was,"* and does not say how you would find out you had failed. - What enforces the boundaries he recommends putting in the master prompt? `[INFERENCE]` **Instructions are not controls.** - `[EVIDENCE]` What is the real adoption rate of master prompts? *"Top one percent"* is offered with *"I swear"* as the evidence. ## 🔮 Future Research - **Whether curation versus passivity determines the direction of the cognitive-agency effect.** `[INFERENCE]` **The most important open question in this batch.** Six sources claim judgment erodes; this is the only one proposing a variable that decides which way it goes. - **Context-quality/model-tier substitution curves.** `[INFERENCE]` Real money rests on this and the public evidence is anecdotal. - **Perspective narrowing under personalisation.** `[INFERENCE]` Structurally adjacent to filter-bubble research and, as far as I can tell, not yet studied for assistant personalisation. - **Whether writing a master prompt improves outcomes by improving the person's own clarity** rather than the model's context. `[INFERENCE]` A confound in H1 and arguably the more interesting finding if it holds. ## 💡 Implications - **Practice.** `[RECOMMEND]` Write the profile sections you are missing. Keep it to things stable over months. Put depth in project context, not standing instructions. **Keep an unpersonalised session for when you want to be surprised.** - **Policy.** `[INFERENCE]` The **AI-usage-intentions** section is the transferable governance idea: a written boundary on where the tool is *not* used, held at the assistant level rather than in a person's head. - **Security.** `[INFERENCE]` His connector framing — *what is the new capability, and how do I counterbalance the risk it introduces* — is exactly right and unimplemented. **Instructions in a prompt are not enforcement**, and this batch contains the architecture that is. - **Education.** `[EVIDENCE]` The War and Peace test is the most efficient teaching device in the session. `[INFERENCE]` And the framework's implicit critique of its own field — that 90% of AI content addresses the lowest-leverage layer — is worth stating out loud when teaching. - **Funding models.** `[INFERENCE]` One observation about the genre: this is a free 55-minute lesson followed by a 35-minute pitch, and **the lesson is genuinely complete** — the builder is given away and works without the cohort. `[INFERENCE]` That is a materially more honest structure than the withheld-value webinar it explicitly positions against, and it is worth naming as a positive example. ## 🧠 Learning Styles Assessment **Reading level estimate.** `[INFERENCE]` **Flesch–Kincaid grade 9–10** — spoken register, plentiful repetition, short declaratives punctuated by long run-ons. Hand estimate on ~13,950 words, ±1.5 grades. `[EVIDENCE]` Technical terms are consistently glossed — *context window*, *system prompt*, *harness*, *token maxing* — which is appropriate for a 500-person mixed audience. #### 👀 Visual learners - **Present:** `[EVIDENCE]` **Heavily visual and central to the argument** — a five-layer pyramid diagram, screen-shared Claude accounts, the builder document, and the PARA folder sidebar. - **Improve:** `[INFERENCE]` **None of it is captured**, which is the acute problem here: the pyramid *is* the framework, and the transcript conveys it only as an ordered list. `[EVIDENCE]` The narration is also deictic throughout — *"look what we have here in the sidebar,"* *"you can see there."* #### 🔊 Aural learners - **Present:** `[EVIDENCE]` Strong verbal signposting (*"layer one… layer two…"*), deliberate repetition of the central claim, and a stated structure at the top. - **Improve:** `[EVIDENCE]` **The video has no captions at all** — confirmed by two independent tools. `[INFERENCE]` So the aural channel is the *only* channel, which is assessed properly below. #### 🗣️ Verbal learners - **Present:** `[EVIDENCE]` **The strongest dimension.** *"The AI's manual for how to work with you, the user."* *"An inch deep and a mile wide."* *"You escape mediocrity through unique context."* `[INFERENCE]` Each compresses a real distinction, and the War and Peace example is the best of them. - **Improve:** `[EVIDENCE]` **The one term that does not survive contact with his own advice is "context maxing"** — it names the opposite of what he recommends, and it is the term most likely to be repeated. #### 🧩 Kinesthetic learners - **Present:** `[EVIDENCE]` **Genuinely well served.