3 ChatGPT Prompts I Use to Standout at Work  🔗 **Source:** [3 ChatGPT Prompts I Use to Standout at Work](https://www.youtube.com/watch?v=pUG7_03G_Ac) > [!note]+ Document Notes > tags: [problem-framing, prompt-engineering, mental-models, critical-thinking, knowledge-work, ai-agents] > author: Vicky Zhao > source: YouTube > reference: https://www.youtube.com/watch?v=pUG7_03G_Ac > date created: 2026-08-31 15:00 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: 3 ChatGPT Prompts I Use to Standout At Work > <!-- 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 - [The video under analysis](https://www.youtube.com/watch?v=pUG7_03G_Ac) — Vicky Zhao, 18:43, published 2025-05-05 - [Quote Investigator on the "55 minutes on the problem" attribution](https://quoteinvestigator.com/2014/05/22/solve/) — the Einstein quote she cites is unsubstantiated - [Wispr Flow](https://wisprflow.ai/) — the system-wide dictation tool she recommends - [Her 17-minute AI workflow video](https://www.youtube.com/watch?v=yqq_U2fxd2U) — the companion piece she references ## 📝 Summary - **What happened.** Zhao argues that the way to stand out with AI at work is not faster answers but better *problems*. She contrasts two user types — those who put ten words in and expect a thousand out, and those who put a thousand in and get a thousand back — and argues the second produces quality the first cannot. She then gives three problem-refinement prompts: surface **assumptions**, go **five levels deep** on why, and ask for **alternatives** ("what else could be true?"), demonstrating all three live on a real problem from a university guest lecture she was preparing. - **Why it matters.** The central claim is correct and under-taught: most people use AI as an answer machine, when the higher-leverage use is as an instrument for reframing the question. Her diagnosis of *why* — that professional training produces excellent problem-solvers who have never practiced problem-identification, because *"most employees are good students"* — is a genuinely sharp observation that connects the school-to-work pipeline to a specific AI-usage failure. - **Key stakeholders.** Knowledge workers at companies making AI use a baseline expectation (she names Google, Amazon, Shopify). Anyone whose value depends on judgment rather than throughput. Wispr Flow, a paid tool she recommends with a link and no stated commercial relationship. And viewers who will not reproduce her results, for a reason she discloses honestly and does not fully follow through. - **Risks and tradeoffs.** The Einstein quote anchoring her argument is apocryphal and easily checked. Her most interesting demo results come from a model that has accumulated months of her own frameworks in memory — she says so plainly, which is rare and creditable, but she does not draw the harder implication: when a personalized model returns your own vocabulary, "alternatives" may be less alternative than they appear. And enthusiasm does a lot of the evaluative work across the demo. - **Bottom line.** The best framing in this batch of *what* to use AI for, paired with three reusable prompts that will improve most people's output immediately. Take the frameworks. Notice that the demo is partly a mirror, and build in your own falsification step, which the video never does. ## 📖 Main Content **The framing, and the good counterintuitive move.** She opens on a real pressure — Google, Amazon, and Shopify treating AI usage as baseline expectation — and then dismisses two things people obsess over. Not syntax: *"It's not about the fancy syntax and the symbols you have to use."* Not tool choice: *"is should I use both? Should I use all of them? Those things don't matter."* What matters is signal-to-noise. Then the sharpest structural observation in the video. There are two kinds of LLM user. The first puts ten words in and expects a thousand out, and experiences that as efficiency. The second puts a thousand words in and gets a thousand back — which sounds like a poor trade until you notice that *"the result that you get from ChatGPT on this will be at such a higher quality than what you can write by yourself within that kind of time."* The reason this lands is the premise underneath it: knowledge workers are not paid by the hour. *"your leverage doesn't come from time... what your value is is the quality of the idea, how far you can take something that others can't see."* If that is true, optimizing an AI interaction for speed optimizes the wrong variable. She names both failure modes fairly — the expert who tries it once, sees mediocre output, and quits; and the person who does not care about quality and checks the box. Neither is a strategy. **The core thesis.** *"most people are looking at ChatGPT as a source for answer but actually what the 1% does is they look at AI as a way to refine the problem."* She anchors it with the line about spending fifty-five minutes on the problem and five on the solution, attributed to Einstein. **That attribution is wrong, and it is checkable.** The quote has been traced repeatedly and has no basis in Einstein's writings or documented speech; Quote Investigator finds no evidence before decades after his death. It is a small thing in a video that does not depend on it — the point stands entirely on its own — but it is the load-bearing rhetorical anchor of the central claim, and it took under a minute to check. **The diagnosis, which is the best idea here.** Her guest-lecture story does real work. The students were smart, fast, articulate, many of them honors students — and *"they are fantastic at taking a problem and solving it. But they don't have experience coming up with problems, identifying problems, reframing problems."* She then extends it to work, and the extension is the insight: *"the boss gives us the problem. We say okay I'm going to tell you exactly how to do this. We don't even think whether the problem brought to us is wrong. Most employees are good students."* That last sentence explains a great deal. Twenty years of schooling optimizes for producing correct answers to well-formed questions supplied by an authority. That skill transfers directly into most jobs and is exactly the skill models now perform cheaply. What has not been trained — and what remains scarce — is deciding whether the question is the right one. Her conclusion follows cleanly: differentiation now lives in problem definition, and the assets AI lacks are *"your experience, your judgment, your intuition that you've built over the years."* **The three prompts.** Concrete, reusable, and in a sensible order. *Assumptions* — surface what is going unsaid. Her prompt asks for ten assumptions the audience would not recognize they hold. Returns include "mental models are not answers, they are lenses," "you're already using mental models, you just don't know it," and the distinction between simplifying and dumbing down. She notes some she would keep and some she would reword, and — importantly — *"you don't stop here... I want to tweak and what I can do is continue to have a conversation."* The prompt is a first draft, not an output. *Five levels deep* — the consulting five-whys, pushed from superficial toward fundamental. The chain it produces is the strongest sequence in the demo: students have not been taught mental models → their environment rewards correctness over framing → uncertainty feels like failure rather than opportunity → they are protecting an identity as smart students → they were trained to optimize for performance rather than judgment. She stops on that last link, and correctly: *"how do I phrase this... all of us are focused or overfocusing on performance"* while judgment — deciding among options without a complete answer — *"is what separates the top performers from the top doers."* That is a real distinction, arrived at through the process she is demonstrating rather than asserted beforehand, which makes the demo more persuasive than a scripted one would be. *Alternatives* — *"what else could be true?"* Deliberate counterweight to narrow framing. **The disclosure that deserves credit, and the implication she misses.