** A downloadable builder, an explicit install path (copy, settings, paste), and assigned homework — run it on a task that previously failed. - **Improve:** `[INFERENCE]` The builder link is spoken and passed through chat, which does not survive into the recording — the session's central artifact has no durable pointer in the video itself. #### 🧠 Logical learners - **Present:** `[EVIDENCE]` The layered model is coherent, the leverage argument follows from it, and the demo is a **deliberately isolated variable**. - **Improve:** `[EVIDENCE]` **Two internal contradictions a logical viewer will catch**: *context maxing* against *"the smallest amount of context"*, and **Gall's Law against the hundred-question maximalist builder offered minutes later.** `[INFERENCE]` And *"a hundred times more valuable"* from a single unblinded comparison is a quantification the evidence does not support. #### 🧑‍🤝‍🧑 Social learners - **Present:** `[EVIDENCE]` **Unusually well served.** ~500 live attendees, an upvote-sorted Q&A panel, and the cohort framed as *"crowdfunded, real-time R&D"* rather than transmission. `[EVIDENCE]` **The best question in the session came from an attendee**, which the format is what surfaced. - **Improve:** `[INFERENCE]` Nothing substantive. The upvote mechanism is a genuinely good design for a session this size. #### 🧘‍♀️ Solitary learners - **Present:** `[EVIDENCE]` The builder is self-contained and works alone, which is the strongest thing here for someone not joining the cohort. - **Improve:** `[EVIDENCE]` **The one citation is unverifiable as given** — no authors, no date. `[INFERENCE]` A solitary learner told to *"really go check it out"* has a title and a search box, and it is the source the argument rests on. #### 🌿 Naturalistic learners - **Present:** `[EVIDENCE]` The recurring metaphor is agricultural and consistent — *"curating and cultivating your own context,"* context as *"fuel,"* *"a treasure trove, an absolute gold mine."* `[EVIDENCE]` And **Gall's Law is itself an evolutionary principle**: *"systems are grown, not installed from scratch."* - **Improve:** `[INFERENCE]` The metaphor is doing real work rather than decorating — **grown-not-installed is the session's soundest organising idea**, and it is the one his maximalist option violates. ## 📊 Accessibility Assessment Assessed against **WCAG 2.2 AA**. `[INFERENCE]` **Scope limit:** recovered transcript and metadata only. I did not view the video, so slide legibility, contrast, and player behaviour are **not assessed**; absence of a finding is not a pass. #### Screen reader semantics - **Not applicable** — video, not document. #### Keyboard navigation - **Not testable from the captured content.** #### Contrast and readability - **Not testable** — no visual content captured. `[INFERENCE]` A session built on shared slides and two screen-shared browser windows has substantial exposure here. #### Text resizing and reflow - **Not applicable** to a video. #### Media and captions - **Improve:** `[EVIDENCE]` **THIS VIDEO HAS NO CAPTIONS AT ALL.** Confirmed by two independent tools: Reader's scrape returned *"Unfortunately, Youtube does not provide subtitles for this video,"* and `yt-dlp --write-auto-subs --sub-langs en` reported *"There are no subtitles for the requested languages."* - `[INFERENCE]` **This is the most severe accessibility failure in this batch, and it is categorically different from the others.** Every other video here had captions that were *wrong* — corrupted proper nouns, mistaken technical terms. **This one has none.** **WCAG 1.2.2 Captions (Prerecorded), Level A — outright failure, not degradation.** - `[INFERENCE]` A deaf or hard-of-hearing viewer gets **nothing** from a 90-minute session with ~500 live attendees, presented by an author whose entire business is knowledge accessibility. `[EVIDENCE]` And it is not a resourcing problem: the session was professionally produced, recorded, and distributed across YouTube, a Circle community, and email — `[STATED]` *"there will be many ways to access this recording."* **There is one way it cannot be accessed.** - **Improve:** `[EVIDENCE]` **Audio description need is also high and unmet.