** Before showing the third result, she stops: *"I do have to say that I use ChatGPT a lot so it knows me. It knows my frameworks. It knows my mental models. That's why it's giving these inversion and red teaming. So if your ChatGPT doesn't do that yet it's not a problem. It's just because it doesn't have the memory."* This is genuinely good practice and rare in the genre. Most creators demonstrate on a heavily personalized account and let viewers assume the results are reproducible. Naming the difference protects the viewer from concluding their own tool is broken. But the harder implication goes unexamined. If the model is returning inversion and red-teaming *because she taught it those frameworks*, then the "alternatives" it generates are partly her own vocabulary reflected back. For the first two prompts that is fine — surfacing your own assumptions is the point. For the third it is a problem, because the entire purpose of asking "what else could be true?" is to escape your existing frame, and a model tuned to your frame is the wrong instrument for that job. She has the fact and does not follow it to the conclusion. **What she does well with the output.** She filters critically rather than accepting. Offered the hypothesis that students resist mental models because they associate them with *"consulting jargon or elite circles,"* she rejects it on domain grounds — these students are in finance and risk management, where that vocabulary is native. That is judgment applied to model output, modeled on camera, and it is more instructive than any of the prompts. She also flags genuine novelty honestly. Of *"they never seen real world ambiguity modeled by authority figures"* she says *"this is something I don't know"* — distinguishing what confirmed her existing view from what did not. Several other results she notes as matching things she had already put in the lecture, which is confirmation rather than discovery, and she is transparent about that too. **The digression worth keeping.** She veers off to argue that articulateness is not fast thinking: *"the best authors when they give interviews is not because they can think on the spot. It's because they spent years writing out these ideas refining them... They put in the work. It's just we don't see it."* That is Cal Newport's invisible staircase from elsewhere in this same batch, arrived at independently and applied to a different phenomenon. Two creators with no apparent connection converging on the same explanatory structure — the hidden accumulation misread as innate capacity — is worth noting. **Tools mentioned.** She recommends ChatGPT's built-in dictation as *"99% accurate"* and useful for getting into conversational mode, which is how ten words become a thousand. She notes its weakness with uncommon spellings, including her own name, and recommends **Wispr Flow** as a system-wide alternative, crediting Terry for the referral and promising a link. There is no statement of a commercial relationship in either direction — not a sponsorship disclosure, and not a denial of one. Given that the other Zhao video in this batch explicitly says *"this isn't sponsored,"* the absence here is a small inconsistency worth noting rather than an accusation. **Contested and uncertain.** "99% of people" and "the 1%" are rhetorical figures, not measurements. The demo never tests whether any surfaced assumption is actually *true* of the students — only whether it is interesting, which is a different standard and the one the video's own thesis should have caught. And the entire validation is her own reaction, in a domain where she is the expert, using a model she has personalized. ### 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=0) The setup, and two dismissals - Context: Google, Amazon, Shopify treating AI usage as baseline expectation. - **Not** about syntax — *"It's not about the fancy syntax and the symbols."* (Offers a screenshot for basics.) - **Not** about tool choice — ChatGPT vs Claude, *"Those things don't matter."* - What matters: improving signal amid noise. #### ⏱️ [00:50](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=50) ⭐ Ten words in vs. a thousand words in - Type one: 10 words in → 1,000 out. Feels efficient. - Type two: 1,000 in → 1,000 out. Sounds like a bad trade; produces quality you could not write alone in that time. - **The premise that makes it work:** knowledge workers are not paid hourly. *"your leverage doesn't come from time... what your value is is the quality of the idea."* - Both failure modes named fairly: the expert who tries once, sees mediocrity, quits; and the box-checker who does not care about quality. #### ⏱️ [02:58](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=178) The thesis - *"most people are looking at ChatGPT as a source for answer but actually what the 1% does is they look at AI as a way to refine the problem."* - ⚠️ **Anchored on an apocryphal quote.** The "55 minutes on the problem, 5 on the solution" line is attributed to Einstein; it has no basis in his writings or documented speech ([Quote Investigator](https://quoteinvestigator.com/2014/05/22/solve/)). The argument does not need it. #### ⏱️ [03:20](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=200) ⭐ Why good students fail at this - Guest lecture: students smart, fast, articulate, many honors — *"fantastic at taking a problem and solving it. But they don't have experience coming up with problems."* - The extension to work is the insight: *"the boss gives us the problem... We don't even think whether the problem brought to us is wrong. **Most employees are good students.**"* - Twenty years of schooling optimizes for correct answers to well-formed questions from an authority — exactly the skill models now perform cheaply. - What remains scarce: *"your experience, your judgment, your intuition that you've built over the years."* #### ⏱️ [04:43](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=283) The three frameworks 1. **Assumptions** — surface what goes unsaid. 2. **Five W's / five levels deep** — consulting method, superficial → fundamental. 3. **Alternatives** — counter narrow focus. #### ⏱️ [05:30](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=330) Dictation as the mechanism - ChatGPT's built-in dictation called *"99% accurate"*; gets you into conversational mode, which is **how 10 words become 1,000**. - Weakness noted: uncommon spellings, including her own name. - Recommends **Wispr Flow** — works system-wide, not only in ChatGPT. Credits Terry for the referral; promises a link. - ⚠️ No statement of commercial relationship either way. Her Napkin video explicitly says *"this isn't sponsored"*; the absence of any such line here is a small inconsistency, not an accusation. #### ⏱️ [06:44](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=404) The worked problem - Real brief: lecture finance/risk-management students on using mental models under uncertainty, where the real world supplies no predefined problems. - Gives full context first, then asks ChatGPT to help her *understand the problem* rather than solve it. - First pass is decent: trained for an academic mind, entering a world that rewards adaptive minds; *"not a lack of intelligence"* but over-reliance on environments where the right answer is known. #### ⏱️ [08:52](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=532) Framework 1 — assumptions - Prompt: list 10 assumptions these students likely hold without knowing. - Returns she keeps: mental models are lenses not answers; *"You're already using mental models. You just don't know it"*; simplifying vs. dumbing down. - Notes one she would reword (mental models as situational rather than universal). - **Process point:** *"you don't stop here"* — keep some, tweak others, continue the conversation and give direction. The output is a first draft. #### ⏱️ [11:07](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=667) Framework 2 — five levels deep - ChatGPT observes she is doing what she teaches; she confirms. - **The chain:** not taught → environment rewards correctness over framing → uncertainty feels like failure → protecting an identity as smart students → trained to optimize for performance, not judgment. - ⭐ Her stop on the last link: judgment — choosing among options without a complete answer — *"is what separates the top performers from the top doers."* Doers operate in low-complexity environments. - Arrived at *through* the process rather than asserted beforehand, which is what makes the demo persuasive. #### ⏱️ [14:30](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=870) Framework 3 — alternatives, and the honest caveat - Prompt: *"what else could be true?"* - **Creditable disclosure:** *"I use ChatGPT a lot so it knows me. It knows my frameworks... That's why it's giving these inversion and red teaming. So if your ChatGPT doesn't do that yet it's not a problem. It's just because it doesn't have the memory."* - Rare and protective — it stops viewers concluding their own tool is broken. - ⚠️ **The implication she does not draw:** if the model returns inversion and red-teaming *because she taught it those*, the "alternatives" are partly her own vocabulary reflected back. Fine for surfacing your own assumptions; a problem for a prompt whose whole purpose is escaping your frame. #### ⏱️ [15:57](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=957) Filtering the output — the part worth copying - Keeps: *"They do use mental models but lack the vocabulary to name them"* — matches workshop feedback that naming enables recall and building. - **Rejects** the "consulting jargon or elite circles" hypothesis on domain grounds — these students are in finance, where that vocabulary is native. Judgment applied to model output, on camera. - Flags genuine novelty honestly: *"they never seen real world ambiguity modeled by authority figures... this is something I don't know."* - Transparent that several other results merely confirmed what she had already put in the lecture. #### ⏱️ [16:50](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=1010) The digression worth keeping - Articulateness is not fast thinking: *"the best authors when they give interviews is not because they can think on the spot. It's because they spent years writing out these ideas... They put in the work. It's just we don't see it."* - **This is Cal Newport's invisible staircase from elsewhere in this batch**, reached independently and applied to a different phenomenon. Two unconnected creators converging on hidden accumulation misread as innate capacity. #### ⏱️ [17:57](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=1077) Close - *"It really is a good student. Depending on what kind of question you throw at it, the answers it give will be very different."* - Points to her 17-minute AI workflow video on improving inputs. #### What the demo never does - **No falsification step.** Nothing tests whether any surfaced assumption is actually *true* of the students — only whether it is interesting. That is a different standard, and the video's own thesis should have caught it. - Validation is her own reaction, in her own domain, on a model she personalized. ## 📚 References - [Vicky Zhao, "3 ChatGPT Prompts I Use to Standout At Work," 2025-05-05](https://www.youtube.com/watch?v=pUG7_03G_Ac) — the source under analysis - [Quote Investigator — "If I had an hour to solve a problem"](https://quoteinvestigator.com/2014/05/22/solve/) — traces the Einstein attribution as unsubstantiated - [Wispr Flow](https://wisprflow.ai/) — the system-wide dictation tool recommended - [Vicky Zhao, "My 17 Minute AI Workflow To Stand Out At Work"](https://www.youtube.com/watch?v=yqq_U2fxd2U) — the companion input-quality argument - [Cal Newport, "How to Learn Any Hard Skill"](https://www.youtube.com/watch?v=RrLKXRXVrKM) — the invisible-staircase argument she independently reaches at 16:50 - [Five whys](https://en.wikipedia.org/wiki/Five_whys) — the root-cause method behind framework two ## 🗣️ Paraphrase - **Opening.** With major companies treating AI use as a baseline expectation, everyone is trying to keep up with changes that outpace what they just learned. The question is which tools matter and which are gimmicks, and the answer is not prompt syntax or which chatbot you pick — it is how you use AI to raise the ratio of signal to noise. - **The efficiency trap.** People fall into two groups. Some put ten words in and expect a thousand out, which feels efficient. Others put a thousand words in and get a thousand back, which sounds unproductive but yields far better quality than they could produce alone in the same time. Most knowledge workers do the former. - **Why that is the wrong optimization.** Knowledge workers are not paid by the hour, so leverage comes from the quality of ideas and from seeing further than others, not from speed. Companies differentiate, and so must individuals. Focus on output quality rather than time. - **The two failure modes.** Experts try the tool, judge the output poor, and abandon it without learning to use it well. Others accept low quality, check the box, and move on. Neither works long term. - **The reframe.** Most people treat AI as a source of answers; a small minority use it to refine the problem — echoing the maxim about spending most of an hour on the problem and only a little on the solution. - **The evidence.** At a university guest lecture, the students were quick, articulate, and academically strong — excellent at solving problems handed to them, inexperienced at identifying or reframing them. The same holds at work: a manager supplies the problem, employees execute, and nobody asks whether the problem is right, because most employees are good students. - **The consequence.** Standing out now depends on defining and reframing problems so everyone works on the right one — which is good news, because AI lacks the experience, judgment and intuition you have built. - **Three refinements.** Surface the assumptions that usually go unsaid. Use the five W's to move from superficial to fundamental. Ask for alternatives, since narrow focus hides other framings. - **A tool aside.** ChatGPT's dictation is highly accurate and encourages conversational back-and-forth, which is how short prompts become long ones, though it stumbles on unusual spellings. Wispr Flow works everywhere rather than only in ChatGPT. - **The example.** Preparing a lecture for finance and risk-management students on using mental models amid uncertainty, she supplies full context and asks for help understanding the students' problem. The first response identifies training for an academic environment where answers are known, entering a world that rewards adaptability. - **Assumptions.** Asked for ten assumptions the students would not recognize, it returns that mental models are lenses rather than answers, that everyone already uses them unknowingly, that they are situational rather than universal, and that simplifying differs from dumbing down. Some are keepers, some need rewording, and the conversation continues from there. - **Five levels deep.** Asked why students do not think this way, the chain runs: never taught, environment rewards correctness over framing, uncertainty feels like failure, identity as a smart student needs protecting, and training optimized performance rather than judgment. That last point stands out, since judgment — choosing well without a complete answer — is what separates top performers from capable doers. - **Alternatives.** Asked what else could be true, the model offers inversion and red teaming, which reflects long familiarity with her frameworks; newer accounts will not behave this way. Results include that students already use mental models but lack vocabulary to name them, that they may associate the idea with consulting jargon, and that they have never seen authority figures model real ambiguity. The jargon hypothesis seems unlikely for finance students; the last is genuinely new to her. - **A digression.** Articulate people are not fast thinkers; the best authors sound fluent in interviews because they spent years refining those ideas in writing. The work is invisible. - **Close.** Focusing on understanding and reframing the problem rather than extracting answers produces far more depth, because the model is a good student whose answers depend entirely on the question asked. ## 📰 Newsletter **Ten words in, a thousand words out. That's the trap.