** The pyramid diagram is the framework, the demo is the evidence, and the narration names neither — *"look what we have here in the sidebar,"* *"you can see there."* - `[RECOMMEND]` **A transcript exists** — recovered here via a third-party API in one call. `[INFERENCE]` Uploading it as a caption track is a minutes-long task, and it would move this from a Level A failure to compliant. #### Forms and interactions - **Not applicable** to the video. `[EVIDENCE]` The live Q&A used Zoom's Q&A panel with upvoting, which is more structured than open chat and likely better for participants who find live chat hard to follow. #### Standards reference - **1.2.2 Captions (Prerecorded), Level A** — **outright failure.** No captions exist. Confirmed by two independent tools. - **1.2.3 / 1.2.5 Audio Description** — **likely failure.** Slide-and-demo-dependent content with deictic narration. - **3.1.3 Unusual Words (AAA, aspirational)** — **satisfied.** Technical terms are glossed on introduction. ## 🌍 DEIB Assessment #### Cultural breadth - **Present:** Global live audience implied by the scale; no region-locked examples. - **Improve:** `[INFERENCE]` The assumed reader is a knowledge worker with a business, an email list, a team, and two paid AI subscriptions. `[EVIDENCE]` The demo question is *"what are the best ways I could grow my business"* and the master-prompt sections include *"who is your manager"* and *"your income sources."* **The framework generalises; the examples assume a narrow professional class.** #### Inclusive language - **Present:** Neutral throughout; questioners addressed by name and treated seriously. - **Improve:** `[INFERENCE]` *"You escape mediocrity through unique context"* and the framing of AI defaults as *"the average of the entire internet, which is another name for mediocrity"* is more dismissive than the argument requires. **The point — that defaults are generic — does not need "mediocrity" to land.** #### Neurodiversity support - **Present:** `[EVIDENCE]` **Strong.** Explicit numbered structure, heavy repetition of the central claim, stated agenda, and a predictable layer-by-layer progression. `[INFERENCE]` The **upvoted Q&A** is also a real accessibility feature — it removes the need to compete in a fast-moving chat, which is a genuine barrier for many people. - **Improve:** `[EVIDENCE]` **The complete absence of captions** removes the option of reading along, which is the accommodation many people rely on for a 90-minute talk. #### Representation - **Present:** `[EVIDENCE]` Attendee questions are surfaced by vote and answered by name — Jared, Anita, Manu, David, Sandy. `[INFERENCE]` **The best insight in the session came from an attendee**, and he credits her for it explicitly. - **Improve:** `[EVIDENCE]` The worked example is entirely his own business, and the sample master prompt is for a fictional character. `[INFERENCE]` **No real second person's master prompt is shown**, so the only demonstrated instance is the instructor's. #### Psychological safety - **Present:** `[EVIDENCE]` **Handled well.** He explicitly disclaims the expert-transmission model — *"this cohort is not in any way like 'I have the knowledge, I am the instructor'"* — treats the hardest question as the best one, and answers a privacy concern by conceding the difficulty rather than reassuring past it: *"I'm personally pretty concerned about security, privacy and risk, and it's a subtle question."* - **Improve:** `[INFERENCE]` *"If you have any version of a master prompt, I swear you're in the top one percent"* is scarcity framing in a sales context. Mild, and it is the one place the pitch shapes the teaching. ## ✅ Logical Validity #### Strong reasoning - **The training-data argument is the session's best, and it is an argument rather than an assertion.** Summarisation works, verbatim recall fails, therefore the model holds compressed structure rather than an indexed corpus, therefore specifics must be supplied. **Each step follows, and the premise is verifiable by the listener in seconds.** - **The layered leverage model is coherent and correctly ordered.