** Vicky Zhao's framing is the most useful thing I have read on *what to actually use these tools for*, and it starts by dismissing the two things everyone argues about. Not prompt syntax. Not ChatGPT versus Claude. *"Those things don't matter."* Here is what does. There are two kinds of user. One puts ten words in and expects a thousand out, and calls that efficiency. The other puts **a thousand words in and gets a thousand back** — which sounds like a terrible trade until you notice the output is better than anything you could write alone in that time. The premise underneath is the part worth internalizing: **you are not paid by the hour.** *"your leverage doesn't come from time... what your value is is the quality of the idea, how far you can take something that others can't see."* If that is true, then optimizing an AI conversation for speed is optimizing the wrong variable entirely. **Then the observation I have not been able to stop thinking about.** She did a guest lecture for finance students — sharp, fast, articulate, mostly honors. And they were *excellent* at solving problems handed to them and had no practice at all identifying or reframing one. Then she extends it: *"the boss gives us the problem... We don't even think whether the problem brought to us is wrong. **Most employees are good students.**"* That explains a great deal. Twenty years of schooling trains you to produce correct answers to well-formed questions supplied by an authority. That skill transfers straight into most jobs — and it is precisely the skill that models now do cheaply. What was never trained, and is now scarce, is deciding whether the question is the right one. **Three prompts, and they work.** Ask for the **assumptions** nobody is stating. Go **five levels deep** on *why*. Ask for **alternatives** — "what else could be true?" Her live demo on the five-whys is the best sequence: not taught → environment rewards correctness over framing → uncertainty feels like failure → they're protecting an identity as smart students → **they were trained to optimize for performance, not judgment.** She stops there, correctly, because judgment — picking well without a complete answer — is exactly what separates top performers from very capable doers. **Two things to carry with you.** First, credit where it is due: before showing her best result she stops and says *"I use ChatGPT a lot so it knows me. It knows my frameworks... if your ChatGPT doesn't do that yet it's not a problem. It's just because it doesn't have the memory."* Almost nobody does this. Most creators demo on a heavily personalized account and let you conclude your tool is broken. But follow it one step further than she does. If the model offers "inversion" and "red teaming" *because she taught it those*, then the alternatives it generates are partly her own vocabulary coming back. That is fine for surfacing your own assumptions — that is the point. It is a real problem for the third prompt, whose entire job is escaping your frame. **A personalized model is the wrong instrument for "what else could be true?"** Ask a cold session that one. Second, and this is the thing her own thesis should have caught: the demo never checks whether any of these assumptions are actually **true** of the students. Only whether they are *interesting*. Those are different standards, and "interesting" is the one that feels like insight. Add a falsification step — go ask three students — or you have refined the problem into a shape you like rather than the shape it has. **One correction.** The Einstein quote anchoring the whole argument — fifty-five minutes on the problem, five on the solution — is apocryphal. [Quote Investigator has traced it](https://quoteinvestigator.com/2014/05/22/solve/); there is no basis in anything Einstein wrote or is documented as saying. The argument does not need it and is stronger without it, which is usually true of borrowed authority. **And a nice piece of convergence.** She veers off to argue that articulate people aren't fast thinkers — *"the best authors when they give interviews... spent years writing out these ideas... They put in the work. It's just we don't see it."* That is Cal Newport's invisible staircase, arrived at independently, in a different video I read the same afternoon. Two people with no apparent connection landing on the same structure: hidden accumulation misread as innate capacity. When that happens it is usually because the idea is right. ## ✅ Action Items #### DevRel - Developer advocacy: before writing any explainer, run the three prompts on the topic to check we are answering the right question. `#action/h1` - Education: teach problem-framing explicitly. It is the scarce skill precisely because schooling never trains it. `#action/h1` - Developer experience: run the assumptions prompt against our own onboarding docs — what do we assume readers already know? `#action/h1` - Developer marketing: not applicable to this source. `#action/h3` - Community management: five-levels-deep is a good tool for recurring community conflicts, which usually have a stated problem and a real one. `#action/h2` - Internal/external/cross-project: normalize asking "is this the right problem?" before scoping. Most roadmap waste starts here. `#action/h1` #### OSPO - Policy & Governance: policy work is problem definition. Apply the three frameworks before drafting, not after. `#action/h2` - License Compliance: not applicable to this source. `#action/h3` - Security Management: "what else could be true?" is a threat-modeling prompt — use a cold session, not a personalized one. `#action/h2` - Software Inventory: not applicable to this source. `#action/h3` - Developer Support: support tickets state a solution more often than a problem. The assumptions prompt surfaces the real ask. `#action/h2` - Community Engagement: not applicable to this source. `#action/h3` - Strategic Alignment: **strong fit.** Most strategic misalignment is teams solving well the problem nobody checked. `#action/h1` #### Community - Maintainers: run the assumptions prompt on your issue triage criteria — what unstated assumptions decide what gets fixed? `#action/h2` - Contributors: not applicable to this source. `#action/h3` - Sponsors: not applicable to this source. `#action/h3` - Ecosystem: recurring ecosystem debates are usually framing failures. Five-levels-deep on one of them would be genuinely clarifying. `#action/h2` #### WordPress - Core: feature requests arrive as solutions. Reframing to the underlying problem is the highest-leverage triage move available. `#action/h1` - Core Performance: not applicable to this source. `#action/h3` - Core AI: any AI feature should be evaluated on whether it helps users frame problems, not only answer them. `#action/h2` - Core Program Team: not applicable to this source. `#action/h3` - Mobile: not applicable to this source. `#action/h3` - Playground: not applicable to this source. `#action/h3` - Community: not applicable to this source. `#action/h3` - Training: **strong fit.** Learn WordPress teaches answers; problem-framing is teachable and absent. `#action/h1` - Marketing: not applicable to this source. `#action/h3` - Support: not applicable to this source. `#action/h3` - Design: design research is problem refinement. These prompts are a cheap first pass before user interviews, not a replacement. `#action/h2` - Documentation: the assumptions prompt is the fastest way to find the prerequisite knowledge a doc silently assumes. `#action/h1` - Photos: not applicable to this source. `#action/h3` - TV: not applicable to this source. `#action/h3` - Polyglots: not applicable to this source. `#action/h3` - Accessibility: "what else could be true about who is using this?" is an accessibility prompt with real reach. `#action/h1` - Meta: not applicable to this source. `#action/h3` - CLI: not applicable to this source. `#action/h3` - Hosting: not applicable to this source. `#action/h3` - Tide: not applicable to this source. `#action/h3` - Openverse: not applicable to this source. `#action/h3` - Themes: not applicable to this source. `#action/h3` - Plugins: not applicable to this source. `#action/h3` #### The WPCC - Programs: program design should start from problem refinement rather than a solution someone already has in mind. `#action/h2` - Projects: not applicable to this source. `#action/h3` - Funding experiments: not applicable to this source. `#action/h3` ### FAIR - Technical independence: not applicable to this source. `#action/h3` - Policies and covenants: not applicable to this source. `#action/h3` - Productized governance artifacts: a problem-framing worksheet built from these three prompts would be a small, genuinely useful artifact. `#action/h2` ## 🧪 Hypotheses 1. **Longer, context-rich inputs produce higher-quality outputs than short prompts, holding model and task fixed.