** Layers 1–2 uncontrollable, 5 low-leverage because four layers have already shaped the signal, so 3–4 is where effort belongs. `[INFERENCE]` The conclusion is genuinely counterintuitive relative to how the field talks about itself, and it is derived rather than asserted. - **The system-prompt observation is a real insight into product design as constraint:** vendors optimise for quick satisfying answers, that is reasonable product design, and it is wrong for hard questions — so part of the master prompt is a correction to a decision made for you. - **The demo isolates its variable deliberately.** Connectors off, memory off, master prompt the only difference. `[INFERENCE]` Rare care for a sales demo. - **The diminishing-returns answer is mechanistic**, not hedging: more candidate details means more selection errors, and past a point the additions are noise. #### Fallacies - **Term contradicts practice — "context maxing."** He advocates it and then recommends *"the smallest amount of context."* **Weakness:** the name will outlive the nuance, and it names the failure mode. **Fix:** call it curation. - **Unverifiable citation carrying the strategic argument** — the Brookings paper, no authors or date. **Weakness:** the two-paths framing is the reason the product exists, and it rests entirely on a source the listener cannot check. - **Overclaim from a single comparison** — *"a hundred times more valuable"* from one unblinded self-assessed test. **Fix:** report the comparison and drop the multiplier. - **Self-contradiction — Gall's Law versus the maximalist builder.** *"A complex system designed from scratch never works,"* then a 100-question, six-to-eight-hour intake as the premium option. **Weakness:** the contradiction is four minutes wide and unacknowledged. - **Unsourced adoption claim** — *"top one percent,"* evidenced by *"I swear."* - **Instruction mistaken for enforcement** — the boundaries proposed for connector risk are *"rules and guidelines"* in a prompt. `[INFERENCE]` **A model can ignore, lose, or be talked out of an instruction.** The risk framing is right; the control is not one. #### Tone of voice `[EVIDENCE]` Enthusiast-teacher — *"I've made it my mission to proclaim the MasterPrompt message to the entire world"*, *"I'm such a fanatic about it."* `[INFERENCE]` The evangelism is self-aware enough to be disarming, and it is the register of someone selling a course, which he says plainly at the outset. `[EVIDENCE]` **Confidence is unevenly calibrated in a specific pattern:** the mechanical claims are stated carefully and hedged where appropriate (*"maybe a quarter,"* *"usually takes something like an hour"*), while the **quantified value claims are the loosest** (*"a hundred times,"* *"top one percent,"* *"infinitely more"*). `[INFERENCE]` **The pattern is diagnostic — precision where he has experience, inflation where he is selling.** A listener can use the hedging as a signal for which is which. `[EVIDENCE]` The clearest mark in his favour is structural: he gives the lesson away complete, including the builder, before pitching. `[INFERENCE]` **The teaching does not depend on the purchase**, which is the difference between a preview and a withheld-value webinar — and he names that distinction explicitly at the start. ## 🛡️ DARVO Analysis **No DARVO pattern present.** `[EVIDENCE]` DARVO — Deny, Attack, Reverse Victim and Offender — describes a response to accusation. This is instructional content with no accusation, no accused party, and no adversarial dimension. Checking each element explicitly rather than omitting the section: - **Deny:** `[EVIDENCE]` **The opposite in the places that matter.** He concedes that master prompts can be too long and degrade results, agrees immediately with the attendee who identifies the perspective-narrowing problem, and refuses to minimise the connector risk — *"they're also like a security and privacy disaster… I'm personally pretty concerned."* `[INFERENCE]` Conceding a structural weakness in your own method, raised by a customer, on a sales call, is the inverse of the denial element. - **Attack:** None. `[EVIDENCE]` The criticism targets **practices** — token maxing, prompt-engineering content, passive default use — and **no person or company is disparaged.