** The video's central claim and directly testable. `#atom/literature/h1` 2. **Problem-reframing prompts produce more novel material than answer-seeking prompts on the same topic.** Testable by having blind raters score outputs for novelty against the asker's prior knowledge. `#atom/literature/h1` 3. **A model personalized to a user returns fewer genuinely alternative framings than a cold session, on "what else could be true?" prompts.** The implication she discloses but does not draw. `#atom/literature/h1` 4. **Professional training predicts strength at problem-solving and weakness at problem-identification.** Her core diagnostic claim, and the one with the widest consequences if true. 5. **Practitioners evaluate surfaced assumptions on interestingness rather than truth, absent an explicit falsification step.** Visible throughout this demo. 6. **Apocryphal quotes propagate faster in advice content than sourced ones, because they are more quotable.** The Einstein attribution here is one specimen. `#atom/literature/h2` ### 📊 Methodology - **H1 (input length).** Fixed task set; vary input from a one-line prompt to a full context brief at matched model and temperature. Blind expert rating for specificity and usefulness. Control for output length, since longer inputs tend to produce longer outputs and raters conflate length with quality — this confounder sinks most casual versions of this test. - **H2 (reframing vs. answering).** Same topic, two prompt families. Have the *asker* rate each output for whether it contained something they did not already believe — novelty relative to prior knowledge is the right measure, not novelty in the abstract. - **H3 (personalization narrows alternatives).** The sharpest experiment available here. Run identical "what else could be true?" prompts through a heavily personalized account and a cold session; have a blind rater judge which outputs fall inside the user's known vocabulary. Prediction: the personalized session returns the user's own frameworks more often. This directly tests the implication she leaves open. - **H4 (training asymmetry).** Present matched cohorts with a well-formed problem to solve and an ill-formed situation requiring framing. Compare performance across education levels and years of professional experience. Confounder: prestige-institution selection effects mean the strongest problem-solvers are also the most selected, so control for institution. - **H5 (interestingness vs. truth).** Give participants AI-surfaced assumptions about a domain they know; ask them to mark which they believe true and which they find interesting. Measure divergence, and whether an explicit falsification instruction closes it. - **H6 (quote propagation).** Corpus of advice videos and posts; code quoted maxims for attribution accuracy against Quote Investigator and primary sources. Measure prevalence. Cheap, and the result would be useful to anyone writing in this genre. ### ❓ Questions - Are any of the surfaced assumptions actually true of those students? The video never asks, and asking three students would settle it. - How much of the third framework's output is her own vocabulary returning? She supplies the fact and not the conclusion. - What is the actual cost of the thousand-word input — does the time spent writing context exceed the quality gained for routine tasks? The claim is stated as general. - Is problem-framing teachable in the way she implies, or does it require the domain experience she says AI lacks? - Would a cold session have produced better alternatives precisely *because* it does not know her frameworks? - What is the relationship between Wispr Flow and this channel? No disclosure appears either way, and her Napkin video explicitly denies sponsorship. - Does the "good students" diagnosis hold across educational cultures, or is it specific to systems that reward correct answers over question formulation? - If everyone adopts problem-framing as the differentiator, does it stop differentiating — and what replaces it? ### 🔮 Future Research - **Personalization as a constraint on divergent thinking.** The most interesting open question here. Memory features improve relevance and may narrow the space of framings a model will offer — a real tension for any use case requiring genuine alternatives, and nobody appears to have measured it. - **Problem-identification as a trainable skill.** Her diagnosis implies a curriculum gap with large consequences. Whether framing can be taught directly, or only accrues through domain experience, is an education-research question with immediate practical value. - **Falsification steps in AI-assisted analysis.** Surfacing plausible assumptions is now trivial; checking them is not, and the asymmetry will produce confident, well-structured, wrong analyses at scale. - **Convergent framings across independent creators.** Zhao and Newport reach the invisible-accumulation idea separately in the same batch. Tracking independent convergence would be a useful signal for which ideas in this genre are real. - **Apocryphal authority in advice media.** Quotable-but-false attributions appear to be selected for by the format itself. Measuring the prevalence would be straightforward and mildly damning. ### 💡 Implications - **Practice.** Adopt the three prompts, and add the step the video omits: after surfacing assumptions, go check whether they are true. Use a cold session for "what else could be true?" — personalization helps everywhere except there. - **Policy.** Organizations should evaluate AI use on whether it improved the framing of decisions, not on output volume. The current metrics reward exactly the ten-words-in behavior she identifies as the trap. - **Security.** Not applicable to this source, beyond noting that threat modeling is a problem-framing exercise and inherits the personalization caveat. - **Education.** The strongest implication in the video. If schooling produces excellent solvers of supplied problems, and that is now the cheap skill, then curricula need to teach question formulation explicitly rather than treating it as something acquired incidentally. - **Funding models.** Not applicable to this source. ### 🧠 Learning Styles Assessment - **Reading level estimate:** spoken, conversational; Flesch-Kincaid Grade Level roughly 8–9, Flesch Reading Ease around 60–68. Plain vocabulary with some business register (leverage, differentiation, 80/20), and long self-interrupting sentences characteristic of unscripted speech. #### 👀 Visual learners - Present: screen recording of the full ChatGPT session; an on-screen prompt-syntax reference she invites viewers to screenshot. - Improve: the three frameworks — the video's actual deliverable — never appear as a single card. The five-levels chain is the strongest content and exists only as scrolling chat text. #### 🔊 Aural learners - Present: energetic narration; genuine real-time reaction that conveys which results mattered to her. - Improve: fast pace with frequent self-interruption (*"how do I phrase this"*), and she reads long model outputs aloud only partially, so listeners get her reaction without the content. #### 🗣️ Verbal learners - Present: strong. *"Most employees are good students"* is compact and lands. The ten-words/thousand-words contrast is memorable and precise. - Improve: the Einstein attribution is false, and "the 1%" and "99% of people" are rhetorical