** The one product judgment (*Gemini's "fall from grace"*) is unsupported but not hostile, and the vendors whose system prompts he critiques are described as making *"totally reasonable"* product decisions. - **Reverse victim/offender:** None. `[INFERENCE]` One adjacent observation, recorded separately because it is **not** DARVO: this is a sales event, and the scarcity line — *"if you have any version of a master prompt, I swear you're in the top one percent"* — is the one place the commercial frame visibly shapes the teaching. **The frame is disclosed in the first two minutes**, the lesson is delivered complete, and the artifact is given away, which is what keeps it transparent. Already logged under Logical Validity. ## 🎯 Conclusion **Bottom line.** `[INFERENCE]` A genuinely useful framework, one badly-chosen term, one unverifiable citation carrying the argument, and the best insight coming from the audience. 1. **The training-data argument is the keeper.** `[STATED]` Summarise *War and Peace* — fine. *"What is the last sentence in chapter three"* — impossible, *"about any book, even the most well-known books, because it doesn't actually have direct access to its training data."* `[INFERENCE]` **This converts note-taking from a discipline into a consequence of retrieval architecture**, and it is checkable in thirty seconds. 2. **The five-layer model is sound and its main claim is counterintuitive:** prompting is the *lowest*-leverage layer, because by the time you type, four layers have already shaped the answer. `[INFERENCE]` That is also an implicit critique of the field — most AI education addresses layer 5 exclusively. 3. **"Context maxing" is the wrong name for what he does.** `[EVIDENCE]` Asked directly, he says a master prompt can absolutely be too long — *"at some point it becomes noise"* — and describes assembling per task as **"the smallest amount of context that I can give it… rather than the biggest."** `[INFERENCE]` **That is curation, and *maxing* names the failure mode he spends the Q&A warning against.** It also matches the standing rule to load the smallest useful context — a third independent convergence in this batch. 4. **The structural cost is identified by an attendee, not the instructor.** `[STATED]` Anita: *"the more information I add… the more Claude sounds like me… but sometimes I do want a genuinely different point of view."* `[INFERENCE]` A master prompt makes a model reason like you, **biases included**, and the better it works the harder it becomes to see what your framing excludes. His AI-advisors answer is a workaround: **a disagreement you designed is not one you did not anticipate.** **The free mitigation is keeping an unpersonalised session for when you want to be surprised.** **Two things to check before repeating.** `[EVIDENCE]` The **Brookings "Context Maxing"** paper — no authors, no date — is carrying the session's entire strategic claim, that passive use degrades cognitive agency while curated use enhances it. `[INFERENCE]` **If it exists and says that, it is the most important source in this batch:** six sources here have claimed judgment erodes under delegation and **none has proposed a variable that determines the direction.** This would be it, and finding it is a citation chase rather than a study. And **Gall's Law** — *"a complex system designed from scratch never works"* — is quoted as a billboard-worthy principle four minutes before a hundred-question builder is offered as the premium path. **Where this lands for the `CLAUDE.md` file.** `[INFERENCE]` It already *is* a master prompt, and it already implements the corrections he recommends — *ask when requirements are unclear* counteracts the do-your-best default, *tell that you're guessing* counteracts the confident-answer default, *load the smallest useful context* is his layer-4 rule. **The gap is the profile, not the corrections:** working style, communication preferences, tool preferences, personal SWOT, what quality means, and explicit **AI-usage intentions** — a written boundary on where the tool is *not* to be used, which is the section least likely to occur to anyone and the most useful to have. **And the accessibility finding, which is the batch's worst.** `[EVIDENCE]` **This video has no captions at all** — confirmed independently by Reader's scrape and by `yt-dlp`. Every other video in this batch had captions that were *wrong*; this one has **none**. **WCAG 1.2.2, Level A, outright failure** — from a professionally produced 90-minute session with ~500 live attendees, distributed across YouTube, Circle, and email, by someone whose business is knowledge accessibility. `[INFERENCE]` A transcript exists — one API call produced it here. **Uploading it is a minutes-long task.** **Practical next steps.** 1. **Find the Brookings paper.** Highest-value follow-up in the batch. 2. Add the missing profile sections; apply the *stable for months, not weeks* filter. 3. Adopt the monthly propose-and-review amendment cycle, expecting ~25% acceptance. 4. **Keep an unpersonalised session available** — the mitigation for the one cost nobody selling this will emphasise. <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/The Master Prompt 2.0 Live Session with Tiago Forte.md|The Master Prompt 2.0: Live Session with Tiago Forte by Tiago Forte (Forte Labs)]] ## Highlights - If you ask, you know, ChatGPT or Claude to summarize the book War and Peace, they can do it. it can summarize it a hundred different ways it can tell you all the themes it can tell you all the lessons and talk about all sorts of the characters and different things but then ask the ai well what is the last word on page five or what is the last sentence in chapter three and it can't tell you it can't tell you that about any book even the most well-known books because it doesn't actually have direct access to its training data ([View Highlight](https://read.readwise.io/read/01m12g9mjm1ay8vg7ycgwyxcf8)) - Note: THE SHARPEST ARGUMENT FOR NOTE-TAKING I'VE SEEN MADE IN AN AI CONTEXT — because it's a fact about RETRIEVAL ARCHITECTURE, not an exhortation about discipline. Summarisation works; verbatim recall fails; therefore the model holds COMPRESSED STATISTICAL STRUCTURE, not an indexed corpus; therefore specifics must be supplied. Each step follows, and THE PREMISE IS VERIFIABLE IN THIRTY SECONDS — ask any model for a specific sentence from a famous book and watch. HIS CONCLUSION: "If you wanted to write something really novel and interesting and unique about the book War and Peace, you would need to read War and Peace and supply notes… Thus the need for note-taking." THIS CONVERTS "TAKE NOTES" FROM ADVICE INTO A CONSEQUENCE. The usual argument is about memory or diligence; this one is about what the machine structurally cannot do for you — which makes it the rare productivity claim that survives contact with how the thing actually works. CONTEXT: this is layer 1 of his five-layer pyramid — training data → vendor system prompt → your master prompt → project knowledge → chat window. Layers 1–2 are uncontrollable; LAYER 5 IS THE LOW-LEVERAGE ONE ("by the time you're typing something into a chat box, it's kind of too late"); the leverage is in 3–4. That ordering is counterintuitive relative to how the field talks about itself — most AI education addresses layer 5 exclusively, which makes the framework's central claim also a critique of its own genre. #atom/zettel/z1 #action/h1 - anita asks i sometimes find that the more information i add to the master prompt and the context layers the more claude sounds like me which is sometimes great but sometimes i do want a genuinely different point of view ([View Highlight](https://read.readwise.io/read/01m12gagrf9vej84084rh2rz1n)) - Note: THE BEST THING SAID IN THE SESSION, AND IT CAME FROM AN ATTENDEE. This identifies the STRUCTURAL COST of the entire approach, and the instructor never raised it. A master prompt is by construction an instrument for making a model reason like you — Forte agrees immediately: "You're almost like giving it your personality, giving it your perspective, GIVING IT YOUR BIASES, giving it your opinions." BIASES INCLUDED, WHICH IS THE PROBLEM. The better the master prompt works, the more reliably the model reproduces your framing, and THE HARDER IT BECOMES TO NOTICE WHAT YOUR FRAMING EXCLUDES. That's an echo chamber built deliberately and installed at the deepest layer you control. HIS ANSWER — a set of AI advisors with contrasting personalities, adjustable along Big Five traits — IS A WORKAROUND, NOT A RESOLUTION, and it's worth being precise about why: THE ADVISORS ARE ALSO AUTHORED BY YOU, so they encode YOUR MODEL OF WHAT DISAGREEMENT LOOKS LIKE. A disagreement you designed is not the same as a disagreement you didn't anticipate. THE HONEST VERSION: a master prompt trades BREADTH OF