figures presented as proportions. #### 🧩 Kinesthetic learners - Present: three prompts demonstrated end-to-end on a real problem, immediately reusable. - Improve: the prompts are spoken and shown on screen but not provided in text, so viewers transcribe. No exercise, and no falsification step to practice. #### 🧠 Logical learners - Present: the argument is well-built — dismiss the distractions, name the two user types, identify the wrong optimization, diagnose the cause, prescribe three tools, demonstrate. Each stage motivates the next. - Improve: the demo validates on interestingness rather than truth, and the personalization disclosure creates a logical gap she leaves open. A logical viewer will notice that the "alternatives" prompt is compromised by exactly the memory she just described. #### 🧑‍🤝‍🧑 Social learners - Present: credits Terry for the tool referral; references workshop and accelerator participants; the guest-lecture story grounds the argument in a real audience. - Improve: the students are discussed entirely in the third person as a problem to be understood. The video's own thesis — refine the problem — would suggest asking them. #### 🧘‍♀️ Solitary learners - Present: entirely self-directed; the frameworks work alone. - Improve: no guidance on when a problem merits this treatment versus a direct question. #### 🌿 Naturalistic learners - Present: nothing environmental. - Improve: not a meaningful gap for this source. ### 📊 Accessibility Assessment *Assessment of the video as published on YouTube.* #### Screen reader semantics - Present: accurate, descriptive title. - Improve: substantial on-screen model output is summarized rather than read, so a non-sighted viewer receives her reactions without the material prompting them. Deictic narration throughout (*"here we go,"* *"look at this one"*). WCAG 2.2 AA 1.2.3 Audio Description or Media Alternative. #### Keyboard navigation - Present: standard YouTube player. - Improve: the dictation workflow she recommends is arguably an accessibility *gain* — voice input lowers a real barrier — and she never frames it that way, which is a missed opportunity. #### Contrast and readability - Present: native-resolution capture of the ChatGPT interface. - Improve: long chat responses shown at full-page scale will be unreadable at phone sizes, and much of an 18-minute video is watched on phones. WCAG 2.2 AA 1.4.3, 1.4.4 Resize Text. #### Text resizing and reflow - Present: YouTube handles player reflow. - Improve: recorded UI does not reflow; zooming into the passage under discussion is used inconsistently. WCAG 2.2 AA 1.4.10 Reflow. #### Media and captions - Present: captions available. - Improve: the caption track garbles the product name throughout — "chat GPT," "Chad GPT," "chatbt," "Chip PT," "tragedy," "Shibb," "CHBT" — in a video whose subject is that product. "Whisper Flow" for Wispr Flow. Sparse punctuation produces run-on text harder to parse than the audio. WCAG 2.2 AA 1.2.2 Captions (Prerecorded). #### Forms and interactions - Present: not applicable to video content. - Improve: the linked tool requires signup; that flow is unassessed. #### Standards reference - Primary gaps: **1.2.2** (product name unreliable throughout the caption track) and **1.2.3** (model output summarized rather than read). Both matter more than usual because the content is a demonstration people are meant to follow. ### 🌍 DEIB Assessment #### Cultural breadth - Present: the problem-framing argument is broadly applicable, and the observation about schooling rewarding correct answers holds across many educational systems. - Improve: the framing is corporate and Western — Google, Amazon, Shopify as the reference employers, consulting methods as the toolkit, and an implied audience with a ChatGPT subscription and discretionary time. The "good students" diagnosis is presented as universal when it is specific to systems built around graded correct answers. #### Inclusive language - Present: warm and encouraging; *"This is such good news"* about experience and judgment being the durable assets is genuinely reassuring to experienced workers anxious about displacement. - Improve: "the 1%" versus "99% of people" creates an in-group framing around a claim with no measurement behind it. "Framework thinker because you're on this side of the internet" is a flattering in-group marker. #### Neurodiversity support - Present: clear enumeration (three frameworks, five levels), consistent structure, and concrete demonstration before abstraction. - Improve: rapid pace, frequent self-interruption, and no chapter markers in an 18-minute video that functions as a reference. The excitement is engaging and adds processing load. #### Representation - Present: credits Terry for a tool referral and references workshop participants and accelerator members. - Improve: the students at the center of the analysis never speak. The video builds an elaborate account of their psychology — identity protection, fear of uncertainty — entirely from a model's inference plus her own observation, and never suggests asking them. That is precisely the gap her own framework should surface. #### Psychological safety - Present: **genuinely strong on one axis.** Naming that uncertainty feels like incompetence, and that people protect an identity as smart students, is compassionate and likely to land for viewers experiencing exactly that. Her note that she uses games in workshops to defuse that fear is a real practice, not a platitude. - Improve: the disclosure about her personalized ChatGPT protects viewers from concluding their tool is broken — good. But viewers are still shown a process that stops at plausible-and-interesting, and a viewer who builds a stakeholder-facing analysis this way inherits a confident, well-structured, unverified account of other people's motivations. ### ✅ Logical Validity #### Strong reasoning - **Correctly identifying the wrong optimization (00:50).** The ten-words/thousand-words contrast rests on a stated premise — knowledge workers are not paid hourly — and the conclusion follows: optimizing an AI interaction for speed optimizes a variable that does not carry your value. Naming both failure modes (the expert who quits, the box-checker) shows she is describing a real distribution rather than a straw man. - **The good-students diagnosis (03:20).** The strongest reasoning in the video. It supplies a *mechanism* — schooling trains answering supplied questions, that skill transfers to work, models now perform it cheaply — that explains both the observed behavior and why it persists. It also generates the correct prediction about where remaining human value sits. - **Demonstrating the process on a live problem (06:44 onward).** Using a real brief she had not yet solved, rather than a worked example, means the insights arrive in the demo rather than being restated. The performance/judgment distinction at 13:33 emerges from the five-whys chain rather than preceding it, which is why the demonstration persuades. - **Judgment applied to output (15:57).** Rejecting the "consulting jargon" hypothesis on domain grounds — these are finance students, for whom that vocabulary is native — models the critical filtering the video is implicitly arguing for. She also separates what confirmed her existing view from what was genuinely new (*"this is something I don't know"*), which is honest epistemic bookkeeping. - **Disclosing the personalization confound (14:30).** *"I use ChatGPT a lot so it knows me... if your ChatGPT doesn't do that yet it's not a problem."