PERSPECTIVE for FIT. That trade is usually correct. THE MITIGATION IS FREE — KEEP AN UNPERSONALISED SESSION AVAILABLE for the cases where you want to be surprised. A second account or a fresh session provides it at zero cost. LANDS ON THIS BATCH'S SYCOPHANCY THREAD: GitHub alt-text (a model reviewing others always finds fault) + Dreams of Code (a model reporting its own work always says done) + THIS — a model that agrees with you because you told it how you think. Three directions, one failure. #atom/zettel/z3 #action/h2 - I'm never giving it this whole stack of folders. That would be crazy. I'm drilling into actually what is the smallest folder? I might even sometimes do a subfolder in here. What is the smallest amount of context that I can give it to solve the task at hand rather than the biggest amount of context that I can give it? ([View Highlight](https://read.readwise.io/read/01m12ga2vr267k64vxgfqsjvcv)) - Note: "CONTEXT MAXING" IS THE WRONG NAME FOR WHAT HE ACTUALLY DOES — and this is the sentence that proves it. He brands the approach "context maxing" against "token maxing," then in the Q&A recommends the opposite of maximising. Asked whether a master prompt can be too long: "Yes, you can absolutely have too much… The more details it has to look through to decide which are relevant, THE HIGHER THE ODDS IT CHOOSES THE WRONG ONES… at some point it becomes noise, and noise actually worsens the quality of your results." THE PRACTICE IS RIGHT AND THE NAME IS WRONG. What he describes is CONTEXT CURATION: a small stable profile at layer 3, plus modular task-scoped bundles at layer 4, assembled on demand. "Maxing" names the failure mode he spends the Q&A warning against — and it's the term people will remember. THIS MATCHES MY OWN CLAUDE.md RULE ALMOST WORD FOR WORD: "load the smallest useful context — grep/index/paths first, targeted reads second, whole files last." THIRD INDEPENDENT CONVERGENCE IN THIS BATCH, after the deterministic-script rule (Serge/Dodds: "turn those manual workflows into deterministic code… to save on that context") and the verify-delegated-work rule (Jacobi's three questions). ALSO NOTE THE MECHANISM HE GIVES, which is better than the usual hand-waving about context windows: it's not that the model runs out of room, it's that MORE CANDIDATES MEANS MORE SELECTION ERRORS. That's a different and more useful claim — it predicts degradation well below any capacity limit. #atom/zettel/z2 - Gall's Law, the short version, is systems are grown, not installed from scratch. Okay. And the full quote, the full version is this. A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works. ([View Highlight](https://read.readwise.io/read/01m12gb0hf2j9rzxcyykerkbk9)) - Note: QUOTED ACCURATELY, TREATED AS BILLBOARD-WORTHY — AND CONTRADICTED FOUR MINUTES LATER. He says if he could take out a Super Bowl commercial this is what he'd put on it. Then: "some of you are maximalists… you want to make the ultimate, totally thorough master prompt. And if that's you, I have a MAXIMALIST VERSION of this master prompt builder that has like 100 QUESTIONS. If you really want to go ham and spend probably SIX TO EIGHT HOURS answering all those." A HUNDRED-QUESTION INTAKE PRODUCING A DOCUMENT DESIGNED FROM SCRATCH IS PRECISELY WHAT GALL'S LAW SAYS WILL NOT WORK, and it's offered as the premium option. The contradiction is four minutes wide and unacknowledged. THE MVP PATH IS THE ONE CONSISTENT WITH HIS OWN PRINCIPLE: 1,000–2,000 words, then grow it. AND THE GROWTH MECHANISM IS THE GOOD HALF — a monthly review where you ask the model to inspect connectors, integrations and local files and propose amendments. His acceptance rate, volunteered: "I ONLY ACCEPT MAYBE A QUARTER of the ones that it recommends." THAT 25% IS THE MOST CALIBRATED NUMBER IN THE SESSION — it says the tool is useful AND unreliable, which is the correct characterisation of automated suggestion, and it's the same human-keeps-judgment / machine-does-volume split this batch found in five other domains. Compare the loosest numbers in the same talk ("a hundred times more valuable," "top one percent," "infinitely more") — THE HEDGING IS DIAGNOSTIC: precision where he has experience, inflation where he's selling. Useful signal for reading anything like this. #atom/zettel/z4 #fallacy/self-contradiction