* Rare, viewer-protective, and against her interest in appearing to demonstrate a universally reproducible result. #### Fallacies - **False attribution** — 02:58: the "fifty-five minutes on the problem, five on the solution" maxim is attributed to Einstein. It is apocryphal, with no basis in his writings or documented speech ([Quote Investigator](https://quoteinvestigator.com/2014/05/22/solve/)). It is the rhetorical anchor for the central thesis, and the thesis is stronger without borrowed authority. **Fix:** state the principle unattributed. `#fallacy/appeal-to-authority` - **Disclosed confound left unresolved** — 14:30: having established that the model returns inversion and red-teaming *because she taught it those frameworks*, she presents its output as alternatives that expand her thinking. A model tuned to your vocabulary is the wrong instrument for a prompt whose purpose is escaping your frame. She supplies the premise and declines the conclusion. **Fix:** run "what else could be true?" in a cold session. `#fallacy/begging-the-question` - **Interestingness substituted for truth** — throughout the demo: no surfaced assumption is ever tested against reality. Whether students actually protect an identity as smart students is treated as settled because the claim is compelling. The video's own thesis — refine the problem properly — should have caught this. **Fix:** one falsification step, e.g. ask three students. `#fallacy/unsubstantiated-claim` - **Invented proportions** — 02:58 and 01:30: *"99% of the people"* and *"what the 1% does"* are presented as descriptions of a distribution with no measurement. They also do in-group work, flattering the viewer into the minority. `#fallacy/unsubstantiated-claim` - **Confirmation read as validation** — 08:52 onward: several outputs are praised specifically because they match what she had already put in the lecture (*"this is actually how I started the lecture"*). A model agreeing with your existing material is weak evidence of quality. She is transparent that this happened, which mitigates it, but still treats the agreement as a success signal. `#fallacy/confirmation-bias` - **Enthusiasm as evidence** — throughout: *"So good, right?"*, *"I love this"*, *"goosebumps"*, *"light bulb moment"*, *"I'm getting too over excited."* Affect substitutes for assessment at the exact points where the output is least verifiable. `#fallacy/appeal-to-emotion` - **Undisclosed relationship status** — 05:30: Wispr Flow is recommended with a promised link and a referral credit, with no statement of whether any commercial relationship exists. Her Napkin video explicitly says *"this isn't sponsored,"* so the absence here is a gap in her own established practice rather than an accusation. `#fallacy/suppressed-evidence` #### Tone of voice Energetic, warm, and thinking-aloud — she is reacting in real time rather than presenting conclusions, and the video is better for it. The self-interruptions (*"how do I phrase this"*), the admission of veering off topic, and the visible excitement when a result surprises her all signal genuine engagement rather than performance. When she says *"this is something I don't know,"* it reads as true. The cost is that enthusiasm carries evaluative weight the content has not earned. *"So good, right?"* and *"I love this"* arrive at moments where the output is plausible and unverified, and the register does not change between a result that confirms what she already wrote and one that is genuinely new — even though she separately, and creditably, notes the difference. The effect is that a viewer absorbs "these prompts produce insight" from the affect rather than from any demonstrated check. The most admirable tonal moment is the personalization disclosure, which is delivered plainly and against interest. The least examined is the enthusiasm immediately following it, which proceeds as though the caveat had not been made. ### 🛡️ DARVO Analysis DARVO describes an accused party's response pattern — Deny, Attack, Reverse Victim and Offender. **No DARVO pattern is present.** This is instructional content with no allegation, accuser, or defensive posture. Component by component: - **Deny:** Absent. Nothing is denied because nothing is alleged. The structurally notable moment runs the other way — at 14:30 she volunteers a confound that weakens her own demonstration (*"I use ChatGPT a lot so it knows me"*), which is the opposite of denial. She also volunteers that several results merely confirmed material she had already written. The failure in this video is that she does not act on a disclosure she made, not that she withheld one. - **Attack:** Absent. No person, company, or competing approach is disparaged. She criticizes two *usage patterns* — the expert who quits after one bad output, and the box-checker — but frames both as understandable mistakes rather than character failings, and includes herself in the population of people learning. Tool comparisons are explicitly declared irrelevant (*"Those things don't matter"*) rather than used to score points. The students at the center of the analysis are described sympathetically, as capable people shaped by their training. - **Reverse victim/offender:** Absent. No grievance is claimed. The framing throughout is that the audience holds an advantage — experience, judgment, intuition — that AI lacks, which is an empowerment frame rather than a victimhood one. **Emotional cues and rhetoric.** High enthusiasm, in-group flattery (*"framework thinker because you're on this side of the internet,"* the 1%/99% split), and a closing pointer to her own companion video. Ordinary creator rhetoric aimed at engagement, none of it defensive. **Related fallacies.** The real weaknesses are captured above: `#fallacy/appeal-to-authority` on the Einstein misattribution, `#fallacy/begging-the-question` on the unresolved personalization confound, and `#fallacy/unsubstantiated-claim` on validating interestingness rather than truth. All are reasoning gaps in good-faith instructional content with no adversary. `#darvo/deny` `#darvo/attack` `#darvo/reverse` — all three assessed. None present. ## 🎯 Conclusion - **Bottom line.** The best framing in this batch of *what to use AI for*, and it works by refusing the two arguments everyone else has. Not syntax, not tool choice — the question is whether you are extracting answers or refining problems. Her ten-words-in versus thousand-words-in contrast is genuinely counterintuitive and rests on a premise worth internalizing: knowledge workers are not paid hourly, so optimizing an AI conversation for speed optimizes a variable that does not carry your value. The diagnosis underneath is sharper still — *"most employees are good students"* — which explains why capable professionals default to solving whatever problem they are handed, and why that skill is precisely the one models now perform cheaply. The three prompts (assumptions, five levels deep, alternatives) are concrete and reusable, and the live demo earns its length because the best insight arrives *during* the process rather than before it. Three things to correct on the way in: the Einstein quote anchoring the thesis is apocryphal; the demo validates on interestingness rather than truth and never once tests whether a surfaced assumption is real; and she honestly discloses that her model has learned her frameworks without following that to its conclusion — a personalized model is the wrong instrument for "what else could be true?" - **Next steps for viewers.** 1. Adopt the three prompts on your next real problem, and give the model a paragraph of context rather than a sentence. The thousand-words-in habit is the whole technique. 2. Run the **alternatives** prompt in a **cold session**. Personalization helps everywhere except the one prompt whose job is escaping your own frame. 3. Add the falsification step the video omits. After surfacing assumptions, go check two or three against reality — ask a person, look at data. Plausible-and-interesting is not the same as true, and it feels identical. 4. Notice which outputs merely confirm what you already wrote. She flags this honestly; treat agreement as weak evidence, not as validation. 5. Drop the Einstein attribution if you repeat the maxim. The principle is sound and does not need the borrowed name. 6. Pair this with her 17-minute workflow video — this one improves the *question*, that one improves the *material*. They are two halves of the same argument. ## 🔄 Refactor Appendix — 3 ChatGPT Prompts I Use to Standout At Work ### 📚 Literature Atoms (H3) #### [H1] Professional training produces solvers, not framers > [!quote] ID:rw_vzprompt_h1 > "the boss gives us the problem. We say okay I'm going to tell you exactly how to do this. We don't even think whether the problem brought to us is wrong. Most employees are good students." **Context & use**: - 03:20 — extending the guest-lecture observation to the workplace. - Supplies a mechanism, not just an observation: schooling trains answering supplied questions, that transfers to work, models now do it cheaply. Explains where remaining human value sits. **Tags**: #xref #articleseed #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=200) <!-- atomicity: 1 idea ✓ --> #### [H2] Optimize the interaction for quality, not for speed > [!quote] ID:rw_vzprompt_h2 > "your leverage doesn't come from time right you're not being paid by the hour what your value is is the quality of the idea" **Context & use**: - 01:30 — the premise beneath the ten-words/thousand-words contrast. - Explains why the efficiency framing misleads: a thousand-word input yielding a thousand-word output is a good trade when the output exceeds what you could write alone in that time. **Tags**: #xref #share **Highlight**: [view highlight](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=90) <!-- atomicity: 1 idea ✓ --> #### [H3] Performance and judgment are different capacities > [!quote] ID:rw_vzprompt_h3 > "they were trained to optimize for performant not judgment" **Context & use**: - 13:33 — the terminal link in the five-levels chain, and the video's best emergent insight. - Judgment is choosing well among options without a complete answer; doers excel in low-complexity environments where that is not required. Arrived at through the demonstrated process rather than asserted beforehand. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=813) <!-- atomicity: 1 idea ✓ --> #### [H4] A personalized model returns your own frame > [!quote] ID:rw_vzprompt_h4 > "I use tragedy a lot so it knows me. It knows my frameworks. It knows my mental models. That's why it's giving these inversion and red teaming." **Context & use**: - 14:30 — an unusually honest reproducibility disclosure ("tragedy" is an ASR error for ChatGPT). - The implication she leaves open: if the model echoes your vocabulary, its "alternatives" are partly your own. Fine for surfacing assumptions; wrong instrument for escaping a frame. **Tags**: #xref #articleseed **Highlight**: [view highlight](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=870) <!-- atomicity: 1 idea ✓ --> #### [H5] Fluency is accumulated work made invisible > [!quote] ID:rw_vzprompt_h5 > "the best authors when they give interviews is not because they can think on the spot. It's because they spent years writing out these ideas refining them" **Context & use**: - 16:50 — a flagged digression. - Independently reaches Cal Newport's invisible-staircase argument from elsewhere in this batch, applied to articulateness rather than expertise. Convergence between unconnected sources is itself evidence. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=pUG7_03G_Ac&t=1010) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Use the model to sharpen the question, not to supply the answer **Claim** (one sentence, Courtney's voice): The scarce contribution is deciding which problem is worth solving, so point the model at the framing rather than at the output. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds wherever the problem arrives pre-specified by someone else — assigned work, feature requests, support tickets, briefs. Strongest when the specifier lacks context the executor has. Does not apply where the problem genuinely is well-formed and the work is execution; not every task needs reframing, and reframing everything is its own failure mode. Diagnostic: can you state why *this* problem rather than an adjacent one? **Links** (≥1 with annotation): - [[Problem framing]] — because extends - [[Grill me]] — because extends the same interrogate-before-executing family - [[Five whys]] — because applies-in root-cause work <!-- atomicity: 1 idea ✓ --> #### [P2] Ask for alternatives somewhere the model does not know you **Claim** (one sentence, Courtney's voice): A model personalized to your vocabulary cannot reliably show you outside your frame, so run divergent prompts in a session with no memory of you. **Evidence pointer**: [[#H4]] **Scope & conditions**: Holds for any prompt whose purpose is escaping an existing frame — alternatives, counterarguments, red-teaming, threat modeling, pre-mortems. Inverts for convergent work, where personalization genuinely helps: summarizing in your voice, applying your established frameworks, drafting in your register. The tell is whether you want the output to sound like you (keep memory) or to surprise you (drop it). **Links** (≥1 with annotation): - [[Confirmation bias]] — because extends - [[Red teaming]] — because applies-in adversarial review - [[Filter bubble]] — because extends the same personalization-narrows-input dynamic `proof_state: working` <!-- atomicity: 1 idea ✓ --> #### [P3] Plausible and interesting is not the same as true **Claim** (one sentence, Courtney's voice): Generated explanations are selected for being compelling rather than correct, so an analysis without a falsification step produces confident structure over unverified content. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for any AI-assisted analysis of causes, motivations, or mechanisms — user psychology, root causes, market behavior, other people's reasoning. Sharpest where the subject is unavailable to check, which is exactly when the temptation is greatest. Does not apply to verifiable outputs (code that runs, arithmetic, citations you open). Practical test: name what would falsify the claim, then go look. **Links** (≥1 with annotation): - [[Falsification]] — because extends - [[A finding is a claim until something runs]] — because extends the same evidence discipline - [[Confirmation loop]] — because contrasts on how the failure completes `proof_state: working` <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Run "what else could be true?" in a cold session vs. my personalized one on the same problem. Compare. Sharpest experiment from this batch. #fleeting #inbox - [F2] Add a falsification step to my own analysis habits — after surfacing assumptions, name what would disprove each. #fleeting #inbox - [F3] Post idea: "most employees are good students" — schooling trains the skill AI just commoditized. Strong thesis, connects to Newport. #fleeting #articleseed - [F4] Zhao and Newport independently reach the invisible-accumulation idea. Convergence across unconnected sources as a signal for what's real in this genre. #fleeting #articleseed - [F5] The Einstein quote is apocryphal — check the other maxims quoted across this batch against Quote Investigator. #fleeting #inbox - [F6] Run the assumptions prompt against my own skill definitions — what do they assume the reader already knows? #fleeting #inbox - [F7] Memory/personalization as a constraint on divergent thinking. Nobody has measured this and it affects every agent I run. #fleeting #articleseed ### → Things 3 (max 2 actions) - [ ] Test cold vs. personalized session on an alternatives prompt and write up the difference — research #things - [ ] Draft the "good students" post connecting schooling, problem-framing, and what AI commoditized — blog #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/3 ChatGPT Prompts I Use to Standout at Work.md|3 ChatGPT Prompts I Use to Standout at Work by Vicky Zhao]] ## Highlights - they were trained to optimize for performing, not judgment. ([View Highlight](https://read.readwise.io/read/01m1c8xndkxt3pjhz7mm6f5stk)) - Note: The best emergent insight in the video — and it arrives THROUGH the five-whys process rather than being asserted beforehand, which is what makes the demo persuasive. The chain: never taught → environment rewards correctness over framing → uncertainty feels like failure → protecting an identity as a smart student → trained to optimize for performance, not judgment. Judgment is choosing well among options without a complete answer, and it's what separates top performers from very capable doers. Pairs with "most employees are good students": schooling trains the exact skill models just commoditized. - Tags: atom/zettel/z1, atom/literature/h1